WEBVTT NOTE This file was exported by MacCaption version 7.0.06 to comply with the WebVTT specification dated March 27, 2017. 00:00:09.543 --> 00:00:11.778 align:center line:-1 position:50% size:34% USGS Landslide Seminar, which is hosted by 00:00:11.778 --> 00:00:13.814 align:center line:-1 position:50% size:42% our Landslide Hazards Program, 00:00:13.814 --> 00:00:15.482 align:center line:-1 position:50% size:43% and coordinated by Matt Thomas, 00:00:15.482 --> 00:00:17.851 align:center line:-1 position:50% size:30% Jaime Kostelnik, and Stephen Slaughter. 00:00:17.851 --> 00:00:20.487 align:center line:-1 position:50% size:37% None of them were available to host today, 00:00:20.487 --> 00:00:23.190 align:center line:-1 position:50% size:31% so you're stuck with me, Ben Mirus. 00:00:23.190 --> 00:00:25.792 align:center line:-1 position:50% size:38% I think they have this well-oiled machine going, 00:00:25.792 --> 00:00:27.327 align:center line:-1 position:50% size:30% so please bear with me 00:00:27.327 --> 00:00:29.363 align:center line:-1 position:50% size:28% if there are any errors or mistakes. 00:00:29.363 --> 00:00:31.131 align:center line:-1 position:50% size:31% One note is we're under a tornado 00:00:31.131 --> 00:00:33.367 align:center line:-1 position:50% size:40% and severe thunderstorm watch here in Golden at the moment, 00:00:33.367 --> 00:00:37.471 align:center line:-1 position:50% size:31% so if everything crashes, that's what's happening. 00:00:38.906 --> 00:00:40.240 align:center line:-1 position:50% size:37% For those of you who are new to this meeting, 00:00:40.240 --> 00:00:42.009 align:center line:-1 position:50% size:30% you will have the ability to submit questions 00:00:42.009 --> 00:00:44.177 align:center line:-1 position:50% size:38% in the chat window or using the "Raise your hand" feature 00:00:44.177 --> 00:00:46.413 align:center line:-1 position:50% size:37% when turning on your microphone afterwards. 00:00:46.413 --> 00:00:49.316 align:center line:-1 position:50% size:40% But we'll wait to keep questions until the end of the talk 00:00:49.316 --> 00:00:50.918 align:center line:-1 position:50% size:36% and have a managed discussion then. 00:00:50.918 --> 00:00:52.219 align:center line:-1 position:50% size:26% So in the meantime, please keep 00:00:52.219 --> 00:00:54.354 align:center line:-1 position:50% size:36% your microphones muted and your cameras turned off 00:00:54.354 --> 00:00:56.523 align:center line:-1 position:50% size:29% to preserve bandwidth. 00:00:56.523 --> 00:00:58.358 align:center line:-1 position:50% size:36% Just a quick announcement: 00:00:58.358 --> 00:01:00.460 align:center line:-1 position:50% size:39% Next week we'll be hearing from Dan Coe 00:01:00.460 --> 00:01:01.929 align:center line:-1 position:50% size:24% of the Washington Geological Survey 00:01:01.929 --> 00:01:04.698 align:center line:-1 position:50% size:43% with a talk entitled, "Communicating Landslide Risk-- 00:01:04.698 --> 00:01:06.533 align:center line:-1 position:50% size:28% Landslide Information and Hazards 00:01:06.533 --> 00:01:07.935 align:center line:-1 position:50% size:31% with Maps and Graphics 00:01:07.935 --> 00:01:11.371 align:center line:-1 position:50% size:25% at the Washington Geological Survey." 00:01:11.371 --> 00:01:13.140 align:center line:-1 position:50% size:27% Today I am taking on the dual role 00:01:13.140 --> 00:01:15.309 align:center line:-1 position:50% size:33% of hosting and introducing today's speaker, 00:01:15.309 --> 00:01:18.211 align:center line:-1 position:50% size:35% Jacob Woodard, who is a Mendenhall fellow 00:01:18.211 --> 00:01:21.214 align:center line:-1 position:50% size:30% with us here at the USGS in Golden. 00:01:21.214 --> 00:01:22.783 align:center line:-1 position:50% size:27% Before coming to us, Jacob pursued 00:01:22.783 --> 00:01:25.452 align:center line:-1 position:50% size:28% his bachelor's degree in Geology at BYU 00:01:25.452 --> 00:01:27.421 align:center line:-1 position:50% size:25% and earned both a master's and PhD 00:01:27.421 --> 00:01:29.923 align:center line:-1 position:50% size:28% from the University of Wisconsin-Madison 00:01:29.923 --> 00:01:34.094 align:center line:-1 position:50% size:41% with an emphasis in Geophysics and Surface Processes. 00:01:34.094 --> 00:01:36.129 align:center line:-1 position:50% size:39% His research there was focused on understanding 00:01:36.129 --> 00:01:38.665 align:center line:-1 position:50% size:38% subglacial erosion processes. 00:01:38.665 --> 00:01:40.867 align:center line:-1 position:50% size:38% And I took him to some of my favorite places 00:01:40.867 --> 00:01:44.471 align:center line:-1 position:50% size:31% in the world, like Alberta and Iceland. 00:01:44.471 --> 00:01:46.606 align:center line:-1 position:50% size:34% But promptly after defending his dissertation, 00:01:46.606 --> 00:01:50.544 align:center line:-1 position:50% size:43% Jacob moved with his family here in the summer of 2020, 00:01:50.544 --> 00:01:52.746 align:center line:-1 position:50% size:38% right in the middle of those pandemic lockdowns 00:01:52.746 --> 00:01:55.916 align:center line:-1 position:50% size:36% and the birth of this seminar. 00:01:55.916 --> 00:01:57.217 align:center line:-1 position:50% size:24% And so it was an interesting time 00:01:57.217 --> 00:01:59.219 align:center line:-1 position:50% size:28% and challenging for us to get research going, 00:01:59.219 --> 00:02:01.722 align:center line:-1 position:50% size:43% but Jacob's really excelled and has done some exciting work 00:02:01.722 --> 00:02:04.424 align:center line:-1 position:50% size:33% on landslide susceptibility with new data 00:02:04.424 --> 00:02:06.760 align:center line:-1 position:50% size:39% and new modeling techniques, and so that's 00:02:06.760 --> 00:02:08.662 align:center line:-1 position:50% size:25% what he's gonna present to us today. 00:02:08.662 --> 00:02:11.965 align:center line:-1 position:50% size:40% It's my pleasure to hand the screen off to Jacob. 00:02:11.965 --> 00:02:14.401 align:center line:-1 position:50% size:28% -Jacob, thank you. -All right, thanks, Ben. 00:02:14.401 --> 00:02:15.869 align:center line:-1 position:50% size:17% Appreciate it. 00:02:15.869 --> 00:02:17.237 align:center line:-1 position:50% size:26% Yeah, it's a pleasure to be here 00:02:17.237 --> 00:02:20.007 align:center line:-1 position:50% size:31% and I'm excited to share with everybody 00:02:20.007 --> 00:02:23.744 align:center line:-1 position:50% size:40% some work that I've been doing here at the USGS. 00:02:23.744 --> 00:02:25.412 align:center line:-1 position:50% size:34% Basically figuring out ways 00:02:25.412 --> 00:02:27.381 align:center line:-1 position:50% size:25% to make meaningful susceptibility maps 00:02:27.381 --> 00:02:30.984 align:center line:-1 position:50% size:25% over large regions with imperfect data. 00:02:30.984 --> 00:02:32.586 align:center line:-1 position:50% size:35% Here's the government disclaimers 00:02:32.586 --> 00:02:34.388 align:center line:-1 position:50% size:25% that are mandatory. 00:02:34.388 --> 00:02:36.623 align:center line:-1 position:50% size:28% For this talk, first gonna go through 00:02:36.623 --> 00:02:38.458 align:center line:-1 position:50% size:28% kinda the problems that we're confronting, 00:02:38.458 --> 00:02:42.462 align:center line:-1 position:50% size:40% and then we're gonna try to answer these two big questions 00:02:42.462 --> 00:02:45.699 align:center line:-1 position:50% size:33% that I keep on running into as I'm trying to develop 00:02:45.699 --> 00:02:47.701 align:center line:-1 position:50% size:32% these susceptibility maps over large regions. 00:02:47.701 --> 00:02:49.770 align:center line:-1 position:50% size:39% First is how to manage regions 00:02:49.770 --> 00:02:52.773 align:center line:-1 position:50% size:26% with limited or no landslide data, 00:02:52.773 --> 00:02:55.709 align:center line:-1 position:50% size:30% and to evaluate this we're gonna investigate 00:02:55.709 --> 00:02:57.310 align:center line:-1 position:50% size:28% the influence of the training domain 00:02:57.310 --> 00:02:59.413 align:center line:-1 position:50% size:30% or the data that you use to develop your models 00:02:59.413 --> 00:03:01.915 align:center line:-1 position:50% size:33% and the limits of that data. 00:03:01.915 --> 00:03:04.017 align:center line:-1 position:50% size:36% And the next question that we keep on confronting 00:03:04.017 --> 00:03:06.553 align:center line:-1 position:50% size:41% is how to manage the imprecise data that we have 00:03:06.553 --> 00:03:09.089 align:center line:-1 position:50% size:34% available to us for creating these susceptibility maps. 00:03:09.089 --> 00:03:11.191 align:center line:-1 position:50% size:35% This isn't a unique problem. 00:03:11.191 --> 00:03:14.127 align:center line:-1 position:50% size:36% Basically any data set that we have available to us 00:03:14.127 --> 00:03:17.631 align:center line:-1 position:50% size:34% is imperfect and imprecise in its very nature. 00:03:17.631 --> 00:03:19.699 align:center line:-1 position:50% size:36% And so how can we mitigate 00:03:19.699 --> 00:03:21.334 align:center line:-1 position:50% size:29% the imprecision with the model output? 00:03:21.334 --> 00:03:24.004 align:center line:-1 position:50% size:43% How to best represent the model. 00:03:24.004 --> 00:03:26.706 align:center line:-1 position:50% size:36% And we're gonna--to do that we're gonna dive into 00:03:26.706 --> 00:03:29.142 align:center line:-1 position:50% size:30% different mapping units, and the pros and cons 00:03:29.142 --> 00:03:33.013 align:center line:-1 position:50% size:32% of the mapping units and how they can adjust, 00:03:33.013 --> 00:03:36.883 align:center line:-1 position:50% size:33% or play a role in displaying the imprecise nature 00:03:36.883 --> 00:03:38.819 align:center line:-1 position:50% size:22% of the input data. 00:03:38.819 --> 00:03:40.587 align:center line:-1 position:50% size:27% So first, introduction. 00:03:40.587 --> 00:03:42.656 align:center line:-1 position:50% size:37% I don't think this needs to be said to this group here, 00:03:42.656 --> 00:03:46.193 align:center line:-1 position:50% size:39% but landslides cause damages and losses of life 00:03:46.193 --> 00:03:49.529 align:center line:-1 position:50% size:38% ever year in the United States and across the world. 00:03:49.529 --> 00:03:53.400 align:center line:-1 position:50% size:28% So here's an example of the 1985 earth-- 00:03:53.400 --> 00:03:57.137 align:center line:-1 position:50% size:36% sorry, landslide that resulted in over 100 fatalities. 00:03:57.137 --> 00:04:01.675 align:center line:-1 position:50% size:40% Then more currently Hurricane Maria in Puerto Rico. 00:04:01.675 --> 00:04:04.644 align:center line:-1 position:50% size:24% And so this causes damages to life, 00:04:04.644 --> 00:04:08.515 align:center line:-1 position:50% size:21% to infrastructure, to residences, 00:04:08.515 --> 00:04:11.952 align:center line:-1 position:50% size:33% and we can expect that with the changing climate 00:04:11.952 --> 00:04:14.621 align:center line:-1 position:50% size:35% the severity of these storms that cause 00:04:14.621 --> 00:04:18.158 align:center line:-1 position:50% size:32% these types of landslides is gonna increase. 00:04:18.158 --> 00:04:21.761 align:center line:-1 position:50% size:25% Here's an example of a damaged road. 00:04:21.761 --> 00:04:24.264 align:center line:-1 position:50% size:40% And this causes both damages to loss and life, 00:04:24.264 --> 00:04:26.933 align:center line:-1 position:50% size:39% which is incalculable, but also, 00:04:26.933 --> 00:04:28.268 align:center line:-1 position:50% size:24% when you damage the infrastructure, 00:04:28.268 --> 00:04:30.770 align:center line:-1 position:50% size:40% there's a lot of economic losses associated with that. 00:04:30.770 --> 00:04:32.239 align:center line:-1 position:50% size:24% If you have a road, a pivotal road, 00:04:32.239 --> 00:04:34.374 align:center line:-1 position:50% size:24% that's closed down for a long time, 00:04:34.374 --> 00:04:37.444 align:center line:-1 position:50% size:37% it's important to try to reduce these losses as best we can, 00:04:37.444 --> 00:04:39.646 align:center line:-1 position:50% size:24% and that's kind of been the objective 00:04:39.646 --> 00:04:41.548 align:center line:-1 position:50% size:24% of my project here. 00:04:41.548 --> 00:04:44.951 align:center line:-1 position:50% size:35% And kinda the first measure that you can do 00:04:44.951 --> 00:04:46.586 align:center line:-1 position:50% size:37% to try to mitigate these losses 00:04:46.586 --> 00:04:49.356 align:center line:-1 position:50% size:32% with the creation of the susceptibility map. 00:04:49.356 --> 00:04:51.992 align:center line:-1 position:50% size:29% So here is an example of a susceptibility map 00:04:51.992 --> 00:04:54.027 align:center line:-1 position:50% size:23% from Puerto Rico. 00:04:54.027 --> 00:04:55.996 align:center line:-1 position:50% size:27% And as I said before, 00:04:55.996 --> 00:04:57.764 align:center line:-1 position:50% size:30% this is kinda the first line of defense. 00:04:57.764 --> 00:05:01.568 align:center line:-1 position:50% size:31% And so what precisely is a susceptibility map? 00:05:01.568 --> 00:05:04.104 align:center line:-1 position:50% size:36% Well, what it does is it measures the likelihood 00:05:04.104 --> 00:05:06.239 align:center line:-1 position:50% size:30% of landslide occurrence at a given location 00:05:06.239 --> 00:05:08.074 align:center line:-1 position:50% size:23% given the local terrain conditions, 00:05:08.074 --> 00:05:09.543 align:center line:-1 position:50% size:24% and that's it, right? 00:05:09.543 --> 00:05:12.312 align:center line:-1 position:50% size:31% It's just a spatial component of-- 00:05:12.312 --> 00:05:14.381 align:center line:-1 position:50% size:31% And typically the outputs are the probability 00:05:14.381 --> 00:05:18.685 align:center line:-1 position:50% size:42% of observing a mapped landslide within a given mapping unit. 00:05:18.685 --> 00:05:22.722 align:center line:-1 position:50% size:35% And so a susceptibility map and a susceptibility model 00:05:22.722 --> 00:05:24.591 align:center line:-1 position:50% size:32% are inherently imprecise, 00:05:24.591 --> 00:05:26.560 align:center line:-1 position:50% size:39% and this is because of the data that we have available 00:05:26.560 --> 00:05:28.195 align:center line:-1 position:50% size:23% for creating these. 00:05:28.195 --> 00:05:29.930 align:center line:-1 position:50% size:41% If we had timing and magnitude, 00:05:29.930 --> 00:05:32.165 align:center line:-1 position:50% size:38% we wouldn't try to create a susceptibility map, 00:05:32.165 --> 00:05:34.467 align:center line:-1 position:50% size:41% we'd try to create a hazard map, 00:05:34.467 --> 00:05:36.570 align:center line:-1 position:50% size:36% which, as the name implies, 00:05:36.570 --> 00:05:39.973 align:center line:-1 position:50% size:31% incorporates magnitude, time, and location. 00:05:39.973 --> 00:05:42.309 align:center line:-1 position:50% size:38% But over much of the United-- 00:05:42.309 --> 00:05:44.211 align:center line:-1 position:50% size:33% most of the United States and the world, 00:05:44.211 --> 00:05:47.047 align:center line:-1 position:50% size:31% we don't have the timing and magnitude 00:05:47.047 --> 00:05:48.381 align:center line:-1 position:50% size:31% of most of the landslides 00:05:48.381 --> 00:05:51.051 align:center line:-1 position:50% size:33% that we have any sort of information on. 00:05:51.051 --> 00:05:54.387 align:center line:-1 position:50% size:36% And so what good is really a susceptibility map 00:05:54.387 --> 00:05:56.523 align:center line:-1 position:50% size:36% if the output's so imprecise? 00:05:56.523 --> 00:05:58.658 align:center line:-1 position:50% size:26% Well, what it does is it highlights areas 00:05:58.658 --> 00:06:00.327 align:center line:-1 position:50% size:26% that are more prone to landsliding 00:06:00.327 --> 00:06:03.830 align:center line:-1 position:50% size:25% by using evidences of past landslides. 00:06:05.232 --> 00:06:08.268 align:center line:-1 position:50% size:30% And again, just kinda emphasizing this point: 00:06:08.268 --> 00:06:12.973 align:center line:-1 position:50% size:34% It's--we're limited to this susceptibility output 00:06:12.973 --> 00:06:14.574 align:center line:-1 position:50% size:30% because we don't have the data available 00:06:14.574 --> 00:06:16.443 align:center line:-1 position:50% size:27% for something better. 00:06:16.443 --> 00:06:19.279 align:center line:-1 position:50% size:38% But people have been making susceptibility maps 00:06:19.279 --> 00:06:20.981 align:center line:-1 position:50% size:19% for a long time, 00:06:20.981 --> 00:06:22.449 align:center line:-1 position:50% size:35% and there's a few common approaches 00:06:22.449 --> 00:06:25.785 align:center line:-1 position:50% size:35% that people use for creating these susceptibility maps. 00:06:25.785 --> 00:06:28.221 align:center line:-1 position:50% size:36% First one is a physically-based model, 00:06:28.221 --> 00:06:30.223 align:center line:-1 position:50% size:27% and these are, as the name implies, 00:06:30.223 --> 00:06:32.659 align:center line:-1 position:50% size:41% they're able to model the actual physical mechanisms 00:06:32.659 --> 00:06:35.328 align:center line:-1 position:50% size:32% that lead to slope failure. 00:06:35.328 --> 00:06:37.831 align:center line:-1 position:50% size:37% The next one is a data-driven or statistical models, 00:06:37.831 --> 00:06:40.100 align:center line:-1 position:50% size:35% and these include methods like machine learning, 00:06:40.100 --> 00:06:43.503 align:center line:-1 position:50% size:38% fuzzy logic, deep learning, AI, those types of methods, 00:06:43.503 --> 00:06:47.007 align:center line:-1 position:50% size:36% where really the output is dictated by the input data. 00:06:47.007 --> 00:06:49.276 align:center line:-1 position:50% size:30% And we'll dig more into exactly how these work 00:06:49.276 --> 00:06:51.911 align:center line:-1 position:50% size:34% because this is gonna be kinda the focus of this talk, 00:06:51.911 --> 00:06:55.115 align:center line:-1 position:50% size:27% are these data-driven statistical models. 00:06:55.115 --> 00:06:57.884 align:center line:-1 position:50% size:33% Next are heuristic models, and what these are, 00:06:57.884 --> 00:07:00.320 align:center line:-1 position:50% size:22% are a user would basically define 00:07:00.320 --> 00:07:02.022 align:center line:-1 position:50% size:35% what dictates susceptibility. 00:07:02.022 --> 00:07:04.858 align:center line:-1 position:50% size:27% So you can imagine somebody could say, 00:07:04.858 --> 00:07:07.260 align:center line:-1 position:50% size:30% "Okay, we're gonna say that susceptibility 00:07:07.260 --> 00:07:09.029 align:center line:-1 position:50% size:38% is gonna increase with slope." 00:07:09.029 --> 00:07:10.864 align:center line:-1 position:50% size:40% And so you could just measure the slope of the terrain, 00:07:10.864 --> 00:07:14.334 align:center line:-1 position:50% size:27% and voila, you'd have a susceptibility map. 00:07:14.334 --> 00:07:16.603 align:center line:-1 position:50% size:34% That would be an example of heuristic model. 00:07:16.603 --> 00:07:18.838 align:center line:-1 position:50% size:33% And then you can also do geomorphic mapping 00:07:18.838 --> 00:07:23.310 align:center line:-1 position:50% size:40% using GIS and field observation to create a susceptibility map. 00:07:23.310 --> 00:07:25.545 align:center line:-1 position:50% size:38% But as I said before, we're gonna be focusing here 00:07:25.545 --> 00:07:27.013 align:center line:-1 position:50% size:23% on data-driven statistical models, 00:07:27.013 --> 00:07:28.348 align:center line:-1 position:50% size:29% and the reason for that is because 00:07:28.348 --> 00:07:31.351 align:center line:-1 position:50% size:33% generally over large areas we don't--once again, 00:07:31.351 --> 00:07:34.120 align:center line:-1 position:50% size:41% we don't have the data available for a physically-based model. 00:07:34.120 --> 00:07:37.090 align:center line:-1 position:50% size:27% But thanks to GI-- or to remote sensing, 00:07:37.090 --> 00:07:38.758 align:center line:-1 position:50% size:27% that's really kinda gained traction 00:07:38.758 --> 00:07:40.327 align:center line:-1 position:50% size:30% over the last few years. 00:07:40.327 --> 00:07:43.163 align:center line:-1 position:50% size:31% We do have a lot of data that can be used 00:07:43.163 --> 00:07:45.432 align:center line:-1 position:50% size:23% with data-driven statistical models. 00:07:47.634 --> 00:07:49.602 align:center line:-1 position:50% size:40% So how do these models work? 00:07:49.602 --> 00:07:51.671 align:center line:-1 position:50% size:34% The basic idea is, you imagine that you have 00:07:51.671 --> 00:07:54.774 align:center line:-1 position:50% size:36% some domain of interest shown here in these figures, 00:07:54.774 --> 00:07:57.210 align:center line:-1 position:50% size:40% and you have some known distribution of landslides 00:07:57.210 --> 00:07:59.979 align:center line:-1 position:50% size:38% shown here by the white dots. 00:07:59.979 --> 00:08:01.348 align:center line:-1 position:50% size:35% And then you can measure the attributes 00:08:01.348 --> 00:08:02.615 align:center line:-1 position:50% size:25% of those landslides. 00:08:02.615 --> 00:08:04.317 align:center line:-1 position:50% size:28% Here on the left we have slope plotted 00:08:04.317 --> 00:08:06.453 align:center line:-1 position:50% size:24% and on the right we have elevation. 00:08:06.453 --> 00:08:08.355 align:center line:-1 position:50% size:30% And so you can gather the slope and elevation 00:08:08.355 --> 00:08:10.490 align:center line:-1 position:50% size:32% of all these different landslide points, 00:08:10.490 --> 00:08:13.026 align:center line:-1 position:50% size:38% and then you can collect data across the domain 00:08:13.026 --> 00:08:15.895 align:center line:-1 position:50% size:30% that doesn't have evidence of landsliding, 00:08:15.895 --> 00:08:17.597 align:center line:-1 position:50% size:30% and collect the slope and elevation 00:08:17.597 --> 00:08:20.467 align:center line:-1 position:50% size:33% of all those non-landsliding locations. 00:08:20.467 --> 00:08:22.569 align:center line:-1 position:50% size:32% So you have these two arrays of data. 00:08:22.569 --> 00:08:25.038 align:center line:-1 position:50% size:25% Landslide data, non-landslide data. 00:08:25.038 --> 00:08:27.707 align:center line:-1 position:50% size:35% You feed that into a model and that model is designed 00:08:27.707 --> 00:08:30.076 align:center line:-1 position:50% size:36% to differentiate as best it can 00:08:30.076 --> 00:08:33.613 align:center line:-1 position:50% size:36% the landslide data from the non-landslide data. 00:08:33.613 --> 00:08:35.815 align:center line:-1 position:50% size:22% And so hopefully it's obvious here 00:08:35.815 --> 00:08:37.884 align:center line:-1 position:50% size:38% the importance of data, right? 00:08:37.884 --> 00:08:40.019 align:center line:-1 position:50% size:32% This is gonna be kind of a common theme 00:08:40.019 --> 00:08:43.256 align:center line:-1 position:50% size:35% or point I'm gonna keep on driving home here. 00:08:43.256 --> 00:08:46.326 align:center line:-1 position:50% size:38% Data reigns supreme with these statistical methods. 00:08:46.326 --> 00:08:48.528 align:center line:-1 position:50% size:22% And so with that, 00:08:48.528 --> 00:08:50.296 align:center line:-1 position:50% size:34% we're running into these two major obstacles. 00:08:50.296 --> 00:08:52.665 align:center line:-1 position:50% size:37% How do we deal if we wanna create a susceptibility map 00:08:52.665 --> 00:08:54.334 align:center line:-1 position:50% size:32% where we might not have 00:08:54.334 --> 00:08:56.302 align:center line:-1 position:50% size:27% very much data or no landslide data? 00:08:56.302 --> 00:08:59.072 align:center line:-1 position:50% size:34% How can we create a susceptibility map there? 00:08:59.072 --> 00:09:02.242 align:center line:-1 position:50% size:32% And then how to manage the imprecise nature 00:09:02.242 --> 00:09:04.177 align:center line:-1 position:50% size:26% of the landslide data that's available to us 00:09:04.177 --> 00:09:06.913 align:center line:-1 position:50% size:26% to create these susceptibility maps? 00:09:06.913 --> 00:09:09.182 align:center line:-1 position:50% size:27% To illustrate this point we're gonna zoom in 00:09:09.182 --> 00:09:12.085 align:center line:-1 position:50% size:34% to the Four Corners region of the United States. 00:09:12.085 --> 00:09:14.487 align:center line:-1 position:50% size:28% Here I have an image 00:09:14.487 --> 00:09:18.258 align:center line:-1 position:50% size:40% from the National Inventory that USGS has been compiling, 00:09:18.258 --> 00:09:21.828 align:center line:-1 position:50% size:37% and the colors here correlate with the level of confidence 00:09:21.828 --> 00:09:23.596 align:center line:-1 position:50% size:30% in the extent and nature 00:09:23.596 --> 00:09:26.366 align:center line:-1 position:50% size:26% of the landslide data that we have. 00:09:26.366 --> 00:09:28.835 align:center line:-1 position:50% size:28% Lighter colors being low confidence, 00:09:28.835 --> 00:09:32.005 align:center line:-1 position:50% size:32% higher--or darker colors being higher confidence. 00:09:32.005 --> 00:09:34.841 align:center line:-1 position:50% size:38% So we're gonna zoom in here to Colorado, New Mexico, 00:09:34.841 --> 00:09:36.576 align:center line:-1 position:50% size:24% Arizona, and Utah. 00:09:36.576 --> 00:09:39.712 align:center line:-1 position:50% size:37% And you can see that there's a lot of heterogeneity here 00:09:39.712 --> 00:09:42.782 align:center line:-1 position:50% size:32% both in the extent of the mapped landslides 00:09:42.782 --> 00:09:46.019 align:center line:-1 position:50% size:22% and the quality of the landslides. 00:09:46.019 --> 00:09:47.620 align:center line:-1 position:50% size:37% And so you see here in Utah 00:09:47.620 --> 00:09:50.056 align:center line:-1 position:50% size:26% you have these Colorado landslides 00:09:50.056 --> 00:09:53.693 align:center line:-1 position:50% size:34% that starkly stop at the municipal boundary, 00:09:53.693 --> 00:09:56.496 align:center line:-1 position:50% size:17% which is odd. 00:09:56.496 --> 00:09:58.097 align:center line:-1 position:50% size:24% And then you have varying degrees 00:09:58.097 --> 00:10:00.033 align:center line:-1 position:50% size:27% of confidence in this. 00:10:00.033 --> 00:10:02.135 align:center line:-1 position:50% size:42% And so how do we manage this? 00:10:02.135 --> 00:10:03.570 align:center line:-1 position:50% size:28% The way we're gonna 00:10:03.570 --> 00:10:05.271 align:center line:-1 position:50% size:29% manage these regions of no landslide data 00:10:05.271 --> 00:10:08.074 align:center line:-1 position:50% size:37% is that we're gonna assess the limits of the training data. 00:10:08.074 --> 00:10:09.709 align:center line:-1 position:50% size:24% And I'll highlight these experiments 00:10:09.709 --> 00:10:11.077 align:center line:-1 position:50% size:26% in the coming slides, but that's gonna be 00:10:11.077 --> 00:10:13.046 align:center line:-1 position:50% size:30% the first half of this talk. 00:10:13.046 --> 00:10:16.149 align:center line:-1 position:50% size:43% The next one is, "How to manage imprecise landslide data?" 00:10:16.149 --> 00:10:18.751 align:center line:-1 position:50% size:30% By "imprecise"-- let me define this here-- 00:10:18.751 --> 00:10:20.887 align:center line:-1 position:50% size:43% we mean heterogeneous formats. 00:10:20.887 --> 00:10:23.189 align:center line:-1 position:50% size:32% So for instance here, the Four Corners region. 00:10:23.189 --> 00:10:26.125 align:center line:-1 position:50% size:34% New Mexico predominantly uses points. 00:10:26.125 --> 00:10:28.194 align:center line:-1 position:50% size:31% The other areas often use a combination 00:10:28.194 --> 00:10:30.396 align:center line:-1 position:50% size:30% of points and polygons. 00:10:30.396 --> 00:10:32.432 align:center line:-1 position:50% size:44% And so how do we manage those? 00:10:32.432 --> 00:10:35.201 align:center line:-1 position:50% size:39% Inaccurate landslide locations, as we already talked about. 00:10:35.201 --> 00:10:37.270 align:center line:-1 position:50% size:35% The colors here correlate with the level of confidence 00:10:37.270 --> 00:10:39.372 align:center line:-1 position:50% size:35% and the extent and location of these landslides. 00:10:39.372 --> 00:10:41.941 align:center line:-1 position:50% size:30% So how do we deal with that heterogeneity? 00:10:41.941 --> 00:10:43.243 align:center line:-1 position:50% size:40% There's no time unit component 00:10:43.243 --> 00:10:44.777 align:center line:-1 position:50% size:30% on basically any of these landslides 00:10:44.777 --> 00:10:46.045 align:center line:-1 position:50% size:30% that are available to us. 00:10:46.045 --> 00:10:48.314 align:center line:-1 position:50% size:28% These are all mapped using predominantly 00:10:48.314 --> 00:10:51.251 align:center line:-1 position:50% size:34% geomorphic mapping tools and GIS. 00:10:51.251 --> 00:10:52.819 align:center line:-1 position:50% size:39% And the level of completeness. 00:10:52.819 --> 00:10:56.589 align:center line:-1 position:50% size:34% We can see that there's relative differences 00:10:56.589 --> 00:10:58.258 align:center line:-1 position:50% size:36% in the level of completeness. 00:10:58.258 --> 00:11:00.393 align:center line:-1 position:50% size:35% And what really does-- what does a complete map 00:11:00.393 --> 00:11:02.195 align:center line:-1 position:50% size:20% really look like? 00:11:02.195 --> 00:11:04.864 align:center line:-1 position:50% size:27% These are all kinda inherent imprecisions 00:11:04.864 --> 00:11:06.766 align:center line:-1 position:50% size:32% that we need to deal with and take account for 00:11:06.766 --> 00:11:10.069 align:center line:-1 position:50% size:26% when we're creating a susceptibility map. 00:11:10.069 --> 00:11:14.240 align:center line:-1 position:50% size:42% And one way that we can kind of accommodate this imprecision 00:11:14.240 --> 00:11:15.742 align:center line:-1 position:50% size:21% is by the choice of mapping unit, 00:11:15.742 --> 00:11:19.345 align:center line:-1 position:50% size:27% and I'll dive into how that plays a role 00:11:19.345 --> 00:11:21.848 align:center line:-1 position:50% size:24% for the second half of this talk. 00:11:21.848 --> 00:11:23.716 align:center line:-1 position:50% size:30% So once again, we're first gonna tackle 00:11:23.716 --> 00:11:25.685 align:center line:-1 position:50% size:30% how to manage regions with limited 00:11:25.685 --> 00:11:28.221 align:center line:-1 position:50% size:31% or no landslide data, and then we'll talk about 00:11:28.221 --> 00:11:30.924 align:center line:-1 position:50% size:30% how to manage imprecision in the data. 00:11:30.924 --> 00:11:33.560 align:center line:-1 position:50% size:27% So to evaluate how to apply a model 00:11:33.560 --> 00:11:37.764 align:center line:-1 position:50% size:30% to areas with poor data, or limited or no data, 00:11:37.764 --> 00:11:39.933 align:center line:-1 position:50% size:28% we're gonna evaluate previous approaches 00:11:39.933 --> 00:11:43.403 align:center line:-1 position:50% size:33% that have been done for mapping susceptibility 00:11:43.403 --> 00:11:45.438 align:center line:-1 position:50% size:33% over areas with poor data. 00:11:45.438 --> 00:11:48.441 align:center line:-1 position:50% size:30% And having reviewed the literature quite a bit, 00:11:48.441 --> 00:11:52.745 align:center line:-1 position:50% size:38% I kinda broke 'em down into these three main approaches. 00:11:52.745 --> 00:11:54.380 align:center line:-1 position:50% size:25% The first one is applying a model 00:11:54.380 --> 00:11:57.483 align:center line:-1 position:50% size:35% trained on data-rich regions to data-poor regions. 00:11:57.483 --> 00:11:58.952 align:center line:-1 position:50% size:32% And the way we're gonna evaluate this approach 00:11:58.952 --> 00:12:00.820 align:center line:-1 position:50% size:29% is that we have four different data sets 00:12:00.820 --> 00:12:02.789 align:center line:-1 position:50% size:39% from across the United States, 00:12:02.789 --> 00:12:05.258 align:center line:-1 position:50% size:39% and we're gonna train a model on three of those four 00:12:05.258 --> 00:12:07.393 align:center line:-1 position:50% size:31% and apply that model and see how it performs 00:12:07.393 --> 00:12:09.128 align:center line:-1 position:50% size:24% on that one region that's left out, 00:12:09.128 --> 00:12:10.964 align:center line:-1 position:50% size:36% and this is basically to mimic 00:12:10.964 --> 00:12:14.233 align:center line:-1 position:50% size:27% what most people do 00:12:14.233 --> 00:12:16.035 align:center line:-1 position:50% size:24% with this approach. 00:12:16.035 --> 00:12:17.537 align:center line:-1 position:50% size:32% We're gonna do this for each of the four areas 00:12:17.537 --> 00:12:20.573 align:center line:-1 position:50% size:38% so we can get kind of a range of the relative performance 00:12:20.573 --> 00:12:22.942 align:center line:-1 position:50% size:19% of this method. 00:12:22.942 --> 00:12:25.178 align:center line:-1 position:50% size:38% The second approach that I've seen commonly used 00:12:25.178 --> 00:12:27.847 align:center line:-1 position:50% size:38% is that they restrict model training and application 00:12:27.847 --> 00:12:31.017 align:center line:-1 position:50% size:41% to regions with shared environmental attributes. 00:12:31.017 --> 00:12:34.921 align:center line:-1 position:50% size:43% And so we selected these regions based off of one, the quality 00:12:34.921 --> 00:12:36.556 align:center line:-1 position:50% size:26% of the landslide data that's available, 00:12:36.556 --> 00:12:39.926 align:center line:-1 position:50% size:26% but also on the relative similarity 00:12:39.926 --> 00:12:44.063 align:center line:-1 position:50% size:34% in environmental attributes on the continental scale. 00:12:44.063 --> 00:12:46.566 align:center line:-1 position:50% size:30% So two of them share physiographic province, 00:12:46.566 --> 00:12:48.968 align:center line:-1 position:50% size:39% and one of them is outside of that physiographic province, 00:12:48.968 --> 00:12:51.771 align:center line:-1 position:50% size:39% but shares a similar ecoregion, 00:12:51.771 --> 00:12:54.340 align:center line:-1 position:50% size:27% and then one region is completely outside 00:12:54.340 --> 00:12:57.977 align:center line:-1 position:50% size:33% of those environmental conditions. 00:12:57.977 --> 00:13:01.180 align:center line:-1 position:50% size:33% And so we'll train a model on one location 00:13:01.180 --> 00:13:03.549 align:center line:-1 position:50% size:36% that should have high environmental similarity 00:13:03.549 --> 00:13:05.051 align:center line:-1 position:50% size:24% and tested on the other locations 00:13:05.051 --> 00:13:08.721 align:center line:-1 position:50% size:39% to see if constraining by these environmental attributes 00:13:08.721 --> 00:13:11.424 align:center line:-1 position:50% size:44% helps improve model performance. 00:13:11.424 --> 00:13:13.192 align:center line:-1 position:50% size:34% So we're gonna do that with these varying degrees 00:13:13.192 --> 00:13:15.294 align:center line:-1 position:50% size:34% of environmental similarity. 00:13:15.294 --> 00:13:17.296 align:center line:-1 position:50% size:23% The last approach 00:13:17.296 --> 00:13:18.798 align:center line:-1 position:50% size:41% is training and applying a model 00:13:18.798 --> 00:13:21.668 align:center line:-1 position:50% size:30% on very sparse, but a uniform inventory. 00:13:21.668 --> 00:13:24.370 align:center line:-1 position:50% size:35% And so what I mean by this is that you have 00:13:24.370 --> 00:13:27.340 align:center line:-1 position:50% size:28% some landslide data across the entire area 00:13:27.340 --> 00:13:29.008 align:center line:-1 position:50% size:26% where you wanna model susceptibility, 00:13:29.008 --> 00:13:31.444 align:center line:-1 position:50% size:35% but it could be very sparse. 00:13:31.444 --> 00:13:36.382 align:center line:-1 position:50% size:42% And often the way people do this is that they'll kinda scour GIS 00:13:36.382 --> 00:13:38.184 align:center line:-1 position:50% size:36% and just find some locations 00:13:38.184 --> 00:13:40.853 align:center line:-1 position:50% size:39% that they know don't have a mapped landslide 00:13:40.853 --> 00:13:43.523 align:center line:-1 position:50% size:26% or do have evidence of landsliding. 00:13:43.523 --> 00:13:48.428 align:center line:-1 position:50% size:38% But they're not worried about getting a complete repository. 00:13:48.428 --> 00:13:50.196 align:center line:-1 position:50% size:28% And to mimic this, what we're gonna do, 00:13:50.196 --> 00:13:52.298 align:center line:-1 position:50% size:32% we're gonna use these same four regions, 00:13:52.298 --> 00:13:55.234 align:center line:-1 position:50% size:36% but only sample five percent of the available data, 00:13:55.234 --> 00:13:57.637 align:center line:-1 position:50% size:27% develop a model with that five percent, 00:13:57.637 --> 00:14:00.006 align:center line:-1 position:50% size:31% and then test that model on the 95 percent 00:14:00.006 --> 00:14:02.008 align:center line:-1 position:50% size:28% we didn't use to develop the model, 00:14:02.008 --> 00:14:03.309 align:center line:-1 position:50% size:26% and we'll repeat this several times 00:14:03.309 --> 00:14:05.344 align:center line:-1 position:50% size:35% so we can get distributions. 00:14:05.344 --> 00:14:07.914 align:center line:-1 position:50% size:29% And so now we kind of understand the theory 00:14:07.914 --> 00:14:09.348 align:center line:-1 position:50% size:30% of how we're gonna approach this problem. 00:14:09.348 --> 00:14:12.185 align:center line:-1 position:50% size:37% Here are the actual data sets that we're looking at here. 00:14:12.185 --> 00:14:17.023 align:center line:-1 position:50% size:35% So we have three locations in Appalachia 00:14:17.023 --> 00:14:19.392 align:center line:-1 position:50% size:35% we have Magoffin County, Kentucky, 00:14:19.392 --> 00:14:20.827 align:center line:-1 position:50% size:42% Doddridge County, West Virginia, 00:14:20.827 --> 00:14:23.730 align:center line:-1 position:50% size:42% which both share the same physiographic province 00:14:23.730 --> 00:14:25.465 align:center line:-1 position:50% size:31% and level two ecoregion. 00:14:25.465 --> 00:14:27.066 align:center line:-1 position:50% size:29% Ecoregions are shown by the colors. 00:14:27.066 --> 00:14:30.269 align:center line:-1 position:50% size:37% The physiographic provinces are shown by outlines. 00:14:30.269 --> 00:14:31.704 align:center line:-1 position:50% size:29% And then we also have Macon County, 00:14:31.704 --> 00:14:34.307 align:center line:-1 position:50% size:30% which is in a different physiographic province, 00:14:34.307 --> 00:14:36.943 align:center line:-1 position:50% size:27% but within the same level two eco region. 00:14:36.943 --> 00:14:39.245 align:center line:-1 position:50% size:38% And then we also have the Elkhorn Ridge Wilderness 00:14:39.245 --> 00:14:40.613 align:center line:-1 position:50% size:16% in California. 00:14:40.613 --> 00:14:42.648 align:center line:-1 position:50% size:40% And so these are the data sets, and once again, 00:14:42.648 --> 00:14:44.751 align:center line:-1 position:50% size:37% we chose these areas due to 00:14:44.751 --> 00:14:46.953 align:center line:-1 position:50% size:33% the variation in environmental similarity 00:14:46.953 --> 00:14:49.088 align:center line:-1 position:50% size:25% and the high quality landslide data 00:14:49.088 --> 00:14:50.389 align:center line:-1 position:50% size:27% that's available here. 00:14:50.389 --> 00:14:51.724 align:center line:-1 position:50% size:37% And we know this because of 00:14:51.724 --> 00:14:55.762 align:center line:-1 position:50% size:39% the high resolution LiDAR imagery that they used, 00:14:55.762 --> 00:14:58.664 align:center line:-1 position:50% size:39% including field reconnaissance for developing these catalogs 00:14:58.664 --> 00:15:00.867 align:center line:-1 position:50% size:28% that we're gonna use. 00:15:00.867 --> 00:15:04.070 align:center line:-1 position:50% size:29% All right, so the models that we're gonna use, 00:15:04.070 --> 00:15:05.938 align:center line:-1 position:50% size:27% the statistical models that we're gonna use 00:15:05.938 --> 00:15:07.573 align:center line:-1 position:50% size:27% is logistic regression. 00:15:07.573 --> 00:15:09.675 align:center line:-1 position:50% size:29% And we're gonna evaluate these models 00:15:09.675 --> 00:15:11.177 align:center line:-1 position:50% size:38% using two primary techniques. 00:15:11.177 --> 00:15:13.179 align:center line:-1 position:50% size:33% The first, commonly used, 00:15:13.179 --> 00:15:14.781 align:center line:-1 position:50% size:22% receiver operator characteristics, 00:15:14.781 --> 00:15:16.415 align:center line:-1 position:50% size:28% area under the curve. 00:15:16.415 --> 00:15:18.151 align:center line:-1 position:50% size:30% Kind of the go-to metric 00:15:18.151 --> 00:15:21.988 align:center line:-1 position:50% size:29% for assessing landslide susceptibility, 00:15:21.988 --> 00:15:24.090 align:center line:-1 position:50% size:31% or how well the models-- 00:15:24.090 --> 00:15:26.058 align:center line:-1 position:50% size:24% But something that's often lacking 00:15:26.058 --> 00:15:29.695 align:center line:-1 position:50% size:27% when we evaluate susceptibility models 00:15:29.695 --> 00:15:31.731 align:center line:-1 position:50% size:32% is actually understanding what leads to 00:15:31.731 --> 00:15:33.733 align:center line:-1 position:50% size:28% that level of accuracy, 00:15:33.733 --> 00:15:37.436 align:center line:-1 position:50% size:34% as reflected in the receiver operator characteristics. 00:15:37.436 --> 00:15:39.705 align:center line:-1 position:50% size:30% And we're gonna dive into understanding 00:15:39.705 --> 00:15:41.941 align:center line:-1 position:50% size:27% this relative accuracy by comparing 00:15:41.941 --> 00:15:46.412 align:center line:-1 position:50% size:29% the logistic coefficients within the models. 00:15:46.412 --> 00:15:48.247 align:center line:-1 position:50% size:29% And we'll get more into what that means 00:15:48.247 --> 00:15:50.516 align:center line:-1 position:50% size:26% in the coming slides. 00:15:50.516 --> 00:15:52.785 align:center line:-1 position:50% size:30% It should be noted here that we're not using 00:15:52.785 --> 00:15:55.988 align:center line:-1 position:50% size:33% just a basic frequentist logistic regression model, 00:15:55.988 --> 00:15:57.557 align:center line:-1 position:50% size:43% we're using a Bayesian approach. 00:15:57.557 --> 00:15:59.392 align:center line:-1 position:50% size:27% And the reason we're (audio drops out) 00:15:59.392 --> 00:16:00.793 align:center line:-1 position:50% size:36% is that a Bayesian approach is able 00:16:00.793 --> 00:16:03.629 align:center line:-1 position:50% size:35% to incorporate uncertainty in the estimated distribution 00:16:03.629 --> 00:16:05.598 align:center line:-1 position:50% size:32% of the model parameters. 00:16:05.598 --> 00:16:08.601 align:center line:-1 position:50% size:37% In this case, we're interested in distribution of beta. 00:16:08.601 --> 00:16:11.370 align:center line:-1 position:50% size:40% And a logistic regression model is shown here, 00:16:11.370 --> 00:16:13.172 align:center line:-1 position:50% size:42% where you have a linear function 00:16:13.172 --> 00:16:16.576 align:center line:-1 position:50% size:23% that's within a logistic function. 00:16:16.576 --> 00:16:18.878 align:center line:-1 position:50% size:40% And that's basically-- and then it outputs the log odds 00:16:18.878 --> 00:16:22.481 align:center line:-1 position:50% size:32% of there being a landslide at a given location. 00:16:22.481 --> 00:16:25.318 align:center line:-1 position:50% size:33% And the basis of the Bayesian approach 00:16:25.318 --> 00:16:26.853 align:center line:-1 position:50% size:27% for logistic regression 00:16:26.853 --> 00:16:29.155 align:center line:-1 position:50% size:25% is what's referred to as Bayes' Law, 00:16:29.155 --> 00:16:30.890 align:center line:-1 position:50% size:31% and what Bayes' Law is, 00:16:30.890 --> 00:16:35.127 align:center line:-1 position:50% size:39% is as a posterior distribution of some parameter of interest, 00:16:35.127 --> 00:16:39.298 align:center line:-1 position:50% size:31% in this case it's the beta, given your input data, 00:16:39.298 --> 00:16:42.435 align:center line:-1 position:50% size:23% is proportional to prior probability 00:16:42.435 --> 00:16:43.736 align:center line:-1 position:50% size:25% times the likelihood of function. 00:16:43.736 --> 00:16:47.173 align:center line:-1 position:50% size:40% The prior probability is an assumed distribution of beta 00:16:47.173 --> 00:16:49.575 align:center line:-1 position:50% size:27% before you even look at the data. 00:16:49.575 --> 00:16:51.644 align:center line:-1 position:50% size:34% And the likelihood function is assumed 00:16:51.644 --> 00:16:52.945 align:center line:-1 position:50% size:37% to be a Bernoulli distribution, 00:16:52.945 --> 00:16:56.382 align:center line:-1 position:50% size:30% another logistic function that's shown here. 00:16:56.382 --> 00:16:58.284 align:center line:-1 position:50% size:27% So for our instances, 00:16:58.284 --> 00:17:00.319 align:center line:-1 position:50% size:23% we're gonna use very vague priors, 00:17:00.319 --> 00:17:02.421 align:center line:-1 position:50% size:31% so that way the data is really what dominates 00:17:02.421 --> 00:17:04.624 align:center line:-1 position:50% size:36% these posterior distributions. 00:17:04.624 --> 00:17:06.158 align:center line:-1 position:50% size:26% And so this gives us a distribution 00:17:06.158 --> 00:17:07.727 align:center line:-1 position:50% size:27% of these coefficients, and that allows us 00:17:07.727 --> 00:17:10.229 align:center line:-1 position:50% size:25% to do more statistical tinkering, 00:17:10.229 --> 00:17:11.731 align:center line:-1 position:50% size:33% as we'll show you, than you can generally do 00:17:11.731 --> 00:17:15.668 align:center line:-1 position:50% size:36% with a frequentist approach. 00:17:15.668 --> 00:17:17.937 align:center line:-1 position:50% size:38% So for preprocessing the data 00:17:17.937 --> 00:17:20.406 align:center line:-1 position:50% size:33% there's a lot that I'm not gonna go into, 00:17:20.406 --> 00:17:22.775 align:center line:-1 position:50% size:39% but we follow kinda the standard procedures 00:17:22.775 --> 00:17:26.345 align:center line:-1 position:50% size:31% that people have had success with in the past. 00:17:26.345 --> 00:17:29.282 align:center line:-1 position:50% size:37% And we ended up evaluating 64 different predictors, 00:17:29.282 --> 00:17:31.651 align:center line:-1 position:50% size:37% and by "predictors" I mean aspects of the terrain, 00:17:31.651 --> 00:17:34.553 align:center line:-1 position:50% size:29% like slope or elevation. 00:17:34.553 --> 00:17:36.689 align:center line:-1 position:50% size:23% And after we did our preprocessing 00:17:36.689 --> 00:17:38.858 align:center line:-1 position:50% size:34% and a correlation analysis, we reduced that number 00:17:38.858 --> 00:17:40.626 align:center line:-1 position:50% size:15% down to 11. 00:17:42.161 --> 00:17:44.363 align:center line:-1 position:50% size:34% And so that's the model approach. 00:17:44.363 --> 00:17:45.765 align:center line:-1 position:50% size:39% How are we evaluating these? 00:17:45.765 --> 00:17:47.900 align:center line:-1 position:50% size:35% So if you're not familiar with 00:17:47.900 --> 00:17:49.468 align:center line:-1 position:50% size:22% receiver operator characteristics, 00:17:49.468 --> 00:17:51.871 align:center line:-1 position:50% size:28% area under the curve, the basic idea is that 00:17:51.871 --> 00:17:53.806 align:center line:-1 position:50% size:30% it's a measure of how well your model 00:17:53.806 --> 00:17:57.543 align:center line:-1 position:50% size:38% is able to distinguish no landslides from landslides. 00:17:57.543 --> 00:17:58.945 align:center line:-1 position:50% size:30% So if you have a model that's able 00:17:58.945 --> 00:18:00.846 align:center line:-1 position:50% size:29% to perfectly distinguish no landslides, 00:18:00.846 --> 00:18:02.949 align:center line:-1 position:50% size:25% shown here in blue, from landslides, 00:18:02.949 --> 00:18:04.283 align:center line:-1 position:50% size:29% shown here in orange, you would get 00:18:04.283 --> 00:18:05.651 align:center line:-1 position:50% size:39% an area-under-the-curve value of one, 00:18:05.651 --> 00:18:08.821 align:center line:-1 position:50% size:28% one being perfect differentiation. 00:18:08.821 --> 00:18:10.957 align:center line:-1 position:50% size:33% We contrast if your model does a very poor job 00:18:10.957 --> 00:18:12.525 align:center line:-1 position:50% size:22% of differentiating, 00:18:12.525 --> 00:18:14.060 align:center line:-1 position:50% size:42% such as the example down here, 00:18:14.060 --> 00:18:17.296 align:center line:-1 position:50% size:24% where it basically randomly chooses. 00:18:17.296 --> 00:18:18.998 align:center line:-1 position:50% size:28% It's equivalent to randomly choosing 00:18:18.998 --> 00:18:20.866 align:center line:-1 position:50% size:25% between landslides and no landslides. 00:18:20.866 --> 00:18:22.868 align:center line:-1 position:50% size:39% You would get an area-under-the-curve value 00:18:22.868 --> 00:18:24.203 align:center line:-1 position:50% size:8% of 0.5, 00:18:24.203 --> 00:18:26.706 align:center line:-1 position:50% size:29% so basically equivalent to randomly guessing. 00:18:26.706 --> 00:18:27.940 align:center line:-1 position:50% size:36% So you want area-under-the-curve values 00:18:27.940 --> 00:18:29.976 align:center line:-1 position:50% size:22% of near 1, ideally, 00:18:29.976 --> 00:18:33.245 align:center line:-1 position:50% size:36% closer to 1, further from 0.5. 00:18:33.245 --> 00:18:34.880 align:center line:-1 position:50% size:42% And once again, the second way 00:18:34.880 --> 00:18:36.749 align:center line:-1 position:50% size:29% we're gonna evaluate these models 00:18:36.749 --> 00:18:39.952 align:center line:-1 position:50% size:36% to understand these area-under-the-curve values 00:18:39.952 --> 00:18:41.887 align:center line:-1 position:50% size:22% are by assessing the coefficients 00:18:41.887 --> 00:18:44.023 align:center line:-1 position:50% size:21% of these models. 00:18:44.023 --> 00:18:46.325 align:center line:-1 position:50% size:32% But because of some inherent properties 00:18:46.325 --> 00:18:48.227 align:center line:-1 position:50% size:32% within logistic regression, 00:18:48.227 --> 00:18:50.730 align:center line:-1 position:50% size:36% we actually can't directly compare coefficients 00:18:50.730 --> 00:18:53.999 align:center line:-1 position:50% size:37% between the different models that we've trained, 00:18:53.999 --> 00:18:56.268 align:center line:-1 position:50% size:42% and so we have to convert these into something called 00:18:56.268 --> 00:18:58.571 align:center line:-1 position:50% size:32% average marginal effects, but all these are, 00:18:58.571 --> 00:19:02.008 align:center line:-1 position:50% size:31% it's a way of assessing the effect of the variable 00:19:02.008 --> 00:19:04.710 align:center line:-1 position:50% size:30% or predictor of interest within the given model. 00:19:04.710 --> 00:19:07.346 align:center line:-1 position:50% size:28% So an example of this is shown here. 00:19:07.346 --> 00:19:10.383 align:center line:-1 position:50% size:39% These different colors correlate with a different model 00:19:10.383 --> 00:19:11.951 align:center line:-1 position:50% size:31% or different training data that was used 00:19:11.951 --> 00:19:13.819 align:center line:-1 position:50% size:29% to develop that model. 00:19:13.819 --> 00:19:15.921 align:center line:-1 position:50% size:42% And we have our predictor slope, 00:19:15.921 --> 00:19:17.390 align:center line:-1 position:50% size:29% and we have these predictor effects, 00:19:17.390 --> 00:19:21.160 align:center line:-1 position:50% size:35% or average marginal effects on the X-axis. 00:19:21.160 --> 00:19:23.863 align:center line:-1 position:50% size:29% And what you can see is that there's variation 00:19:23.863 --> 00:19:28.167 align:center line:-1 position:50% size:38% in the sine and the magnitude of these predictor effects. 00:19:28.167 --> 00:19:30.403 align:center line:-1 position:50% size:36% And so if we take, for instance, Macon County, 00:19:30.403 --> 00:19:34.206 align:center line:-1 position:50% size:29% you can see it has a negative AME value, 00:19:34.206 --> 00:19:37.109 align:center line:-1 position:50% size:36% and what that tells you is that as you increase slope 00:19:37.109 --> 00:19:40.379 align:center line:-1 position:50% size:34% at that area, the model's gonna cause susceptibility 00:19:40.379 --> 00:19:42.148 align:center line:-1 position:50% size:26% to actually go down. 00:19:42.148 --> 00:19:44.417 align:center line:-1 position:50% size:32% So that's what the negative sine means. 00:19:44.417 --> 00:19:46.218 align:center line:-1 position:50% size:27% In contrast to all the other locations 00:19:46.218 --> 00:19:48.788 align:center line:-1 position:50% size:30% or all the other models, as you increase slope, 00:19:48.788 --> 00:19:50.723 align:center line:-1 position:50% size:25% you get an increase in susceptibility 00:19:50.723 --> 00:19:53.659 align:center line:-1 position:50% size:32% or in probability, which the model outputs. 00:19:53.659 --> 00:19:56.862 align:center line:-1 position:50% size:34% But the magnitude of that probability increase 00:19:56.862 --> 00:19:59.331 align:center line:-1 position:50% size:37% is variable depending on the magnitude 00:19:59.331 --> 00:20:01.734 align:center line:-1 position:50% size:39% of this average marginal effect. 00:20:01.734 --> 00:20:05.204 align:center line:-1 position:50% size:33% So how is this useful to us for understanding 00:20:05.204 --> 00:20:06.539 align:center line:-1 position:50% size:24% a model accuracy? 00:20:06.539 --> 00:20:10.776 align:center line:-1 position:50% size:42% Imagine that we develop a model only using slope, 00:20:10.776 --> 00:20:13.179 align:center line:-1 position:50% size:26% and we only have Macon County data, 00:20:13.179 --> 00:20:15.181 align:center line:-1 position:50% size:41% and we're gonna train our model on Macon County, 00:20:15.181 --> 00:20:18.484 align:center line:-1 position:50% size:37% and we end up with this negative predictor effect, 00:20:18.484 --> 00:20:21.187 align:center line:-1 position:50% size:28% negative average- marginal-effect value. 00:20:21.187 --> 00:20:22.555 align:center line:-1 position:50% size:25% And then we try to apply that model 00:20:22.555 --> 00:20:24.457 align:center line:-1 position:50% size:40% to any of these other data sets. 00:20:24.457 --> 00:20:28.861 align:center line:-1 position:50% size:38% You're gonna get a horrible fit 00:20:28.861 --> 00:20:30.596 align:center line:-1 position:50% size:36% for these areas where you apply that model, 00:20:30.596 --> 00:20:32.565 align:center line:-1 position:50% size:35% because it's gonna actually do the opposite 00:20:32.565 --> 00:20:34.900 align:center line:-1 position:50% size:34% of what it should be for assessing susceptibility 00:20:34.900 --> 00:20:36.836 align:center line:-1 position:50% size:23% in those locations. 00:20:36.836 --> 00:20:39.738 align:center line:-1 position:50% size:39% So in this way, these average marginal effects 00:20:39.738 --> 00:20:41.974 align:center line:-1 position:50% size:30% can give us information as to whether or not 00:20:41.974 --> 00:20:44.577 align:center line:-1 position:50% size:40% these different models are compatible with each other, 00:20:44.577 --> 00:20:46.278 align:center line:-1 position:50% size:18% or better said, 00:20:46.278 --> 00:20:48.647 align:center line:-1 position:50% size:43% the location where these models were trained 00:20:48.647 --> 00:20:50.516 align:center line:-1 position:50% size:42% are compatible with one another. 00:20:51.884 --> 00:20:53.919 align:center line:-1 position:50% size:38% And so using these two tools, 00:20:53.919 --> 00:20:57.890 align:center line:-1 position:50% size:33% average marginal effects and area under the curve, 00:20:57.890 --> 00:21:00.993 align:center line:-1 position:50% size:38% we're gonna assess these three different methods. 00:21:00.993 --> 00:21:04.430 align:center line:-1 position:50% size:34% So first is applying models 00:21:04.430 --> 00:21:06.732 align:center line:-1 position:50% size:28% from data-rich regions to data-poor, 00:21:06.732 --> 00:21:08.334 align:center line:-1 position:50% size:25% and then also we're gonna look at 00:21:08.334 --> 00:21:10.936 align:center line:-1 position:50% size:32% restricting by environmental attributes. 00:21:10.936 --> 00:21:12.471 align:center line:-1 position:50% size:27% And so the next plot, I'm gonna show you 00:21:12.471 --> 00:21:15.407 align:center line:-1 position:50% size:32% the area-under-the-curve distributions 00:21:15.407 --> 00:21:16.909 align:center line:-1 position:50% size:40% for these different experiments. 00:21:16.909 --> 00:21:18.277 align:center line:-1 position:50% size:38% So there's a lot going on here. 00:21:18.277 --> 00:21:20.713 align:center line:-1 position:50% size:29% We're gonna break this figure down. 00:21:20.713 --> 00:21:22.615 align:center line:-1 position:50% size:23% All right, so first up here at the top 00:21:22.615 --> 00:21:24.817 align:center line:-1 position:50% size:40% we have what we're referring to as resubstitution, 00:21:24.817 --> 00:21:27.019 align:center line:-1 position:50% size:29% that is the training data 00:21:27.019 --> 00:21:30.623 align:center line:-1 position:50% size:33% is then evaluated on itself. 00:21:30.623 --> 00:21:32.091 align:center line:-1 position:50% size:30% And so this you would expect to get 00:21:32.091 --> 00:21:34.360 align:center line:-1 position:50% size:37% relatively high area-under-the-curve values, 00:21:34.360 --> 00:21:37.596 align:center line:-1 position:50% size:33% and we do indeed get area under the curves 00:21:37.596 --> 00:21:40.199 align:center line:-1 position:50% size:15% around 0.7. 00:21:40.199 --> 00:21:42.067 align:center line:-1 position:50% size:35% Lowest value is about 0.65. 00:21:42.067 --> 00:21:45.137 align:center line:-1 position:50% size:41% But relatively high values compared to the other examples 00:21:45.137 --> 00:21:47.072 align:center line:-1 position:50% size:27% I'm gonna show you. 00:21:47.072 --> 00:21:49.875 align:center line:-1 position:50% size:42% Down here is where we evaluate the first quest-- 00:21:49.875 --> 00:21:51.777 align:center line:-1 position:50% size:30% the first method that we're interested in. 00:21:51.777 --> 00:21:53.812 align:center line:-1 position:50% size:24% Training model on data-rich areas, 00:21:53.812 --> 00:21:56.482 align:center line:-1 position:50% size:26% and then applying to a data-poor area. 00:21:56.482 --> 00:21:59.084 align:center line:-1 position:50% size:28% And these are models that were trained 00:21:59.084 --> 00:22:02.021 align:center line:-1 position:50% size:34% on all the data, except for a given location, 00:22:02.021 --> 00:22:04.924 align:center line:-1 position:50% size:32% so all the data except for the Magoffin location. 00:22:04.924 --> 00:22:06.992 align:center line:-1 position:50% size:36% And then we test that model on that one area 00:22:06.992 --> 00:22:09.562 align:center line:-1 position:50% size:33% that was left out, in this case the Magoffins. 00:22:09.562 --> 00:22:10.996 align:center line:-1 position:50% size:37% You can see here that we get 00:22:10.996 --> 00:22:14.033 align:center line:-1 position:50% size:37% quite a bit lower area-under-the-curve values. 00:22:14.033 --> 00:22:16.969 align:center line:-1 position:50% size:36% So this suggests that maybe this isn't the best approach. 00:22:16.969 --> 00:22:19.338 align:center line:-1 position:50% size:39% Okay, how about constricting-- or restricting 00:22:19.338 --> 00:22:21.840 align:center line:-1 position:50% size:33% based off of environmental attributes? 00:22:21.840 --> 00:22:23.776 align:center line:-1 position:50% size:28% That's shown up here. 00:22:23.776 --> 00:22:26.679 align:center line:-1 position:50% size:35% So here's the two data sets that have shared ecoregion 00:22:26.679 --> 00:22:28.714 align:center line:-1 position:50% size:37% and physiographic provinces. 00:22:28.714 --> 00:22:30.149 align:center line:-1 position:50% size:42% And once again, the highest area under the curve 00:22:30.149 --> 00:22:32.551 align:center line:-1 position:50% size:33% we're getting is about 0.6, 00:22:32.551 --> 00:22:34.887 align:center line:-1 position:50% size:30% but we're also getting another example of this 00:22:34.887 --> 00:22:38.824 align:center line:-1 position:50% size:42% below 0.5, so very poor model performance. 00:22:38.824 --> 00:22:40.993 align:center line:-1 position:50% size:40% Shared ecoregion and different physiographic province, 00:22:40.993 --> 00:22:42.494 align:center line:-1 position:50% size:28% once again, we're getting very low 00:22:42.494 --> 00:22:43.996 align:center line:-1 position:50% size:37% area-under-the-curve values. 00:22:43.996 --> 00:22:46.198 align:center line:-1 position:50% size:35% And really these--restricting by these environments 00:22:46.198 --> 00:22:48.100 align:center line:-1 position:50% size:43% are no better than when you have different ecoregion 00:22:48.100 --> 00:22:49.468 align:center line:-1 position:50% size:37% and physiographic provinces, 00:22:49.468 --> 00:22:52.204 align:center line:-1 position:50% size:29% and those distributions are shown down here. 00:22:52.204 --> 00:22:55.007 align:center line:-1 position:50% size:37% And so it appears that neither of these approaches, 00:22:55.007 --> 00:22:56.842 align:center line:-1 position:50% size:28% training on data-rich, applying to data-poor, 00:22:56.842 --> 00:22:59.245 align:center line:-1 position:50% size:33% or restricting by environmental similarities, 00:22:59.245 --> 00:23:01.113 align:center line:-1 position:50% size:32% are effective in this case. 00:23:01.113 --> 00:23:02.281 align:center line:-1 position:50% size:18% And so, why? 00:23:02.281 --> 00:23:04.250 align:center line:-1 position:50% size:27% What's going on here with this data 00:23:04.250 --> 00:23:05.517 align:center line:-1 position:50% size:29% and with these models that cause them 00:23:05.517 --> 00:23:09.888 align:center line:-1 position:50% size:33% not to be transferable between these locations? 00:23:09.888 --> 00:23:12.558 align:center line:-1 position:50% size:39% Average marginal effects are gonna give us information. 00:23:12.558 --> 00:23:15.995 align:center line:-1 position:50% size:24% So here hopefully the slope segment 00:23:15.995 --> 00:23:17.830 align:center line:-1 position:50% size:22% is familiar to you, and once again, 00:23:17.830 --> 00:23:19.898 align:center line:-1 position:50% size:24% if we want a model to perform well 00:23:19.898 --> 00:23:21.767 align:center line:-1 position:50% size:30% in a different location, you would expect these 00:23:21.767 --> 00:23:25.771 align:center line:-1 position:50% size:36% to have similar average- marginal-effect distributions. 00:23:25.771 --> 00:23:28.507 align:center line:-1 position:50% size:35% And across the parameter-- or the predictors 00:23:28.507 --> 00:23:30.109 align:center line:-1 position:50% size:22% that we include in these models-- 00:23:30.109 --> 00:23:32.011 align:center line:-1 position:50% size:33% here's a subset of those-- 00:23:32.011 --> 00:23:34.013 align:center line:-1 position:50% size:29% you can see that there's no consistency 00:23:34.013 --> 00:23:35.581 align:center line:-1 position:50% size:22% in sine generally, 00:23:35.581 --> 00:23:37.883 align:center line:-1 position:50% size:41% and definitely no consistency in the magnitude 00:23:37.883 --> 00:23:39.852 align:center line:-1 position:50% size:32% of these average marginal effects. 00:23:39.852 --> 00:23:41.754 align:center line:-1 position:50% size:31% And if we look at the rest of the predictors 00:23:41.754 --> 00:23:44.189 align:center line:-1 position:50% size:28% included in the model, it's the same story. 00:23:44.189 --> 00:23:46.725 align:center line:-1 position:50% size:40% And so because you don't have any consistency 00:23:46.725 --> 00:23:48.627 align:center line:-1 position:50% size:32% in these average marginal effects, 00:23:48.627 --> 00:23:51.297 align:center line:-1 position:50% size:39% a model trained in one location isn't gonna perform well 00:23:51.297 --> 00:23:52.931 align:center line:-1 position:50% size:28% in the other locations. 00:23:55.401 --> 00:23:57.803 align:center line:-1 position:50% size:36% And so the poor area-under-the-curve values 00:23:57.803 --> 00:23:59.505 align:center line:-1 position:50% size:36% that we're getting are due to 00:23:59.505 --> 00:24:01.740 align:center line:-1 position:50% size:22% these divergent predictor effects, 00:24:01.740 --> 00:24:03.075 align:center line:-1 position:50% size:19% as related by-- 00:24:03.075 --> 00:24:07.980 align:center line:-1 position:50% size:37% as manifest by the average marginal effects. 00:24:07.980 --> 00:24:09.948 align:center line:-1 position:50% size:36% How about this last method, 00:24:09.948 --> 00:24:11.216 align:center line:-1 position:50% size:37% training and applying models 00:24:11.216 --> 00:24:13.886 align:center line:-1 position:50% size:28% on very sparse but uniform inventory. 00:24:13.886 --> 00:24:16.755 align:center line:-1 position:50% size:32% Again, here, we're only sampling 5% of the data, 00:24:16.755 --> 00:24:19.024 align:center line:-1 position:50% size:25% developing a model with that 5%, 00:24:19.024 --> 00:24:22.161 align:center line:-1 position:50% size:28% and applying it to the rest of the data. 00:24:22.161 --> 00:24:24.129 align:center line:-1 position:50% size:28% Here's the distribution of those areas 00:24:24.129 --> 00:24:26.031 align:center line:-1 position:50% size:27% and the curve values we're getting. 00:24:26.031 --> 00:24:28.701 align:center line:-1 position:50% size:36% For reference, when we use all the available data, 00:24:28.701 --> 00:24:30.936 align:center line:-1 position:50% size:35% 100% of the available data, 00:24:30.936 --> 00:24:33.872 align:center line:-1 position:50% size:40% we're getting an area under the curve values of 0.65. 00:24:33.872 --> 00:24:35.574 align:center line:-1 position:50% size:34% But when we only use 5%, 00:24:35.574 --> 00:24:38.510 align:center line:-1 position:50% size:38% on average, we get an area under the curve value of 0.63. 00:24:38.510 --> 00:24:42.681 align:center line:-1 position:50% size:26% The drop is not that significant here. 00:24:42.681 --> 00:24:46.118 align:center line:-1 position:50% size:25% Why is this method of only using 5% 00:24:46.118 --> 00:24:49.555 align:center line:-1 position:50% size:26% of the available data effective? 00:24:49.555 --> 00:24:51.690 align:center line:-1 position:50% size:39% Average marginal effects are gonna give us this information. 00:24:51.690 --> 00:24:54.226 align:center line:-1 position:50% size:22% Here, this might look intimidating, 00:24:54.226 --> 00:24:55.594 align:center line:-1 position:50% size:34% but we'll walk through this. 00:24:55.594 --> 00:24:57.596 align:center line:-1 position:50% size:37% These different colored lines, 00:24:57.596 --> 00:24:58.530 align:center line:-1 position:50% size:29% as I mentioned before, 00:24:58.530 --> 00:25:02.835 align:center line:-1 position:50% size:22% we sampled 5% about 100 times. 00:25:02.835 --> 00:25:04.837 align:center line:-1 position:50% size:28% These are those posterior distributions 00:25:04.837 --> 00:25:06.972 align:center line:-1 position:50% size:35% for each of those iterations. 00:25:06.972 --> 00:25:08.474 align:center line:-1 position:50% size:30% The black line here is showing the average 00:25:08.474 --> 00:25:09.675 align:center line:-1 position:50% size:27% of all those iterations, 00:25:09.675 --> 00:25:11.877 align:center line:-1 position:50% size:29% and the red is showing 00:25:11.877 --> 00:25:14.146 align:center line:-1 position:50% size:35% the average marginal effect distribution 00:25:14.146 --> 00:25:18.083 align:center line:-1 position:50% size:34% for a model when we used 100% of the data. 00:25:18.083 --> 00:25:19.918 align:center line:-1 position:50% size:33% If we expect the model to perform well, 00:25:19.918 --> 00:25:25.023 align:center line:-1 position:50% size:40% we would expect these average marginal effect distributions 00:25:25.023 --> 00:25:27.493 align:center line:-1 position:50% size:24% to be well aligned with this red curve. 00:25:27.493 --> 00:25:28.794 align:center line:-1 position:50% size:19% And generally, 00:25:28.794 --> 00:25:33.165 align:center line:-1 position:50% size:40% across all the predictors that we included in our models, 00:25:33.165 --> 00:25:34.400 align:center line:-1 position:50% size:25% that's what we see. 00:25:34.400 --> 00:25:36.502 align:center line:-1 position:50% size:33% We see consistency here between these red curves 00:25:36.502 --> 00:25:39.104 align:center line:-1 position:50% size:34% and the multicolored lines. 00:25:39.104 --> 00:25:41.173 align:center line:-1 position:50% size:29% There's consistency in the predictor effects. 00:25:41.173 --> 00:25:42.775 align:center line:-1 position:50% size:38% Even though we're only using 00:25:42.775 --> 00:25:47.379 align:center line:-1 position:50% size:25% a very small subset of the data, 5%, 00:25:47.379 --> 00:25:50.883 align:center line:-1 position:50% size:30% it is still able to capture the influence 00:25:50.883 --> 00:25:52.384 align:center line:-1 position:50% size:26% of the landslide data appropriately 00:25:52.384 --> 00:26:01.193 align:center line:-1 position:50% size:21% to better assess the susceptibility 00:26:01.193 --> 00:26:03.962 align:center line:-1 position:50% size:38% across the domain of interest. 00:26:03.962 --> 00:26:06.031 align:center line:-1 position:50% size:28% So to just kind of drive home this point, 00:26:06.031 --> 00:26:07.699 align:center line:-1 position:50% size:27% on the right, we can see similarity 00:26:07.699 --> 00:26:10.903 align:center line:-1 position:50% size:33% and the AME distributions per slope, 00:26:10.903 --> 00:26:12.137 align:center line:-1 position:50% size:36% and when you train a model 00:26:12.137 --> 00:26:15.107 align:center line:-1 position:50% size:29% on each given location individually, 00:26:15.107 --> 00:26:18.644 align:center line:-1 position:50% size:24% you get quite a bit of dispersion here. 00:26:18.644 --> 00:26:20.679 align:center line:-1 position:50% size:36% To wrap up this first section, 00:26:20.679 --> 00:26:22.948 align:center line:-1 position:50% size:29% the unique geospatial attributes of landslides 00:26:22.948 --> 00:26:24.550 align:center line:-1 position:50% size:15% (inaudible). 00:26:24.550 --> 00:26:26.285 align:center line:-1 position:50% size:23% You can't really leave an area out 00:26:26.285 --> 00:26:28.120 align:center line:-1 position:50% size:33% and then expect that model to perform well 00:26:28.120 --> 00:26:30.122 align:center line:-1 position:50% size:15% in that area you left out. 00:26:30.122 --> 00:26:31.423 align:center line:-1 position:50% size:35% Limiting by the environment 00:26:31.423 --> 00:26:33.025 align:center line:-1 position:50% size:26% doesn't seem to improve anything. 00:26:33.025 --> 00:26:36.628 align:center line:-1 position:50% size:36% You can't expect to use Level II ecoregions and say, 00:26:36.628 --> 00:26:38.130 align:center line:-1 position:50% size:38% we're gonna develop a model 00:26:38.130 --> 00:26:40.432 align:center line:-1 position:50% size:41% on this area where we have really good data 00:26:40.432 --> 00:26:41.700 align:center line:-1 position:50% size:22% and apply it everywhere else, 00:26:41.700 --> 00:26:46.004 align:center line:-1 position:50% size:27% because we assume landslide similarities. 00:26:46.004 --> 00:26:48.307 align:center line:-1 position:50% size:24% We've shown that that's not effective, 00:26:48.307 --> 00:26:51.376 align:center line:-1 position:50% size:32% at least over the domains that we analyzed. 00:26:51.376 --> 00:26:54.146 align:center line:-1 position:50% size:34% But using a sparse uniform landslide inventory 00:26:54.146 --> 00:26:55.747 align:center line:-1 position:50% size:27% is relatively effective. 00:26:55.747 --> 00:26:57.516 align:center line:-1 position:50% size:32% So even if you only have a little bit of data 00:26:57.516 --> 00:27:00.719 align:center line:-1 position:50% size:34% that you can collect across your entire domain, 00:27:00.719 --> 00:27:02.588 align:center line:-1 position:50% size:37% that's better than trying to get 00:27:02.588 --> 00:27:05.257 align:center line:-1 position:50% size:24% a very detailed landslide inventory 00:27:05.257 --> 00:27:11.129 align:center line:-1 position:50% size:33% on a very small subset of your domain of interest. 00:27:11.129 --> 00:27:13.899 align:center line:-1 position:50% size:28% We've looked at how to manage areas 00:27:13.899 --> 00:27:16.702 align:center line:-1 position:50% size:22% with limited or no landslide data. 00:27:16.702 --> 00:27:19.738 align:center line:-1 position:50% size:26% Now we're gonna go into how to manage 00:27:19.738 --> 00:27:21.573 align:center line:-1 position:50% size:21% the imprecision of landslide data 00:27:21.573 --> 00:27:26.445 align:center line:-1 position:50% size:33% that's inherent with these susceptibility maps. 00:27:26.445 --> 00:27:29.481 align:center line:-1 position:50% size:40% Once again, we're gonna dive into the choice of mapping unit. 00:27:29.481 --> 00:27:30.716 align:center line:-1 position:50% size:22% By mapping unit, 00:27:30.716 --> 00:27:34.520 align:center line:-1 position:50% size:31% generally, there's two major mapping units 00:27:34.520 --> 00:27:35.888 align:center line:-1 position:50% size:21% that people use. 00:27:35.888 --> 00:27:38.790 align:center line:-1 position:50% size:45% The first and most common method is the raster-based approach, 00:27:38.790 --> 00:27:41.560 align:center line:-1 position:50% size:33% where you can characterize susceptibility 00:27:41.560 --> 00:27:46.665 align:center line:-1 position:50% size:32% for each pixel or grid unit of your raster 00:27:46.665 --> 00:27:49.635 align:center line:-1 position:50% size:31% with an output that looks something like up here, 00:27:49.635 --> 00:27:51.603 align:center line:-1 position:50% size:21% in the upper left. 00:27:51.603 --> 00:27:52.838 align:center line:-1 position:50% size:35% In contrast, another method 00:27:52.838 --> 00:27:55.307 align:center line:-1 position:50% size:24% that's been gaining more traction 00:27:55.307 --> 00:27:57.009 align:center line:-1 position:50% size:29% is the slope unit-based approach. 00:27:57.009 --> 00:28:00.879 align:center line:-1 position:50% size:29% What a slope unit is is a way of delineating 00:28:00.879 --> 00:28:02.848 align:center line:-1 position:50% size:30% your domain of interest 00:28:02.848 --> 00:28:05.017 align:center line:-1 position:50% size:33% according to drainage and divide lines. 00:28:05.017 --> 00:28:07.619 align:center line:-1 position:50% size:34% Here we have a watershed from Oregon. 00:28:07.619 --> 00:28:08.921 align:center line:-1 position:50% size:30% You can see that each of these lines 00:28:08.921 --> 00:28:15.394 align:center line:-1 position:50% size:33% correlates with a drainage and/or divide line. 00:28:15.394 --> 00:28:17.362 align:center line:-1 position:50% size:27% You can then analyze susceptibility 00:28:17.362 --> 00:28:19.698 align:center line:-1 position:50% size:20% over the whole area of interest. 00:28:19.698 --> 00:28:22.868 align:center line:-1 position:50% size:30% This gives us quite a bit of advantage 00:28:22.868 --> 00:28:25.837 align:center line:-1 position:50% size:27% when we're facing imprecise input data. 00:28:25.837 --> 00:28:27.039 align:center line:-1 position:50% size:37% What are these advantages? 00:28:27.039 --> 00:28:30.409 align:center line:-1 position:50% size:32% First, because we're dealing with areas, 00:28:30.409 --> 00:28:32.511 align:center line:-1 position:50% size:24% we're able to apply more statistics 00:28:32.511 --> 00:28:34.146 align:center line:-1 position:50% size:28% to that area of interest 00:28:34.146 --> 00:28:37.482 align:center line:-1 position:50% size:39% instead of just taking the mean slope value of each pixel, 00:28:37.482 --> 00:28:39.351 align:center line:-1 position:50% size:27% which is basically what you're limited to 00:28:39.351 --> 00:28:40.986 align:center line:-1 position:50% size:36% with a grid-based approach. 00:28:40.986 --> 00:28:42.187 align:center line:-1 position:50% size:26% Because you're dealing with an area 00:28:42.187 --> 00:28:44.456 align:center line:-1 position:50% size:20% with an array of slope values, 00:28:44.456 --> 00:28:45.857 align:center line:-1 position:50% size:29% you can then measure the standard deviation 00:28:45.857 --> 00:28:47.059 align:center line:-1 position:50% size:16% of the slope, 00:28:47.059 --> 00:28:48.293 align:center line:-1 position:50% size:26% or the max, the min, 00:28:48.293 --> 00:28:51.263 align:center line:-1 position:50% size:37% whatever metric you think is most appropriate 00:28:51.263 --> 00:28:56.635 align:center line:-1 position:50% size:34% for assessing susceptibility over your area of interest. 00:28:56.635 --> 00:28:58.437 align:center line:-1 position:50% size:17% Alternatively, 00:28:58.437 --> 00:29:00.505 align:center line:-1 position:50% size:30% it's a better way of accommodating 00:29:00.505 --> 00:29:02.841 align:center line:-1 position:50% size:26% the imprecise nature of your input data. 00:29:02.841 --> 00:29:05.043 align:center line:-1 position:50% size:26% And once again, by imprecise I mean 00:29:05.043 --> 00:29:08.180 align:center line:-1 position:50% size:31% heterogeneous formats, inaccurate locations, 00:29:08.180 --> 00:29:10.582 align:center line:-1 position:50% size:30% or no timing component 00:29:10.582 --> 00:29:15.887 align:center line:-1 position:50% size:20% and incomplete landslide data. 00:29:15.887 --> 00:29:17.489 align:center line:-1 position:50% size:27% This is gonna be a little bit theoretical, 00:29:17.489 --> 00:29:20.192 align:center line:-1 position:50% size:29% but I hope it kind of drives home the point. 00:29:20.192 --> 00:29:22.027 align:center line:-1 position:50% size:25% We have imprecise input data. 00:29:22.027 --> 00:29:24.563 align:center line:-1 position:50% size:31% The very nature of the data that we have 00:29:24.563 --> 00:29:26.365 align:center line:-1 position:50% size:35% across the U.S. for creating susceptibility maps 00:29:26.365 --> 00:29:28.233 align:center line:-1 position:50% size:16% is imprecise. 00:29:28.233 --> 00:29:29.701 align:center line:-1 position:50% size:33% You feed it into the model, 00:29:29.701 --> 00:29:34.039 align:center line:-1 position:50% size:33% you want the outputs to reflect that imprecision. 00:29:34.039 --> 00:29:37.676 align:center line:-1 position:50% size:36% But by using very high resolution rasters, 00:29:37.676 --> 00:29:42.114 align:center line:-1 position:50% size:30% you're actually showing a precise output 00:29:42.114 --> 00:29:45.450 align:center line:-1 position:50% size:31% that doesn't really match the input data. 00:29:45.450 --> 00:29:48.353 align:center line:-1 position:50% size:36% A metaphor that I thought of is that it's kind of like 00:29:48.353 --> 00:29:51.323 align:center line:-1 position:50% size:30% putting a go-kart engine in a pickup truck. 00:29:51.323 --> 00:29:53.692 align:center line:-1 position:50% size:33% There's an incompatibility there. 00:29:53.692 --> 00:29:55.427 align:center line:-1 position:50% size:21% If you try to use this pickup truck 00:29:55.427 --> 00:29:57.996 align:center line:-1 position:50% size:26% to go and tow a boat or whatever, 00:29:57.996 --> 00:29:59.431 align:center line:-1 position:50% size:37% it's not gonna work very well, 00:29:59.431 --> 00:30:01.900 align:center line:-1 position:50% size:20% it's just gonna crash and burn. 00:30:01.900 --> 00:30:03.735 align:center line:-1 position:50% size:31% That's kind of equivalent to trying to get 00:30:03.735 --> 00:30:07.606 align:center line:-1 position:50% size:26% super-detailed, very precise outputs 00:30:07.606 --> 00:30:08.874 align:center line:-1 position:50% size:38% of these susceptibility models 00:30:08.874 --> 00:30:11.410 align:center line:-1 position:50% size:28% when the input data really is kind of limited 00:30:11.410 --> 00:30:15.247 align:center line:-1 position:50% size:28% because it's imprecise by nature. 00:30:15.247 --> 00:30:18.784 align:center line:-1 position:50% size:32% To further dive into explicitly how slope units 00:30:18.784 --> 00:30:21.119 align:center line:-1 position:50% size:27% account for these first two points, 00:30:21.119 --> 00:30:22.754 align:center line:-1 position:50% size:28% I'm going to show a couple of examples. 00:30:22.754 --> 00:30:26.024 align:center line:-1 position:50% size:36% The first one is how to deal with heterogeneous formats 00:30:26.024 --> 00:30:29.127 align:center line:-1 position:50% size:26% and then inaccurate landslide locations. 00:30:29.127 --> 00:30:30.962 align:center line:-1 position:50% size:35% The grid-based approach is the most common way 00:30:30.962 --> 00:30:33.765 align:center line:-1 position:50% size:31% of dealing with this variation in formats. 00:30:33.765 --> 00:30:38.170 align:center line:-1 position:50% size:31% It's to convert the polygons into points, 00:30:38.170 --> 00:30:40.305 align:center line:-1 position:50% size:40% so therefore, they're compatible with the other points. 00:30:40.305 --> 00:30:42.274 align:center line:-1 position:50% size:38% Again, this is really applicable for large regions 00:30:42.274 --> 00:30:46.645 align:center line:-1 position:50% size:30% that we're talking about through this whole talk. 00:30:46.645 --> 00:30:48.647 align:center line:-1 position:50% size:23% You can imagine that by converting 00:30:48.647 --> 00:30:51.717 align:center line:-1 position:50% size:34% these whole polygons into a representative point, 00:30:51.717 --> 00:30:55.454 align:center line:-1 position:50% size:24% you're losing a lot of information 00:30:55.454 --> 00:30:56.855 align:center line:-1 position:50% size:37% that would be available to us 00:30:56.855 --> 00:30:58.690 align:center line:-1 position:50% size:28% if we were able to use the whole polygon, 00:30:58.690 --> 00:31:00.559 align:center line:-1 position:50% size:36% but that's really not tractable 00:31:00.559 --> 00:31:05.063 align:center line:-1 position:50% size:26% when you're dealing with large domains. 00:31:05.063 --> 00:31:07.099 align:center line:-1 position:50% size:30% By contrast, the slope unit approach 00:31:07.099 --> 00:31:09.534 align:center line:-1 position:50% size:36% converts everything to areas rather than points. 00:31:09.534 --> 00:31:12.704 align:center line:-1 position:50% size:34% That is, any landslide point or landslide polygon 00:31:12.704 --> 00:31:15.807 align:center line:-1 position:50% size:22% that intersects a slope unit area 00:31:15.807 --> 00:31:20.378 align:center line:-1 position:50% size:33% would then be mapped as containing a landslide, 00:31:20.378 --> 00:31:23.815 align:center line:-1 position:50% size:33% and then you can develop your model based off that. 00:31:23.815 --> 00:31:25.884 align:center line:-1 position:50% size:37% As far as landslide accuracy, 00:31:25.884 --> 00:31:27.719 align:center line:-1 position:50% size:27% at least as far as the position goes, 00:31:27.719 --> 00:31:31.189 align:center line:-1 position:50% size:33% Jacobs et al. back in 2020 tackled this problem, 00:31:31.189 --> 00:31:33.391 align:center line:-1 position:50% size:30% where they compared grid-based approaches 00:31:33.391 --> 00:31:35.360 align:center line:-1 position:50% size:32% to slope unit approaches, 00:31:35.360 --> 00:31:39.030 align:center line:-1 position:50% size:42% and they measured how resilient these methods were 00:31:39.030 --> 00:31:41.867 align:center line:-1 position:50% size:29% for imprecise locations of the landslides. 00:31:41.867 --> 00:31:43.268 align:center line:-1 position:50% size:22% They did this by-- 00:31:43.268 --> 00:31:45.003 align:center line:-1 position:50% size:28% they had polygons for all their landslides, 00:31:45.003 --> 00:31:48.406 align:center line:-1 position:50% size:39% and they then buffered those polygons by cell amount 00:31:48.406 --> 00:31:50.475 align:center line:-1 position:50% size:21% and randomly sampled a point 00:31:50.475 --> 00:31:51.977 align:center line:-1 position:50% size:36% within that buffered polygon. 00:31:51.977 --> 00:31:54.312 align:center line:-1 position:50% size:34% That's basically replicating what it'd be like 00:31:54.312 --> 00:31:57.415 align:center line:-1 position:50% size:30% if you had slightly imprecise or inaccurate 00:31:57.415 --> 00:31:59.284 align:center line:-1 position:50% size:19% landslide data. 00:31:59.284 --> 00:32:01.253 align:center line:-1 position:50% size:35% That's in contrast to these two other methods 00:32:01.253 --> 00:32:06.291 align:center line:-1 position:50% size:36% where you sampled within the landslide polygon. 00:32:06.291 --> 00:32:07.559 align:center line:-1 position:50% size:27% What they found was that they got 00:32:07.559 --> 00:32:10.195 align:center line:-1 position:50% size:29% a larger drop in model performance. 00:32:10.195 --> 00:32:12.430 align:center line:-1 position:50% size:29% Here we have the area under the curve metric 00:32:12.430 --> 00:32:15.534 align:center line:-1 position:50% size:27% using the pixel-based approaches, 00:32:15.534 --> 00:32:17.269 align:center line:-1 position:50% size:34% compared to the slope unit approaches, 00:32:17.269 --> 00:32:19.704 align:center line:-1 position:50% size:33% where, here on the right, you can see that the drop 00:32:19.704 --> 00:32:24.509 align:center line:-1 position:50% size:28% in model performance is less drastic. 00:32:24.509 --> 00:32:26.645 align:center line:-1 position:50% size:27% If slope units have all these advantages, 00:32:26.645 --> 00:32:29.514 align:center line:-1 position:50% size:31% hopefully, I've illustrated the advantages to you 00:32:29.514 --> 00:32:31.550 align:center line:-1 position:50% size:38% and you're a little bit convinced that these might be 00:32:31.550 --> 00:32:33.618 align:center line:-1 position:50% size:38% something worth entertaining, 00:32:33.618 --> 00:32:35.854 align:center line:-1 position:50% size:29% why don't more people use these? 00:32:35.854 --> 00:32:37.322 align:center line:-1 position:50% size:34% Well, after I was convinced 00:32:37.322 --> 00:32:39.157 align:center line:-1 position:50% size:25% that slope units are the way forward, 00:32:39.157 --> 00:32:41.827 align:center line:-1 position:50% size:39% at least for the size of domains we're interested in 00:32:41.827 --> 00:32:44.629 align:center line:-1 position:50% size:28% and the data we have available to us, 00:32:44.629 --> 00:32:46.531 align:center line:-1 position:50% size:29% I ran into the problem of the current methods 00:32:46.531 --> 00:32:49.167 align:center line:-1 position:50% size:32% for delineating these are either very laborious, 00:32:49.167 --> 00:32:52.070 align:center line:-1 position:50% size:29% you can manually delineate these things, 00:32:52.070 --> 00:32:55.340 align:center line:-1 position:50% size:41% very slow, and/or if you use an automatic method, 00:32:55.340 --> 00:32:57.642 align:center line:-1 position:50% size:18% most of them, 00:32:57.642 --> 00:33:01.146 align:center line:-1 position:50% size:28% the size of the sloping is arbitrarily set. 00:33:01.146 --> 00:33:04.616 align:center line:-1 position:50% size:37% There are examples of those that are not arbitrary. 00:33:04.616 --> 00:33:07.652 align:center line:-1 position:50% size:35% They do parameter sweeps that are very, very slow 00:33:07.652 --> 00:33:09.688 align:center line:-1 position:50% size:30% for optimizing the scale of these slope units. 00:33:09.688 --> 00:33:12.123 align:center line:-1 position:50% size:25% So, there isn't really anything out there 00:33:12.123 --> 00:33:15.827 align:center line:-1 position:50% size:33% that is fast and effective at delineating slope units. 00:33:15.827 --> 00:33:19.631 align:center line:-1 position:50% size:33% So we wrote an algorithm that I'm calling SUMak, 00:33:19.631 --> 00:33:21.132 align:center line:-1 position:50% size:26% or Slope Unit Maker, 00:33:21.132 --> 00:33:22.968 align:center line:-1 position:50% size:22% and it overcomes these challenges 00:33:22.968 --> 00:33:25.337 align:center line:-1 position:50% size:30% of the past approaches. 00:33:25.337 --> 00:33:26.504 align:center line:-1 position:50% size:23% First, it's very fast. 00:33:26.504 --> 00:33:28.673 align:center line:-1 position:50% size:26% It automatically runs in parallel. 00:33:28.673 --> 00:33:30.508 align:center line:-1 position:50% size:21% You can throw it on a cluster, 00:33:30.508 --> 00:33:31.843 align:center line:-1 position:50% size:30% and it's parameter free. 00:33:31.843 --> 00:33:34.012 align:center line:-1 position:50% size:35% There isn't anything that a user needs to dictate 00:33:34.012 --> 00:33:35.747 align:center line:-1 position:50% size:26% within this algorithm. 00:33:35.747 --> 00:33:38.817 align:center line:-1 position:50% size:34% It automatically adjusts the scale of the slope units 00:33:38.817 --> 00:33:41.486 align:center line:-1 position:50% size:27% according to the local terrain morphology. 00:33:41.486 --> 00:33:43.321 align:center line:-1 position:50% size:32% I'll get into more explicitly how that works 00:33:43.321 --> 00:33:46.791 align:center line:-1 position:50% size:26% in the coming slides. 00:33:46.791 --> 00:33:48.326 align:center line:-1 position:50% size:40% There's minimal preprocessing. 00:33:48.326 --> 00:33:51.096 align:center line:-1 position:50% size:25% You give it a DEM and area of interest 00:33:51.096 --> 00:33:53.265 align:center line:-1 position:50% size:20% and it'll spit out the slope units. 00:33:53.265 --> 00:33:54.799 align:center line:-1 position:50% size:24% It's all open source programming, 00:33:54.799 --> 00:33:58.036 align:center line:-1 position:50% size:37% GRASS and R and TauDEM, 00:33:58.036 --> 00:34:02.173 align:center line:-1 position:50% size:30% so a user could adapt it to their needs. 00:34:02.173 --> 00:34:05.076 align:center line:-1 position:50% size:39% How does this algorithm work? 00:34:05.076 --> 00:34:09.247 align:center line:-1 position:50% size:37% It uses something referred to as a constant drop law. 00:34:09.247 --> 00:34:12.017 align:center line:-1 position:50% size:32% A constant drop law says 00:34:12.017 --> 00:34:14.452 align:center line:-1 position:50% size:30% as you go up Strahler stream orders, 00:34:14.452 --> 00:34:18.390 align:center line:-1 position:50% size:38% shown here in this figure, left, 00:34:18.390 --> 00:34:20.825 align:center line:-1 position:50% size:35% if you have a fluvial dominated system, 00:34:20.825 --> 00:34:22.994 align:center line:-1 position:50% size:29% where the morphology is controlled 00:34:22.994 --> 00:34:24.963 align:center line:-1 position:50% size:26% by fluvial processes, 00:34:24.963 --> 00:34:27.532 align:center line:-1 position:50% size:42% you expect that as you go across these Strahler stream orders, 00:34:27.532 --> 00:34:30.936 align:center line:-1 position:50% size:34% you would get a consistent drop in elevation 00:34:30.936 --> 00:34:33.204 align:center line:-1 position:50% size:38% between these stream orders. 00:34:33.204 --> 00:34:35.106 align:center line:-1 position:50% size:25% Once you get to the stream order 00:34:35.106 --> 00:34:38.243 align:center line:-1 position:50% size:28% that you get a statistical difference 00:34:38.243 --> 00:34:40.111 align:center line:-1 position:50% size:30% in the drop in elevation, 00:34:40.111 --> 00:34:42.380 align:center line:-1 position:50% size:30% it's been shown by Tarbotton and others 00:34:42.380 --> 00:34:45.684 align:center line:-1 position:50% size:34% that you're no longer in a fluvial dominated system 00:34:45.684 --> 00:34:49.621 align:center line:-1 position:50% size:38% but rather a hillslope dominated system. 00:34:49.621 --> 00:34:52.557 align:center line:-1 position:50% size:36% So what we do is that we go through different thresholds 00:34:52.557 --> 00:34:55.026 align:center line:-1 position:50% size:28% for flow accumulation, 00:34:55.026 --> 00:34:59.898 align:center line:-1 position:50% size:38% then we find where this constant drop law breaks. 00:34:59.898 --> 00:35:00.999 align:center line:-1 position:50% size:28% And right at that point, 00:35:00.999 --> 00:35:02.400 align:center line:-1 position:50% size:40% we delineate these watersheds. 00:35:02.400 --> 00:35:04.402 align:center line:-1 position:50% size:28% That's basically the largest watershed 00:35:04.402 --> 00:35:07.505 align:center line:-1 position:50% size:25% that captures hillslope processes, 00:35:07.505 --> 00:35:10.675 align:center line:-1 position:50% size:40% then we split these watersheds by their longest flow paths. 00:35:10.675 --> 00:35:13.678 align:center line:-1 position:50% size:40% So this output is what would be recognized in the field 00:35:13.678 --> 00:35:16.114 align:center line:-1 position:50% size:14% as a slope. 00:35:16.114 --> 00:35:19.451 align:center line:-1 position:50% size:36% Just to compare our method to the other method 00:35:19.451 --> 00:35:20.785 align:center line:-1 position:50% size:29% that's commonly used, 00:35:20.785 --> 00:35:23.455 align:center line:-1 position:50% size:26% it's a great algorithm developed by Alvioli, 00:35:23.455 --> 00:35:25.457 align:center line:-1 position:50% size:34% and it's called r.slopeunits. 00:35:25.457 --> 00:35:27.225 align:center line:-1 position:50% size:25% It's kind of really laid the foundation, 00:35:27.225 --> 00:35:29.861 align:center line:-1 position:50% size:32% really allowed slope units to gain traction. 00:35:29.861 --> 00:35:32.430 align:center line:-1 position:50% size:22% But it's very slow. 00:35:32.430 --> 00:35:34.632 align:center line:-1 position:50% size:24% So we delineated the island of Sicily, 00:35:34.632 --> 00:35:36.968 align:center line:-1 position:50% size:35% which is an area that was already delineated 00:35:36.968 --> 00:35:38.837 align:center line:-1 position:50% size:23% using r.slopeunits, 00:35:38.837 --> 00:35:41.473 align:center line:-1 position:50% size:35% and we wanted to compare how our algorithm 00:35:41.473 --> 00:35:42.574 align:center line:-1 position:50% size:25% compared to theirs. 00:35:42.574 --> 00:35:44.542 align:center line:-1 position:50% size:25% And you can see the relative outputs. 00:35:44.542 --> 00:35:47.412 align:center line:-1 position:50% size:24% TauDEM here should be SUMak. 00:35:47.412 --> 00:35:49.514 align:center line:-1 position:50% size:33% So ours are shown in red, theirs is shown in blue. 00:35:49.514 --> 00:35:53.485 align:center line:-1 position:50% size:34% There isn't a perfect match between the two, 00:35:53.485 --> 00:35:56.321 align:center line:-1 position:50% size:24% but if you look at the whole island 00:35:56.321 --> 00:35:59.457 align:center line:-1 position:50% size:32% and quantify a metric that they use to optimize, 00:35:59.457 --> 00:36:02.794 align:center line:-1 position:50% size:30% r.slopeunits uses to optimize their model, 00:36:02.794 --> 00:36:04.129 align:center line:-1 position:50% size:19% which is this V. 00:36:04.129 --> 00:36:06.464 align:center line:-1 position:50% size:33% It's an area of normalized local variance. 00:36:06.464 --> 00:36:07.732 align:center line:-1 position:50% size:28% What that basically is, 00:36:07.732 --> 00:36:11.236 align:center line:-1 position:50% size:37% it's a measure of the homogeneity of aspect 00:36:11.236 --> 00:36:12.971 align:center line:-1 position:50% size:29% within each slope unit, 00:36:12.971 --> 00:36:15.306 align:center line:-1 position:50% size:27% which is argued to be a good slope unit 00:36:15.306 --> 00:36:18.009 align:center line:-1 position:50% size:28% should have pretty uniform aspect. 00:36:18.009 --> 00:36:19.577 align:center line:-1 position:50% size:34% These are the distributions that we get. 00:36:19.577 --> 00:36:22.380 align:center line:-1 position:50% size:23% So using SUMak, using r.slopeunits, 00:36:22.380 --> 00:36:25.316 align:center line:-1 position:50% size:41% there's quite a bit of overlap between these two distributions, 00:36:25.316 --> 00:36:26.851 align:center line:-1 position:50% size:21% suggesting that, 00:36:26.851 --> 00:36:29.454 align:center line:-1 position:50% size:20% while theirs is a little bit better 00:36:29.454 --> 00:36:32.424 align:center line:-1 position:50% size:25% as far as this metric is concerned, 00:36:32.424 --> 00:36:34.192 align:center line:-1 position:50% size:32% ours is remarkably faster. 00:36:34.192 --> 00:36:36.928 align:center line:-1 position:50% size:34% In fact, across the country of Italy, 00:36:36.928 --> 00:36:39.264 align:center line:-1 position:50% size:34% it would take our algorithm about three days 00:36:39.264 --> 00:36:41.232 align:center line:-1 position:50% size:33% to process the same area 00:36:41.232 --> 00:36:44.969 align:center line:-1 position:50% size:33% that took r.slopeunits three months to delineate, 00:36:44.969 --> 00:36:46.838 align:center line:-1 position:50% size:30% and they had four times the number of cores 00:36:46.838 --> 00:36:48.373 align:center line:-1 position:50% size:34% and five times the memory that we had. 00:36:48.373 --> 00:36:51.776 align:center line:-1 position:50% size:30% I delineated this just on my local laptop. 00:36:51.776 --> 00:36:56.948 align:center line:-1 position:50% size:26% So, really, SUMak is an effective means 00:36:56.948 --> 00:37:00.819 align:center line:-1 position:50% size:33% of delineating large areas for slope units, 00:37:00.819 --> 00:37:04.389 align:center line:-1 position:50% size:39% and you get very good results. 00:37:04.389 --> 00:37:07.125 align:center line:-1 position:50% size:29% Now I want to illustrate the uses of slope units 00:37:07.125 --> 00:37:09.794 align:center line:-1 position:50% size:29% and bring it back to how it's able to deal 00:37:09.794 --> 00:37:12.363 align:center line:-1 position:50% size:34% with imprecise data, again, 00:37:12.363 --> 00:37:14.766 align:center line:-1 position:50% size:32% putting the go-kart motor in a pickup truck. 00:37:14.766 --> 00:37:17.769 align:center line:-1 position:50% size:34% We're going to try to illustrate that point here. 00:37:17.769 --> 00:37:20.371 align:center line:-1 position:50% size:38% We're going to do this with three different locations-- 00:37:20.371 --> 00:37:23.441 align:center line:-1 position:50% size:29% two locations from Western Oregon, 00:37:23.441 --> 00:37:26.878 align:center line:-1 position:50% size:42% we have the Umpqua Watershed 00:37:26.878 --> 00:37:30.315 align:center line:-1 position:50% size:38% and the Calapooia Watershed in Western Oregon, 00:37:30.315 --> 00:37:34.719 align:center line:-1 position:50% size:36% and also, we're going to use the island of Puerto Rico. 00:37:34.719 --> 00:37:35.954 align:center line:-1 position:50% size:37% For the Puerto Rico data set, 00:37:35.954 --> 00:37:38.156 align:center line:-1 position:50% size:32% we used an event-based landslide catalogue 00:37:38.156 --> 00:37:41.793 align:center line:-1 position:50% size:40% from the 2017 Hurricane Maria. 00:37:41.793 --> 00:37:44.129 align:center line:-1 position:50% size:36% We'll dig into how slope units are effective 00:37:44.129 --> 00:37:49.334 align:center line:-1 position:50% size:31% for event-based susceptibility maps also. 00:37:49.334 --> 00:37:51.336 align:center line:-1 position:50% size:38% As I said, we're going to compare slope unit outputs 00:37:51.336 --> 00:37:52.670 align:center line:-1 position:50% size:28% to grid-based outputs, 00:37:52.670 --> 00:37:56.007 align:center line:-1 position:50% size:34% and because we're dealing with variable formats 00:37:56.007 --> 00:37:58.376 align:center line:-1 position:50% size:33% in these data sets, both points and polygons, 00:37:58.376 --> 00:38:01.346 align:center line:-1 position:50% size:37% we need to use some sort of standardization procedure. 00:38:01.346 --> 00:38:05.717 align:center line:-1 position:50% size:26% We use an array of different methods. 00:38:05.717 --> 00:38:07.685 align:center line:-1 position:50% size:22% First, for using a 10 meter DEM, 00:38:07.685 --> 00:38:10.288 align:center line:-1 position:50% size:25% we take the highest elevation point 00:38:10.288 --> 00:38:11.923 align:center line:-1 position:50% size:26% within each polygon 00:38:11.923 --> 00:38:14.526 align:center line:-1 position:50% size:21% and convert that into a point. 00:38:14.526 --> 00:38:17.095 align:center line:-1 position:50% size:39% We also take the middle value of each polygon 00:38:17.095 --> 00:38:19.264 align:center line:-1 position:50% size:34% and convert that into a representative point. 00:38:19.264 --> 00:38:20.632 align:center line:-1 position:50% size:28% Then we take the highest point also, 00:38:20.632 --> 00:38:22.400 align:center line:-1 position:50% size:30% using a 30 meter DEM, 00:38:22.400 --> 00:38:25.436 align:center line:-1 position:50% size:30% to make a slightly more conservative map 00:38:25.436 --> 00:38:29.340 align:center line:-1 position:50% size:35% using larger mapping units. 00:38:29.340 --> 00:38:31.176 align:center line:-1 position:50% size:28% Then also we're going to take multiple points 00:38:31.176 --> 00:38:33.444 align:center line:-1 position:50% size:35% from each polygon using a 200 meter spacing, 00:38:33.444 --> 00:38:35.747 align:center line:-1 position:50% size:36% again, with a 10 meter DEM 00:38:35.747 --> 00:38:38.116 align:center line:-1 position:50% size:32% just to see if any of these can approach 00:38:38.116 --> 00:38:41.519 align:center line:-1 position:50% size:43% the model performance of the slope unit-based approaches. 00:38:41.519 --> 00:38:43.421 align:center line:-1 position:50% size:36% For our models that we're going to compare, 00:38:43.421 --> 00:38:44.489 align:center line:-1 position:50% size:24% we're going to use the frequentist 00:38:44.489 --> 00:38:46.157 align:center line:-1 position:50% size:25% Logistic Regression and XGBoost. 00:38:46.157 --> 00:38:48.259 align:center line:-1 position:50% size:36% We use two different models just so we can see 00:38:48.259 --> 00:38:49.527 align:center line:-1 position:50% size:23% that we're getting consistent results 00:38:49.527 --> 00:38:50.828 align:center line:-1 position:50% size:22% between the two. 00:38:50.828 --> 00:38:53.031 align:center line:-1 position:50% size:31% As far as predictor data, 00:38:53.031 --> 00:38:54.766 align:center line:-1 position:50% size:27% we're going to keep it very simple. 00:38:54.766 --> 00:38:57.569 align:center line:-1 position:50% size:31% We're only using elevation, slope, aspect, 00:38:57.569 --> 00:39:00.338 align:center line:-1 position:50% size:34% roughness, and curvature, as far as static variables. 00:39:00.338 --> 00:39:02.407 align:center line:-1 position:50% size:37% For the Puerto Rico data set, 00:39:02.407 --> 00:39:04.709 align:center line:-1 position:50% size:34% because it is an event-based catalogue, 00:39:04.709 --> 00:39:08.313 align:center line:-1 position:50% size:34% we're also going to include a trigger parameter 00:39:08.313 --> 00:39:10.415 align:center line:-1 position:50% size:28% that's soil moisture using the SMAP data, 00:39:10.415 --> 00:39:11.916 align:center line:-1 position:50% size:33% which is shown by people 00:39:11.916 --> 00:39:13.651 align:center line:-1 position:50% size:39% who have looked more closely at Puerto Rico 00:39:13.651 --> 00:39:17.989 align:center line:-1 position:50% size:39% as being an effective measure for susceptibility 00:39:17.989 --> 00:39:20.225 align:center line:-1 position:50% size:26% for Hurricane Maria. 00:39:20.225 --> 00:39:22.560 align:center line:-1 position:50% size:36% For the two slope unit maps, 00:39:22.560 --> 00:39:25.897 align:center line:-1 position:50% size:39% we're going to develop models using only the median values, 00:39:25.897 --> 00:39:28.566 align:center line:-1 position:50% size:29% predictor values, within each slope unit, 00:39:28.566 --> 00:39:29.867 align:center line:-1 position:50% size:37% and then also another model 00:39:29.867 --> 00:39:33.071 align:center line:-1 position:50% size:34% using the median standard deviation values. 00:39:33.071 --> 00:39:35.640 align:center line:-1 position:50% size:38% Model evaluation, we're going to use area under the curve. 00:39:35.640 --> 00:39:37.709 align:center line:-1 position:50% size:36% We're also going to throw in an additional metric, 00:39:37.709 --> 00:39:39.644 align:center line:-1 position:50% size:20% which is called the Brier score, 00:39:39.644 --> 00:39:41.212 align:center line:-1 position:50% size:30% which is just the mean squared error 00:39:41.212 --> 00:39:44.315 align:center line:-1 position:50% size:32% between the observation and our model outputs. 00:39:44.315 --> 00:39:46.517 align:center line:-1 position:50% size:36% We want lower values there. 00:39:46.517 --> 00:39:47.785 align:center line:-1 position:50% size:31% Let's look at the outputs, 00:39:47.785 --> 00:39:50.788 align:center line:-1 position:50% size:36% let's look at the results here. 00:39:50.788 --> 00:39:54.359 align:center line:-1 position:50% size:43% Up here at the top we have the two raster-based approaches. 00:39:54.359 --> 00:39:57.061 align:center line:-1 position:50% size:41% I'm showing the best-performing results here 00:39:57.061 --> 00:39:58.396 align:center line:-1 position:50% size:28% from the raster-based approaches. 00:39:58.396 --> 00:40:00.665 align:center line:-1 position:50% size:40% On the bottom, we have the slope unit outputs. 00:40:00.665 --> 00:40:04.702 align:center line:-1 position:50% size:40% These are outputs using median and standard deviation 00:40:04.702 --> 00:40:07.071 align:center line:-1 position:50% size:24% of each slope unit. 00:40:07.071 --> 00:40:08.539 align:center line:-1 position:50% size:32% A few observations here. 00:40:08.539 --> 00:40:12.277 align:center line:-1 position:50% size:37% Generally, they both highlight similar locations 00:40:12.277 --> 00:40:17.282 align:center line:-1 position:50% size:35% of having both low and high susceptibility values. 00:40:17.282 --> 00:40:21.019 align:center line:-1 position:50% size:37% But the raster-based approaches are able to show 00:40:21.019 --> 00:40:24.222 align:center line:-1 position:50% size:27% a lot more variability in these probabilities 00:40:24.222 --> 00:40:26.057 align:center line:-1 position:50% size:29% across these domains. 00:40:26.057 --> 00:40:28.359 align:center line:-1 position:50% size:30% While this level of detail can be advantageous, 00:40:28.359 --> 00:40:32.664 align:center line:-1 position:50% size:42% the high resolution rasters may be requiring too much detail 00:40:32.664 --> 00:40:35.967 align:center line:-1 position:50% size:33% from the models trained with imprecise input data. 00:40:35.967 --> 00:40:38.569 align:center line:-1 position:50% size:42% This can lead to poor confidence in the model outputs. 00:40:38.569 --> 00:40:41.839 align:center line:-1 position:50% size:31% Again, these maps were created with data 00:40:41.839 --> 00:40:45.610 align:center line:-1 position:50% size:36% with heterogeneous formats, points and polygons, 00:40:45.610 --> 00:40:48.479 align:center line:-1 position:50% size:41% somewhat inaccurate, probably, locations. 00:40:48.479 --> 00:40:50.948 align:center line:-1 position:50% size:36% There's no timing with any of these landslides. 00:40:50.948 --> 00:40:54.352 align:center line:-1 position:50% size:37% There's no real measure of the level of completeness. 00:40:54.352 --> 00:41:00.591 align:center line:-1 position:50% size:37% Once again, I don't even really know what that means. 00:41:00.591 --> 00:41:02.093 align:center line:-1 position:50% size:25% There's no way of really measuring 00:41:02.093 --> 00:41:05.163 align:center line:-1 position:50% size:38% how complete an inventory is. 00:41:05.163 --> 00:41:06.964 align:center line:-1 position:50% size:40% In contrast, the slope unit-based approach, 00:41:06.964 --> 00:41:09.434 align:center line:-1 position:50% size:28% because it uses a larger mapping unit, 00:41:09.434 --> 00:41:11.269 align:center line:-1 position:50% size:29% it's more conservative, 00:41:11.269 --> 00:41:12.337 align:center line:-1 position:50% size:24% and in my opinion, 00:41:12.337 --> 00:41:13.938 align:center line:-1 position:50% size:30% it's able to better accommodate 00:41:13.938 --> 00:41:16.240 align:center line:-1 position:50% size:26% the imprecise nature of this input data-- 00:41:16.240 --> 00:41:23.114 align:center line:-1 position:50% size:33% again, the go-kart engine inside of the pickup truck. 00:41:23.114 --> 00:41:26.651 align:center line:-1 position:50% size:30% That's what we're trying to not have here. 00:41:26.651 --> 00:41:29.587 align:center line:-1 position:50% size:31% Just to further compare the Umpqua Watershed, 00:41:29.587 --> 00:41:32.156 align:center line:-1 position:50% size:28% here's the distribution of known landslides 00:41:32.156 --> 00:41:34.225 align:center line:-1 position:50% size:22% over this domain. 00:41:34.225 --> 00:41:36.661 align:center line:-1 position:50% size:34% You can see that the landslide deposits here 00:41:36.661 --> 00:41:39.230 align:center line:-1 position:50% size:37% cover almost the whole area. 00:41:39.230 --> 00:41:42.200 align:center line:-1 position:50% size:18% In my opinion, 00:41:42.200 --> 00:41:45.703 align:center line:-1 position:50% size:40% what you would expect a useful susceptibility map to do, 00:41:45.703 --> 00:41:47.438 align:center line:-1 position:50% size:30% at least over the scales we're interested in here 00:41:47.438 --> 00:41:48.806 align:center line:-1 position:50% size:25% with our input data, 00:41:48.806 --> 00:41:51.242 align:center line:-1 position:50% size:22% is to model most of this area 00:41:51.242 --> 00:41:53.244 align:center line:-1 position:50% size:35% as being highly susceptible. 00:41:53.244 --> 00:41:54.879 align:center line:-1 position:50% size:31% Looking at the Calapooia data sets, 00:41:54.879 --> 00:41:56.280 align:center line:-1 position:50% size:29% we see a similar thing. 00:41:56.280 --> 00:41:58.015 align:center line:-1 position:50% size:37% Again, here's the distribution of landslides, 00:41:58.015 --> 00:42:00.051 align:center line:-1 position:50% size:41% here are the two model outputs. 00:42:00.051 --> 00:42:03.654 align:center line:-1 position:50% size:37% It we want a more quantitative measure, 00:42:03.654 --> 00:42:05.757 align:center line:-1 position:50% size:24% this effect we've here plotted, 00:42:05.757 --> 00:42:09.227 align:center line:-1 position:50% size:29% the percent area against the probability. 00:42:09.227 --> 00:42:10.895 align:center line:-1 position:50% size:26% If we look at the Umpqua data, 00:42:10.895 --> 00:42:13.464 align:center line:-1 position:50% size:36% the two slope unit-based approaches are shown here 00:42:13.464 --> 00:42:14.832 align:center line:-1 position:50% size:26% in green and yellow. 00:42:14.832 --> 00:42:19.404 align:center line:-1 position:50% size:40% You can see that it gets a much higher proportion of the area 00:42:19.404 --> 00:42:23.474 align:center line:-1 position:50% size:28% at higher probabilities, 00:42:23.474 --> 00:42:25.309 align:center line:-1 position:50% size:35% in contrast to the grid-based approaches, 00:42:25.309 --> 00:42:27.412 align:center line:-1 position:50% size:28% where, really, the highest proportion 00:42:27.412 --> 00:42:29.080 align:center line:-1 position:50% size:16% is around .5. 00:42:29.080 --> 00:42:32.150 align:center line:-1 position:50% size:24% It's more nebulous in that sense. 00:42:32.150 --> 00:42:34.852 align:center line:-1 position:50% size:22% At the Calapooia, it's less stark, 00:42:34.852 --> 00:42:37.088 align:center line:-1 position:50% size:28% but you can also see kind of a similar thing, 00:42:37.088 --> 00:42:40.091 align:center line:-1 position:50% size:39% where both the highs and lows are more extreme 00:42:40.091 --> 00:42:41.692 align:center line:-1 position:50% size:32% with the slope unit-based approaches 00:42:41.692 --> 00:42:44.061 align:center line:-1 position:50% size:32% as opposed to the raster. 00:42:44.061 --> 00:42:45.630 align:center line:-1 position:50% size:34% Let's look at the event-based catalogue 00:42:45.630 --> 00:42:47.899 align:center line:-1 position:50% size:23% from Puerto Rico. 00:42:47.899 --> 00:42:50.334 align:center line:-1 position:50% size:28% Here's the distribution of landslides 00:42:50.334 --> 00:42:51.936 align:center line:-1 position:50% size:28% from Hurricane Maria. 00:42:51.936 --> 00:42:53.971 align:center line:-1 position:50% size:34% At the top, we have the raster-based approach 00:42:53.971 --> 00:42:55.873 align:center line:-1 position:50% size:30% using a 30 meter DEM, 00:42:55.873 --> 00:42:57.041 align:center line:-1 position:50% size:37% and then down at the bottom, 00:42:57.041 --> 00:42:59.110 align:center line:-1 position:50% size:40% we have the slope unit-based approach. 00:42:59.110 --> 00:43:02.713 align:center line:-1 position:50% size:28% I think we're seeing the same things here. 00:43:02.713 --> 00:43:06.617 align:center line:-1 position:50% size:36% Looking at the percent area, once again, 00:43:06.617 --> 00:43:09.287 align:center line:-1 position:50% size:38% it's less stark here than it was for the Umpqua Watershed, 00:43:09.287 --> 00:43:11.622 align:center line:-1 position:50% size:23% but you do still get higher proportion 00:43:11.622 --> 00:43:12.990 align:center line:-1 position:50% size:33% with the slope unit-based, 00:43:12.990 --> 00:43:15.092 align:center line:-1 position:50% size:41% as opposed to the raster-based. 00:43:15.092 --> 00:43:18.763 align:center line:-1 position:50% size:38% Looking at the model metrics, area under the curve, 00:43:18.763 --> 00:43:21.466 align:center line:-1 position:50% size:37% here we have all the different grid-based approaches 00:43:21.466 --> 00:43:22.967 align:center line:-1 position:50% size:17% that we used. 00:43:22.967 --> 00:43:24.402 align:center line:-1 position:50% size:35% That is the different standardization procedures 00:43:24.402 --> 00:43:27.905 align:center line:-1 position:50% size:34% we did to get the polygons and points consistent. 00:43:27.905 --> 00:43:31.275 align:center line:-1 position:50% size:35% In bold here, we have the slope unit-based maps. 00:43:31.275 --> 00:43:33.444 align:center line:-1 position:50% size:33% We're looking at Logistic Regression in red 00:43:33.444 --> 00:43:36.981 align:center line:-1 position:50% size:27% and XGBoost in blue. 00:43:36.981 --> 00:43:39.917 align:center line:-1 position:50% size:29% Hopefully, you can see that generally, 00:43:39.917 --> 00:43:42.453 align:center line:-1 position:50% size:35% across each of the domains that we studied, 00:43:42.453 --> 00:43:43.955 align:center line:-1 position:50% size:28% you get a higher area under the curve, 00:43:43.955 --> 00:43:46.491 align:center line:-1 position:50% size:20% which suggests better model fit 00:43:46.491 --> 00:43:50.094 align:center line:-1 position:50% size:33% when you use slope units, 00:43:50.094 --> 00:43:54.499 align:center line:-1 position:50% size:26% in contrast to grid-based maps. 00:43:54.499 --> 00:43:57.101 align:center line:-1 position:50% size:38% Brier score is another metric-- again, mean squared error, 00:43:57.101 --> 00:43:59.737 align:center line:-1 position:50% size:36% so we want low error values. 00:43:59.737 --> 00:44:04.876 align:center line:-1 position:50% size:42% You can see that in bold we have lower Brier scores generally 00:44:04.876 --> 00:44:06.177 align:center line:-1 position:50% size:32% with the slope unit-based approaches, 00:44:06.177 --> 00:44:08.746 align:center line:-1 position:50% size:37% in contrast to the raster-based approaches 00:44:08.746 --> 00:44:12.517 align:center line:-1 position:50% size:32% for both XGBoost and Logistic Regression. 00:44:12.517 --> 00:44:15.052 align:center line:-1 position:50% size:32% Wrapping up this section, 00:44:15.052 --> 00:44:17.088 align:center line:-1 position:50% size:37% hopefully, I've convinced you that slope units are 00:44:17.088 --> 00:44:18.422 align:center line:-1 position:50% size:28% a powerful alternative 00:44:18.422 --> 00:44:19.557 align:center line:-1 position:50% size:26% when you're dealing with the scales 00:44:19.557 --> 00:44:22.260 align:center line:-1 position:50% size:29% that we're dealing with, very large regions. 00:44:22.260 --> 00:44:23.895 align:center line:-1 position:50% size:30% Again, this might not be as applicable 00:44:23.895 --> 00:44:27.465 align:center line:-1 position:50% size:35% when you're trying to analyze the susceptibility 00:44:27.465 --> 00:44:28.666 align:center line:-1 position:50% size:25% over a smaller area 00:44:28.666 --> 00:44:30.034 align:center line:-1 position:50% size:25% and you might have more detailed data. 00:44:30.034 --> 00:44:31.502 align:center line:-1 position:50% size:28% But the scales we're looking at here, 00:44:31.502 --> 00:44:33.070 align:center line:-1 position:50% size:16% large areas, 00:44:33.070 --> 00:44:35.339 align:center line:-1 position:50% size:25% it is a very powerful alternative. 00:44:35.339 --> 00:44:37.808 align:center line:-1 position:50% size:26% It's able to provide an effective solution 00:44:37.808 --> 00:44:42.513 align:center line:-1 position:50% size:38% for inconsistent and imprecise landslide data. 00:44:42.513 --> 00:44:44.382 align:center line:-1 position:50% size:41% It increases the model accuracy, 00:44:44.382 --> 00:44:47.585 align:center line:-1 position:50% size:29% as shown by the two metrics before. 00:44:47.585 --> 00:44:51.355 align:center line:-1 position:50% size:38% The conservative output better matches the input data, 00:44:51.355 --> 00:44:53.858 align:center line:-1 position:50% size:37% so you're keeping the go-kart a go-kart, 00:44:53.858 --> 00:44:58.195 align:center line:-1 position:50% size:35% you're not trying to put your go-kart engine in the truck. 00:44:58.195 --> 00:45:01.499 align:center line:-1 position:50% size:27% Let's wrap things up. 00:45:01.499 --> 00:45:03.234 align:center line:-1 position:50% size:18% In conclusion, 00:45:03.234 --> 00:45:05.303 align:center line:-1 position:50% size:35% developing a meaningful landslide susceptibility map 00:45:05.303 --> 00:45:06.737 align:center line:-1 position:50% size:23% over large regions 00:45:06.737 --> 00:45:08.306 align:center line:-1 position:50% size:33% is a very difficult problem. 00:45:08.306 --> 00:45:10.174 align:center line:-1 position:50% size:31% We've investigated here 00:45:10.174 --> 00:45:12.643 align:center line:-1 position:50% size:35% where susceptibility models can be applied. 00:45:12.643 --> 00:45:14.045 align:center line:-1 position:50% size:23% We've found that, 00:45:14.045 --> 00:45:16.747 align:center line:-1 position:50% size:35% at least over the study sites that we looked at, 00:45:16.747 --> 00:45:19.684 align:center line:-1 position:50% size:35% every region needs to have some sort of representation 00:45:19.684 --> 00:45:22.453 align:center line:-1 position:50% size:28% in your training model 00:45:22.453 --> 00:45:27.258 align:center line:-1 position:50% size:37% for it to be applicable over the entire domain of interest. 00:45:27.258 --> 00:45:29.827 align:center line:-1 position:50% size:31% Using one concentrated landslide data set 00:45:29.827 --> 00:45:31.429 align:center line:-1 position:50% size:25% and then applying it to other areas 00:45:31.429 --> 00:45:35.967 align:center line:-1 position:50% size:41% with assumed similarities to where you trained your model 00:45:35.967 --> 00:45:37.368 align:center line:-1 position:50% size:27% probably isn't a very safe approach 00:45:37.368 --> 00:45:40.471 align:center line:-1 position:50% size:26% when you're dealing with large scales. 00:45:40.471 --> 00:45:42.106 align:center line:-1 position:50% size:32% Slope units are an effective mapping unit 00:45:42.106 --> 00:45:44.442 align:center line:-1 position:50% size:23% for imprecise data 00:45:44.442 --> 00:45:45.843 align:center line:-1 position:50% size:33% that's commonly available for creating 00:45:45.843 --> 00:45:47.178 align:center line:-1 position:50% size:33% these susceptibility maps, 00:45:47.178 --> 00:45:49.246 align:center line:-1 position:50% size:38% and I've shown how SUMak is 00:45:49.246 --> 00:45:53.284 align:center line:-1 position:50% size:35% a new, fast method of implementing slope units 00:45:53.284 --> 00:45:56.087 align:center line:-1 position:50% size:37% to hopefully allow slope units to gain more traction 00:45:56.087 --> 00:46:00.257 align:center line:-1 position:50% size:36% and for more people to take advantage of these benefits 00:46:00.257 --> 00:46:04.428 align:center line:-1 position:50% size:31% that slope units provide. 00:46:04.428 --> 00:46:07.231 align:center line:-1 position:50% size:29% With that, I'd be happy to take any questions. 00:46:07.231 --> 00:46:09.567 align:center line:-1 position:50% size:19% You can read-- 00:46:09.567 --> 00:46:11.836 align:center line:-1 position:50% size:36% the paper just got published on the first half of this talk 00:46:11.836 --> 00:46:14.905 align:center line:-1 position:50% size:28% in JGR Earth Surface, and hopefully, 00:46:14.905 --> 00:46:16.607 align:center line:-1 position:50% size:42% the SUMak algorithm manuscript 00:46:16.607 --> 00:46:20.411 align:center line:-1 position:50% size:32% will be out in not too long as well. 00:46:20.411 --> 00:46:22.446 align:center line:-1 position:50% size:28% I'd be happy to take any questions.