0 00:00:04,500 --> 00:00:05,790 Awesome, thanks for the introduction. 2 00:00:05,790 --> 00:00:07,710 Thank you everyone for coming to the talk, 3 00:00:07,710 --> 00:00:09,153 here and virtually. 4 00:00:10,020 --> 00:00:13,740 And first, yeah, I'm Larry, I'm with the Vermont Co-Op Unit. 5 00:00:13,740 --> 00:00:17,107 I am an ecologist, and I am a computer scientist. 6 00:00:17,107 --> 00:00:19,253 And I work with a lot of great people on this project 7 00:00:19,253 --> 00:00:21,750 who are my co-authors on this who helped to collect data, 8 00:00:21,750 --> 00:00:23,250 folks from Green Mountain National Forest, 9 00:00:23,250 --> 00:00:25,950 Vermont Fish Wildlife Service, 10 00:00:25,950 --> 00:00:28,560 as well as we received support from FEMC 11 00:00:28,560 --> 00:00:29,833 and Ruffed Grouse Society 12 00:00:29,833 --> 00:00:32,130 in pulling this project off. 13 00:00:32,130 --> 00:00:33,157 And so the title 14 00:00:33,157 --> 00:00:35,293 "Forest Monitoring for Early Successional Species," 15 00:00:35,293 --> 00:00:37,200 I could have said "Early Successional Bird Species," 16 00:00:37,200 --> 00:00:38,970 if you didn't figure out that was the direction 17 00:00:38,970 --> 00:00:40,260 this was going already. 18 00:00:40,260 --> 00:00:41,610 And this is an ongoing project. 19 00:00:41,610 --> 00:00:43,645 So I'm gonna give you a bit of a progress update 20 00:00:43,645 --> 00:00:46,290 on what we've done so far and where we're at, 21 00:00:46,290 --> 00:00:48,933 and what our plans are for what comes next. 22 00:00:50,160 --> 00:00:53,370 So first, just to place this in a particular bird species, 23 00:00:53,370 --> 00:00:54,900 the ruffed grouse. 24 00:00:54,900 --> 00:00:57,750 Anyone here ever see a ruffed grouse out in the woods? 25 00:00:57,750 --> 00:00:58,860 -Usually you hear 'em first -(students laughing) 26 00:00:58,860 --> 00:01:01,140 before you see 'em and they scare the heck out of ya. 27 00:01:01,140 --> 00:01:03,600 They flush out from under your feet, 28 00:01:03,600 --> 00:01:05,580 but maybe you don't see 'em as much as you used to. 29 00:01:05,580 --> 00:01:09,150 This is could be considered an early successional species. 30 00:01:09,150 --> 00:01:10,650 They really do well in habitats 31 00:01:10,650 --> 00:01:12,930 that have a mixture of different forest ages. 32 00:01:12,930 --> 00:01:14,910 Young forest as well as old forest together is 33 00:01:14,910 --> 00:01:16,620 where you tend to get a lot of ruffed grouse. 34 00:01:16,620 --> 00:01:18,720 And as we all know, the Northeast has become 35 00:01:18,720 --> 00:01:19,980 much more forested over time. 36 00:01:19,980 --> 00:01:21,120 And so, we might expect 37 00:01:21,120 --> 00:01:23,550 that perhaps ruffed grouse would decline 38 00:01:23,550 --> 00:01:26,010 as a result of that change in habitat. 39 00:01:26,010 --> 00:01:27,900 This ruffed grouse I actually saw many years ago 40 00:01:27,900 --> 00:01:29,760 in Garrison, New York, where I don't think they are 41 00:01:29,760 --> 00:01:31,770 as common as they used to be, and it actually attacked me. 42 00:01:31,770 --> 00:01:33,906 It tried to, it tried to take off my fingers, 43 00:01:33,906 --> 00:01:35,040 -(students laughing) -which is a really cool thing 44 00:01:35,040 --> 00:01:35,873 about grouse. 45 00:01:35,873 --> 00:01:36,783 They can be really tame. 46 00:01:38,878 --> 00:01:40,500 But if you look at data to see, 47 00:01:40,500 --> 00:01:42,630 you know, how are ruffed grouse doing? 48 00:01:42,630 --> 00:01:44,130 It tends to paint, at least in the northeast, 49 00:01:44,130 --> 00:01:46,170 a not so great picture. 50 00:01:46,170 --> 00:01:48,780 This is a graph of data from 1960 to present 51 00:01:48,780 --> 00:01:50,280 from the Burlington Christmas Bird Count, 52 00:01:50,280 --> 00:01:53,043 which is taking place for the 75th year this Sunday. 53 00:01:53,043 --> 00:01:53,907 [man in audience] Oh, wow. 54 00:01:53,907 --> 00:01:56,160 And what we see is that they were fairly regular, 55 00:01:56,160 --> 00:02:00,180 not common, but regular up until about 10 years ago. 56 00:02:00,180 --> 00:02:02,760 At which point they've become really, really hard to find. 57 00:02:02,760 --> 00:02:03,660 And this is a trend. 58 00:02:03,660 --> 00:02:06,030 I picked Burlington 'cause we're all here, 59 00:02:06,030 --> 00:02:07,950 virtually or in person, 60 00:02:07,950 --> 00:02:09,937 but this trend is consistent across 61 00:02:09,937 --> 00:02:13,980 other Christmas bird counts throughout the region. 62 00:02:13,980 --> 00:02:15,150 We look at other data streams, 63 00:02:15,150 --> 00:02:17,430 such as the Breeding Bird Survey, 64 00:02:17,430 --> 00:02:21,840 we see similar declines in ruffed grouse through time, 65 00:02:21,840 --> 00:02:23,280 but as we try to get more nuanced, 66 00:02:23,280 --> 00:02:25,110 this is a really new feature for eBird. 67 00:02:25,110 --> 00:02:26,970 If you're not birders and not familiar with eBird, 68 00:02:26,970 --> 00:02:27,960 it's a really cool resource. 69 00:02:27,960 --> 00:02:29,520 But they've put together all their data, 70 00:02:29,520 --> 00:02:33,780 I think from 2007 to 2020 to build these trend maps, 71 00:02:33,780 --> 00:02:35,670 which really give you a fine scaled vision 72 00:02:35,670 --> 00:02:37,620 of where these species are increasing 73 00:02:37,620 --> 00:02:38,520 and where they're decreasing. 74 00:02:38,520 --> 00:02:41,490 And it paints a much more nuanced picture for ruffed grouse. 75 00:02:41,490 --> 00:02:44,040 We see that, yeah, in huge swaths of the Northeast, 76 00:02:44,040 --> 00:02:46,390 they're declining, but in some parts farther to our north, 77 00:02:46,390 --> 00:02:48,960 they're actually increasing, which, you know, 78 00:02:48,960 --> 00:02:50,257 begs the question, 79 00:02:50,257 --> 00:02:53,010 "Could change in climate be in play here too?" 80 00:02:53,010 --> 00:02:54,360 So, as we try to get more fine-grained 81 00:02:54,360 --> 00:02:56,460 about what is going on with ruffed grouse, 82 00:02:56,460 --> 00:02:58,320 and we look more fine-grained in Vermont 83 00:02:58,320 --> 00:03:00,330 at the evidence from the Breeding Bird Atlas, 84 00:03:00,330 --> 00:03:01,860 we see that their decline in Vermont 85 00:03:01,860 --> 00:03:03,960 actually isn't consistent throughout the state despite 86 00:03:03,960 --> 00:03:06,341 what the previous slide made it seem like that. 87 00:03:06,341 --> 00:03:08,760 And they've really declined, declined a lot 88 00:03:08,760 --> 00:03:10,710 in the Champlain Valley and Connecticut River Valley, 89 00:03:10,710 --> 00:03:12,420 but in other parts of the state they've held steady 90 00:03:12,420 --> 00:03:15,210 or maybe even increased a teeny bit. 91 00:03:15,210 --> 00:03:18,450 So, and, if we broaden our our scope 92 00:03:18,450 --> 00:03:20,100 to other early successional species, 93 00:03:20,100 --> 00:03:21,960 we see a really mixed bag. 94 00:03:21,960 --> 00:03:25,350 Some species that have declined, you know, precipitously, 95 00:03:25,350 --> 00:03:27,060 others of which have held fairly steady 96 00:03:27,060 --> 00:03:29,280 and yet others of which have, have slightly increased. 97 00:03:29,280 --> 00:03:32,190 And obviously there's lots of factors at play. 98 00:03:32,190 --> 00:03:33,810 Forest type is one of those 99 00:03:33,810 --> 00:03:35,160 and habitat type is one of those. 100 00:03:35,160 --> 00:03:37,200 And these different species have all, you know, 101 00:03:37,200 --> 00:03:39,870 all sorts of different micro type habitats that they, 102 00:03:39,870 --> 00:03:40,770 that they can inhabit. 103 00:03:40,770 --> 00:03:42,603 So, you know, it begs the question, you know, 104 00:03:42,603 --> 00:03:45,630 what, what is the influence of forest, 105 00:03:45,630 --> 00:03:48,030 of forest type, early successional habitat, 106 00:03:48,030 --> 00:03:49,770 and in particular we look at it through the lens 107 00:03:49,770 --> 00:03:51,930 of forest management, on the abundance 108 00:03:51,930 --> 00:03:53,490 and the presence of these species. 109 00:03:53,490 --> 00:03:55,710 So, that's the question that we're trying to answer. 110 00:03:55,710 --> 00:03:58,140 For, you know, what is the impact of timber harvest 111 00:03:58,140 --> 00:04:00,180 on the Avifauna, in particular, 112 00:04:00,180 --> 00:04:02,040 these early successional species 113 00:04:02,040 --> 00:04:04,470 that will come in after a cut? 114 00:04:04,470 --> 00:04:08,400 And the way we went about trying to answer this question 115 00:04:08,400 --> 00:04:10,350 is through an acoustic survey. 116 00:04:10,350 --> 00:04:13,470 So, starting in the late spring and early summer 117 00:04:13,470 --> 00:04:16,860 of this year, we deployed roughly 50 microphones. 118 00:04:16,860 --> 00:04:19,830 They were Song Meter Micros from Wildlife Acoustics, 119 00:04:19,830 --> 00:04:21,570 and we, we deployed them throughout the, 120 00:04:21,570 --> 00:04:23,910 the Green Mountain National Forest, 121 00:04:23,910 --> 00:04:24,990 some in the Southern Green Mountains, 122 00:04:24,990 --> 00:04:26,430 some in the Northern Green Mountains. 123 00:04:26,430 --> 00:04:29,543 And that map represents all of those locations. 124 00:04:29,543 --> 00:04:31,260 And we left them out there 125 00:04:31,260 --> 00:04:33,480 as long as their batteries lasted, 126 00:04:33,480 --> 00:04:35,190 which was really through late fall. 127 00:04:35,190 --> 00:04:38,220 We got, because of the timing of, of the start of the study, 128 00:04:38,220 --> 00:04:39,747 we got a little bit of a late start for the, 129 00:04:39,747 --> 00:04:41,940 the really vocal period for some species, 130 00:04:41,940 --> 00:04:43,920 like ruffed grouse, but still within the window 131 00:04:43,920 --> 00:04:46,080 for other species like mourning warblers for example. 132 00:04:46,080 --> 00:04:47,700 But we left those units out 133 00:04:47,700 --> 00:04:49,350 through basically November, December. 134 00:04:49,350 --> 00:04:50,550 There's a couple of them still out there 135 00:04:50,550 --> 00:04:52,380 that have yet to be collected. 136 00:04:52,380 --> 00:04:55,200 And I'd say about two-thirds of these units are paired 137 00:04:55,200 --> 00:04:58,440 with trail cameras that were part of a preexisting study. 138 00:04:58,440 --> 00:04:59,730 So especially for these new sites, 139 00:04:59,730 --> 00:05:01,710 we really tried to target early successional species. 140 00:05:01,710 --> 00:05:03,990 So we get a mix of different habitat types, 141 00:05:03,990 --> 00:05:06,570 some of these acoustic recorders 142 00:05:06,570 --> 00:05:09,300 that are close to recent timber cuts, 143 00:05:09,300 --> 00:05:12,150 others that are in more well-established forest. 144 00:05:12,150 --> 00:05:17,150 And so we began collecting data, you know, just, you know, 145 00:05:17,850 --> 00:05:20,100 over the last month really. 146 00:05:20,100 --> 00:05:22,657 And this is sort of, well, let's just get a look from, 147 00:05:22,657 --> 00:05:24,780 you know, from the field. 148 00:05:24,780 --> 00:05:26,850 So this is the same site in the spring 149 00:05:26,850 --> 00:05:29,362 when we set the camera and in the fall... 150 00:05:29,362 --> 00:05:31,470 (or I say camera cause I do a lot of camera trap stuff too) 151 00:05:31,470 --> 00:05:33,660 We set the, the acoustic recorders in the spring; 152 00:05:33,660 --> 00:05:34,890 same site in the fall. 153 00:05:34,890 --> 00:05:36,030 It's hard to see, but there it is 154 00:05:36,030 --> 00:05:37,350 on the tree right there. 155 00:05:37,350 --> 00:05:38,370 Oh, I have a laser pointer, cool. 156 00:05:38,370 --> 00:05:39,360 -There's, -(students laughing) 157 00:05:39,360 --> 00:05:40,890 there's the acoustic recorder 158 00:05:40,890 --> 00:05:43,140 and there's the same recorder on the same tree. 159 00:05:43,140 --> 00:05:45,630 And so we tried to set these as close as we could, 160 00:05:45,630 --> 00:05:47,310 in most cases, to these recent cuts. 161 00:05:47,310 --> 00:05:48,810 The cut is a little hard to make out, 162 00:05:48,810 --> 00:05:50,130 but it's in the distance. 163 00:05:50,130 --> 00:05:52,800 You see where the tall trees sort of stop right here. 164 00:05:52,800 --> 00:05:53,820 This was the recent cut. 165 00:05:53,820 --> 00:05:56,310 So, we tried to place them as close as we could 166 00:05:56,310 --> 00:05:58,650 while also trying to find spots that were viable 167 00:05:58,650 --> 00:06:01,710 for this paired camera recorder setup. 168 00:06:01,710 --> 00:06:04,950 And, you know, in a lot of these regenerating stands, 169 00:06:04,950 --> 00:06:06,540 the brush is too thick to put one 170 00:06:06,540 --> 00:06:08,070 of these recorders right in the middle, 171 00:06:08,070 --> 00:06:09,510 'cause our trail cams will be useless, 172 00:06:09,510 --> 00:06:10,740 the vegetation is just too thick. 173 00:06:10,740 --> 00:06:12,210 So, we often think, you know, 174 00:06:12,210 --> 00:06:14,160 vocal repertoire is carried a little bit 175 00:06:14,160 --> 00:06:15,870 into the surrounding forest. 176 00:06:15,870 --> 00:06:18,450 So, so we placed them as close to the edges as we could, 177 00:06:18,450 --> 00:06:22,050 in most cases and, and let them sit. 178 00:06:22,050 --> 00:06:23,460 And so the, you know, 179 00:06:23,460 --> 00:06:24,709 that was the first part is just get, 180 00:06:24,709 --> 00:06:26,340 just get these devices out there 181 00:06:26,340 --> 00:06:27,690 so we can start collecting data. 182 00:06:27,690 --> 00:06:29,626 Oh, you gotta be kidding me. 183 00:06:29,626 --> 00:06:30,600 (woman chuckles) 184 00:06:30,600 --> 00:06:31,800 Okay, so we'll skip this one. 185 00:06:31,800 --> 00:06:33,503 What this was gonna show is just a picture 186 00:06:33,503 --> 00:06:37,800 of a recorder and a camera together on the same tree. 187 00:06:37,800 --> 00:06:39,270 So pretend, pretend that's, 188 00:06:39,270 --> 00:06:41,633 -what you're looking at. -(students laughing) 189 00:06:41,633 --> 00:06:43,800 -[woman in audience] Wow. -(students laugh) 190 00:06:43,800 --> 00:06:46,920 It worked on the test run, so I don't know. 191 00:06:46,920 --> 00:06:51,060 But, but, so anyways, so here's a recorder on a tree 192 00:06:51,060 --> 00:06:53,520 and then we had a data collection protocol 193 00:06:53,520 --> 00:06:55,216 because we had these, you know, 194 00:06:55,216 --> 00:06:56,520 this study's actually part 195 00:06:56,520 --> 00:06:59,370 of a larger regional monitoring effort, 196 00:06:59,370 --> 00:07:00,960 called the Northeast Wildlife Monitoring Network, 197 00:07:00,960 --> 00:07:02,220 and so we've adopted a 198 00:07:02,220 --> 00:07:03,870 standardized data collection protocol. 199 00:07:03,870 --> 00:07:04,703 And this is a picture 200 00:07:04,703 --> 00:07:06,960 of my buddy John Peckham collecting data. 201 00:07:06,960 --> 00:07:08,070 So it's a bit hard to make out, 202 00:07:08,070 --> 00:07:10,740 but on the tree here, here's a recorder 203 00:07:10,740 --> 00:07:13,080 and in his hand is his phone. 204 00:07:13,080 --> 00:07:15,510 We use Survey123, which a bunch 205 00:07:15,510 --> 00:07:16,360 of people are probably familiar with, 206 00:07:16,360 --> 00:07:19,140 is an Esri product that allows you to do digital forms 207 00:07:19,140 --> 00:07:20,217 that can be filled out from the field. 208 00:07:20,217 --> 00:07:23,940 And so John's filling out the survey form on Survey 123 209 00:07:23,940 --> 00:07:25,920 to collect all of the site characteristics, 210 00:07:25,920 --> 00:07:27,630 all of the site visit metadata, 211 00:07:27,630 --> 00:07:30,780 including what the equipment was, where we were, 212 00:07:30,780 --> 00:07:31,680 what the date was, 213 00:07:31,680 --> 00:07:33,240 what are the surrounding conditions, 214 00:07:33,240 --> 00:07:34,290 what's the condition of the equipment. 215 00:07:34,290 --> 00:07:35,970 So, we collect a lot of different types 216 00:07:35,970 --> 00:07:38,490 of metadata related to the site visit. 217 00:07:38,490 --> 00:07:42,990 And then, after John clicks send on that survey form, 218 00:07:42,990 --> 00:07:44,310 it, we use Webhooks, 219 00:07:44,310 --> 00:07:47,310 which are a tool that you can use to automate workflows. 220 00:07:47,310 --> 00:07:49,920 It's a Microsoft product, so it works well with SharePoint, 221 00:07:49,920 --> 00:07:52,620 which is our cloud storage, as well as Survey123. 222 00:07:52,620 --> 00:07:55,200 So when John clicks send on his survey form, 223 00:07:55,200 --> 00:07:58,950 the first thing that happens is a folder is created 224 00:07:58,950 --> 00:08:01,165 in the cloud, in SharePoint, 225 00:08:01,165 --> 00:08:03,540 and the name of that folder, that directory, 226 00:08:03,540 --> 00:08:05,700 is just a concatenation of the location name, 227 00:08:05,700 --> 00:08:06,810 the equipment name, and the date. 228 00:08:06,810 --> 00:08:08,100 So it makes it really easy for John, 229 00:08:08,100 --> 00:08:09,210 when he gets back from the field, 230 00:08:09,210 --> 00:08:11,610 to figure out where he uploads his data too. 231 00:08:11,610 --> 00:08:13,770 The next thing that happens is John gets an email, 232 00:08:13,770 --> 00:08:15,480 it says, you know, thank you for your field visit. 233 00:08:15,480 --> 00:08:18,000 And that email includes a link to that SharePoint folder 234 00:08:18,000 --> 00:08:19,530 where he can upload his data. 235 00:08:19,530 --> 00:08:22,530 And then finally, John's able to upload his data 236 00:08:22,530 --> 00:08:25,320 and into that folder is also copied this survey response, 237 00:08:25,320 --> 00:08:27,240 which puts all the site metadata 238 00:08:27,240 --> 00:08:30,300 from that data collection protocol in the field, 239 00:08:30,300 --> 00:08:32,640 all of the media data that he's collected from the device, 240 00:08:32,640 --> 00:08:35,370 all in one spot in a really neat organized way. 241 00:08:35,370 --> 00:08:39,540 And so this, this procedure's been worked out through the, 242 00:08:39,540 --> 00:08:40,820 through the Vermont co-op unit in, 243 00:08:40,820 --> 00:08:43,980 in the project that I'm involved with called AMMonitor, 244 00:08:43,980 --> 00:08:47,160 which is an R package that we are, it's been developed, 245 00:08:47,160 --> 00:08:49,470 we're continuing to improve it. 246 00:08:49,470 --> 00:08:50,880 And we're sort of branding it 247 00:08:50,880 --> 00:08:52,890 as a data management ecosystem. 248 00:08:52,890 --> 00:08:55,200 So it incorporates, you know, this, this first part, 249 00:08:55,200 --> 00:08:57,180 the remote data collection, a piece of the workflow, 250 00:08:57,180 --> 00:08:58,620 which is often neglected, 251 00:08:58,620 --> 00:08:59,730 but an important part for us 252 00:08:59,730 --> 00:09:01,260 to really make sure is well established 253 00:09:01,260 --> 00:09:02,760 so that we can scale this and look 254 00:09:02,760 --> 00:09:05,550 at a regional type model, where people, you know, 255 00:09:05,550 --> 00:09:07,680 in really spread out, disparate locations, 256 00:09:07,680 --> 00:09:09,900 can use the same protocol to collect data 257 00:09:09,900 --> 00:09:11,910 under the same, yeah, under the same format. 258 00:09:11,910 --> 00:09:13,620 And really be able to merge that data 259 00:09:13,620 --> 00:09:15,030 and look at the big picture 260 00:09:15,030 --> 00:09:16,830 and really alleviate some of those problems 261 00:09:16,830 --> 00:09:19,020 of having data in different formats collected 262 00:09:19,020 --> 00:09:20,130 by different people. 263 00:09:20,130 --> 00:09:21,990 So that's sort of the first part of the wheel that we, 264 00:09:21,990 --> 00:09:22,823 that we've just looked at, 265 00:09:22,823 --> 00:09:24,420 is the remote data collection piece. 266 00:09:24,420 --> 00:09:25,694 And then the cloud storage piece, 267 00:09:25,694 --> 00:09:27,360 currently we've been using SharePoint, 268 00:09:27,360 --> 00:09:29,310 but we're moving towards Amazon Web Services, 269 00:09:29,310 --> 00:09:31,260 which affords us great opportunities 270 00:09:31,260 --> 00:09:32,760 to apply models on our data, 271 00:09:32,760 --> 00:09:34,830 which we'll look at in a minute. 272 00:09:34,830 --> 00:09:36,406 Which brings us to the machine learning piece, 273 00:09:36,406 --> 00:09:38,610 because we collect data at volumes 274 00:09:38,610 --> 00:09:42,510 that's just far too vast for us to manually inspect, 275 00:09:42,510 --> 00:09:43,560 you know, one by one, 276 00:09:43,560 --> 00:09:45,150 so we do need to employ machine learning. 277 00:09:45,150 --> 00:09:47,400 And so, my talk will take us through, you know, 278 00:09:47,400 --> 00:09:50,067 these first three rungs of the wheel, but the, 279 00:09:50,067 --> 00:09:52,080 the real concept behind AMMonitor is 280 00:09:52,080 --> 00:09:54,870 that it's all these different aspects of the workflow 281 00:09:54,870 --> 00:09:57,360 that you need in a remote wireless monitoring project 282 00:09:57,360 --> 00:09:59,940 to take it from the, from the field to the file system, 283 00:09:59,940 --> 00:10:01,920 to the decision tools that you need 284 00:10:01,920 --> 00:10:03,570 in order to inform management decisions. 285 00:10:03,570 --> 00:10:05,790 So we're measure conservation benchmarks. 286 00:10:05,790 --> 00:10:08,877 So, so these later parts of the wheel are, 287 00:10:08,877 --> 00:10:10,290 are still to come. 288 00:10:10,290 --> 00:10:12,060 Right now we're still just collecting data, 289 00:10:12,060 --> 00:10:13,470 but I didn't wanna leave it 290 00:10:13,470 --> 00:10:15,360 just at the data collection part. 291 00:10:15,360 --> 00:10:17,130 I guess one more thing that I sort of said this, 292 00:10:17,130 --> 00:10:18,840 but to just visualize it, 293 00:10:18,840 --> 00:10:21,180 everybody who uses this AMMonitor management system, 294 00:10:21,180 --> 00:10:22,950 we have a number of different projects, 295 00:10:22,950 --> 00:10:24,450 they have the same underlying database 296 00:10:24,450 --> 00:10:25,283 which stores the data, 297 00:10:25,283 --> 00:10:27,060 they're using the same Survey123 form. 298 00:10:27,060 --> 00:10:29,550 So we have a bunch of independent projects 299 00:10:29,550 --> 00:10:31,530 that on their own are collecting data, 300 00:10:31,530 --> 00:10:33,660 but that data gets collected in the same format 301 00:10:33,660 --> 00:10:35,640 and then stored in the same location 302 00:10:35,640 --> 00:10:37,470 so that we can look at that data together 303 00:10:37,470 --> 00:10:38,700 from a regional perspective 304 00:10:38,700 --> 00:10:40,440 and start to build better models 305 00:10:40,440 --> 00:10:41,730 because we have just, you know, 306 00:10:41,730 --> 00:10:42,900 a greater wealth of data. 307 00:10:42,900 --> 00:10:45,270 And yeah, a species that might be rare in one region, 308 00:10:45,270 --> 00:10:46,530 might be more common in another. 309 00:10:46,530 --> 00:10:49,110 So you end up with, you know, better sample sizes 310 00:10:49,110 --> 00:10:50,310 for the species you're trying to model. 311 00:10:50,310 --> 00:10:51,150 It's just this concept 312 00:10:51,150 --> 00:10:53,490 of a regional network is really important. 313 00:10:53,490 --> 00:10:55,470 But, the regional network, up until this point, 314 00:10:55,470 --> 00:10:56,670 has been mostly trail cameras. 315 00:10:56,670 --> 00:11:00,210 This was really a pilot for incorporating acoustic data. 316 00:11:00,210 --> 00:11:02,490 And acoustic data comes with its own sets of challenges. 317 00:11:02,490 --> 00:11:06,150 So, yeah, onto the data analysis. 318 00:11:06,150 --> 00:11:10,290 So far, we've collected almost all of our 50 ish microphones 319 00:11:10,290 --> 00:11:12,090 from the previous slide that we're each recording 320 00:11:12,090 --> 00:11:15,870 about 37 minutes of audio per day, every day, 321 00:11:15,870 --> 00:11:17,280 for months until the batteries died. 322 00:11:17,280 --> 00:11:19,230 So if you add that up over the course of the season, 323 00:11:19,230 --> 00:11:20,490 we could be looking at, you know, 324 00:11:20,490 --> 00:11:24,150 upwards of 90 hours of recordings per site, 325 00:11:24,150 --> 00:11:26,850 multiply that by, you know, roughly 50 sites. 326 00:11:26,850 --> 00:11:29,490 It it would take you 130 days, 24 hours a day 327 00:11:29,490 --> 00:11:31,020 to listen to all these audio recordings. 328 00:11:31,020 --> 00:11:33,210 So obviously the, you know, that's not gonna work. 329 00:11:33,210 --> 00:11:35,610 We need a different approach in order to deal 330 00:11:35,610 --> 00:11:37,290 with this vast volume of data. 331 00:11:37,290 --> 00:11:41,041 And so, our first crack at this was to use BirdNET. 332 00:11:41,041 --> 00:11:42,870 Not many birders in, in the room, 333 00:11:42,870 --> 00:11:44,400 -so I'm not feeling optimistic -(audience laughs) 334 00:11:44,400 --> 00:11:45,510 about a lot of use going on, 335 00:11:45,510 --> 00:11:48,873 but has anyone used the, the app called Merlin? 336 00:11:50,070 --> 00:11:50,970 Wow, okay, so a lot 337 00:11:50,970 --> 00:11:52,350 -of non-birders using Merlin. -(audience laughing) 338 00:11:52,350 --> 00:11:53,760 Well that's all, well maybe that's your, 339 00:11:53,760 --> 00:11:55,650 your ticket into, into the birding world. 340 00:11:55,650 --> 00:11:57,210 It's a really cool app that basically, 341 00:11:57,210 --> 00:11:59,130 you can hold up your phone to the world, 342 00:11:59,130 --> 00:12:01,590 it'll record the, you know, whatever birds are singing 343 00:12:01,590 --> 00:12:03,210 and then in real time tell you 344 00:12:03,210 --> 00:12:04,200 what those birds are, 345 00:12:04,200 --> 00:12:06,270 and it actually works really well. 346 00:12:06,270 --> 00:12:07,800 I mean, occasionally it'll say you have, 347 00:12:07,800 --> 00:12:09,540 like, a Bar-headed goose flying overhead. 348 00:12:09,540 --> 00:12:10,797 -You know that's not right, -(laughing) 349 00:12:10,797 --> 00:12:12,360 but a lot of times it gets it right. 350 00:12:12,360 --> 00:12:13,830 And so, that was our first crack, 351 00:12:13,830 --> 00:12:15,330 was to use BirdNET to try to, 352 00:12:15,330 --> 00:12:17,880 you know, make this data manageable. 353 00:12:17,880 --> 00:12:20,880 And so just, you know, really, over the last couple days, 354 00:12:20,880 --> 00:12:22,950 we've gotten BirdNET running on a small sample 355 00:12:22,950 --> 00:12:25,590 of the data, really just to showcase to you all here today 356 00:12:25,590 --> 00:12:27,450 so that we can see what that looks like. 357 00:12:27,450 --> 00:12:29,340 And still to be determined, 358 00:12:29,340 --> 00:12:31,530 is how good are these BirdNET results? 359 00:12:31,530 --> 00:12:33,390 Like, how much can we really bank on this? 360 00:12:33,390 --> 00:12:35,070 And Angela really set it up well 361 00:12:35,070 --> 00:12:38,130 in terms of talking about like the, the false negatives 362 00:12:38,130 --> 00:12:39,420 or the, you know, the omissions 363 00:12:39,420 --> 00:12:42,360 where, yeah, the species, the focal species was there 364 00:12:42,360 --> 00:12:43,193 but you missed it. 365 00:12:43,193 --> 00:12:44,100 And so we still have to assess 366 00:12:44,100 --> 00:12:45,240 how well it's doing with that. 367 00:12:45,240 --> 00:12:47,550 But just to see, you know, what it looks like 368 00:12:47,550 --> 00:12:50,610 I ran BirdNET, which is, you know, the model 369 00:12:50,610 --> 00:12:53,520 that we looked at, it sort of takes a subset 370 00:12:53,520 --> 00:12:54,990 of all the species it can ID. 371 00:12:54,990 --> 00:12:57,662 It narrowed it down to just 200 species that it can predict, 372 00:12:57,662 --> 00:13:00,780 the species that are likely to be encountered in Vermont. 373 00:13:00,780 --> 00:13:04,151 And so I ran it on all 87 hours of data 374 00:13:04,151 --> 00:13:07,800 for one particular site and just looked at how it did 375 00:13:07,800 --> 00:13:10,200 in predicting the presence of certain species. 376 00:13:10,200 --> 00:13:12,510 It does give a confidence level 377 00:13:12,510 --> 00:13:15,180 in terms of how likely it is 378 00:13:15,180 --> 00:13:17,040 that it's detection is accurate. 379 00:13:17,040 --> 00:13:19,020 And so, I just filtered out any detections 380 00:13:19,020 --> 00:13:21,240 that had a confidence of 0.5 or above. 381 00:13:21,240 --> 00:13:23,010 It's on a zero to one scale. 382 00:13:23,010 --> 00:13:25,380 And so, you know, for woodcock, for example, 383 00:13:25,380 --> 00:13:27,390 these were the detections we have for American woodcock, 384 00:13:27,390 --> 00:13:30,309 which is another species that can tend to favor open areas. 385 00:13:30,309 --> 00:13:33,060 And so, in the very first couple days it picked up a bunch 386 00:13:33,060 --> 00:13:34,410 of woodcocks and I listened to these 387 00:13:34,410 --> 00:13:36,780 and confirmed they were actually woodcocks. 388 00:13:36,780 --> 00:13:38,250 And woodcock is a cool species 389 00:13:38,250 --> 00:13:41,430 because it actually, they tend to continually do, 390 00:13:41,430 --> 00:13:43,080 they continue doing their display, 391 00:13:43,080 --> 00:13:44,190 well into the breeding season. 392 00:13:44,190 --> 00:13:46,260 Way past the point where breeding is complete. 393 00:13:46,260 --> 00:13:48,780 And so, we were actually picking up woodcocks periodically 394 00:13:48,780 --> 00:13:51,107 throughout the season, which makes sense and, 395 00:13:51,107 --> 00:13:53,100 and is sort of what we'd expect. 396 00:13:53,100 --> 00:13:55,950 I haven't manually verified all of these yet 397 00:13:55,950 --> 00:13:57,000 'cuz this data is pretty new, 398 00:13:57,000 --> 00:13:58,290 and you'll notice there is a chunk 399 00:13:58,290 --> 00:14:01,350 where the model hasn't been run yet. 400 00:14:01,350 --> 00:14:04,230 So, we still have a few blanks, but so far, 401 00:14:04,230 --> 00:14:05,760 you know, pretty promising for woodcock. 402 00:14:05,760 --> 00:14:07,883 But how does it do with other species? 403 00:14:07,883 --> 00:14:09,570 I tried mourning warbler, 404 00:14:09,570 --> 00:14:11,430 another early successional 11th species 405 00:14:11,430 --> 00:14:12,540 that really comes in and likes 406 00:14:12,540 --> 00:14:14,880 that early regenerating forest. 407 00:14:14,880 --> 00:14:16,320 And we have, like, a lot 408 00:14:16,320 --> 00:14:17,640 of morning warbler detections, 409 00:14:17,640 --> 00:14:19,320 like right during the peak of their breeding season, 410 00:14:19,320 --> 00:14:21,270 and right around when they start getting quiet, 411 00:14:21,270 --> 00:14:22,800 we stop detecting them. 412 00:14:22,800 --> 00:14:24,540 Which, you know, it's a good sanity check. 413 00:14:24,540 --> 00:14:26,400 It's sort of makes sense; we'd expect that. 414 00:14:26,400 --> 00:14:27,810 I, again, I haven't verified all of these, 415 00:14:27,810 --> 00:14:29,740 but I did verify a bunch of them and, and it was accurate; 416 00:14:29,740 --> 00:14:32,010 they, they picked out the morning warblers. 417 00:14:32,010 --> 00:14:34,650 But how does it do with ruffed grouse? 418 00:14:34,650 --> 00:14:37,623 Grouse are species that, okay, do another hand raising. 419 00:14:37,623 --> 00:14:39,750 How, how many folks have heard ruffed grouse 420 00:14:39,750 --> 00:14:40,620 out in the woods? 421 00:14:40,620 --> 00:14:42,360 The sound that they make. 422 00:14:42,360 --> 00:14:43,620 I think of it as more, like, a, 423 00:14:43,620 --> 00:14:44,940 it's a feeling than something 424 00:14:44,940 --> 00:14:46,507 -that you hear. -(audience laughing) 425 00:14:46,507 --> 00:14:48,690 It sort of reverberates through your body. 426 00:14:48,690 --> 00:14:50,340 And so it was really uncertain. 427 00:14:50,340 --> 00:14:52,590 How well are these microphones gonna pick up grouse? 428 00:14:52,590 --> 00:14:54,990 How well are our algorithms gonna detect grouse? 429 00:14:54,990 --> 00:14:57,030 We did a test, we brought the microphones out 430 00:14:57,030 --> 00:14:59,220 before the study started to a forest 431 00:14:59,220 --> 00:15:00,420 where we knew there's plenty of grouse 432 00:15:00,420 --> 00:15:03,000 and recorded them and listening back on my headphones, 433 00:15:03,000 --> 00:15:04,380 I could not detect the grouse, 434 00:15:04,380 --> 00:15:05,780 but then when I looked at a spectrogram, there it was. 435 00:15:05,780 --> 00:15:08,310 And they're so ultra low frequency 436 00:15:08,310 --> 00:15:09,360 that I think sometimes you need 437 00:15:09,360 --> 00:15:12,990 -like a subwoofer. -(audience laughs) 438 00:15:12,990 --> 00:15:14,340 And our microphones, of course, 439 00:15:14,340 --> 00:15:16,350 because of the timing of the study, went out a little bit 440 00:15:16,350 --> 00:15:18,750 after, really, their peak drumming period, 441 00:15:18,750 --> 00:15:20,700 in like, you know, really late March 442 00:15:20,700 --> 00:15:22,049 through through late April. 443 00:15:22,049 --> 00:15:23,940 So, so we missed that real drumming period. 444 00:15:23,940 --> 00:15:25,830 So that was the question, is are we gonna get any grouse? 445 00:15:25,830 --> 00:15:28,710 And so BirdNET actually detected two grouse 446 00:15:28,710 --> 00:15:30,337 with accuracy over 0.5. 447 00:15:31,266 --> 00:15:32,970 The second, was actually just wind, 448 00:15:32,970 --> 00:15:35,250 like, wind can make a low pitch rumble. 449 00:15:35,250 --> 00:15:37,860 So, it got it wrong once and then the second... 450 00:15:37,860 --> 00:15:39,390 Oh, I have another species here. 451 00:15:39,390 --> 00:15:40,223 We'll we'll come back to that. 452 00:15:40,223 --> 00:15:42,870 The second one was a true detection. 453 00:15:42,870 --> 00:15:44,880 So this is the spectrogram. 454 00:15:44,880 --> 00:15:47,370 August 20th, 6:20 it picked out one ruffed grouse 455 00:15:47,370 --> 00:15:49,500 out of 87 hours of recordings. 456 00:15:49,500 --> 00:15:51,390 I'm gonna try to play it to see if we can hear it, 457 00:15:51,390 --> 00:15:54,121 but just pointing at the screens, all these little, 458 00:15:54,121 --> 00:15:57,210 little signals here are grouse. 459 00:15:57,210 --> 00:15:59,393 And it was actually, it wasn't the drumming. 460 00:16:01,530 --> 00:16:02,363 [man in audience] Hm. 461 00:16:06,420 --> 00:16:07,680 It's not playing particularly well 462 00:16:07,680 --> 00:16:08,527 through the speaker, is it? 463 00:16:08,527 --> 00:16:11,047 [woman in audience] You hear it. 464 00:16:12,540 --> 00:16:14,190 I think we need the subwoofer for this one too. 465 00:16:14,190 --> 00:16:15,450 I actually can't even really hear it, 466 00:16:15,450 --> 00:16:16,620 -but take my word. -(audience laughs) 467 00:16:16,620 --> 00:16:18,510 I can send a link to this recording after. 468 00:16:18,510 --> 00:16:19,920 And actually what you, what you can do is, 469 00:16:19,920 --> 00:16:21,960 especially if you're at home on your computer already, 470 00:16:21,960 --> 00:16:25,020 is go to Peterson Field Guide to bird sounds, 471 00:16:25,020 --> 00:16:28,491 look up ruffed grouse and look up the bark vocalization. 472 00:16:28,491 --> 00:16:29,700 Like this is, like, it's unusual. 473 00:16:29,700 --> 00:16:32,280 This isn't the drumming, I've never heard this myself. 474 00:16:32,280 --> 00:16:33,270 I haven't done a lot of birding, 475 00:16:33,270 --> 00:16:35,100 but it actually manages to pick out a ruffed grouse. 476 00:16:35,100 --> 00:16:35,970 So that's pretty cool. 477 00:16:35,970 --> 00:16:37,860 I think it's a nice proof of concept 478 00:16:37,860 --> 00:16:41,010 that BirdNET can detect our target species. 479 00:16:41,010 --> 00:16:42,390 And then, of course, white-throated sparrow, 480 00:16:42,390 --> 00:16:43,590 is just another sanity check. 481 00:16:43,590 --> 00:16:45,210 These are a much more vocal species, 482 00:16:45,210 --> 00:16:46,350 and they'll actually make noise. 483 00:16:46,350 --> 00:16:47,760 They'll do a lot of the call notes, 484 00:16:47,760 --> 00:16:50,430 which BirdNET's actually pretty good at identifying, 485 00:16:50,430 --> 00:16:51,480 during the fall migration. 486 00:16:51,480 --> 00:16:54,600 So, you see a boost for white-throated sparrow in the fall 487 00:16:54,600 --> 00:16:56,370 and we do see presence earlier in the season, 488 00:16:56,370 --> 00:16:57,780 which makes sense given the habitat. 489 00:16:57,780 --> 00:16:59,760 So, the sanity checks are working out, 490 00:16:59,760 --> 00:17:01,950 it's still a little bit to be determined how, 491 00:17:01,950 --> 00:17:03,450 how good these models are gonna be, 492 00:17:03,450 --> 00:17:06,196 but the next steps are really to run this BirdNET model 493 00:17:06,196 --> 00:17:08,010 on all of our data, 494 00:17:08,010 --> 00:17:10,793 all 130 days worth of audio recordings 495 00:17:10,793 --> 00:17:13,440 and they, these microphones will go out again next season 496 00:17:13,440 --> 00:17:16,680 so we can get that early golden period for ruffed grouse. 497 00:17:16,680 --> 00:17:19,070 And with that... 498 00:17:21,990 --> 00:17:23,250 With that, I think I'm just gonna wrap up 499 00:17:23,250 --> 00:17:24,600 and we'll have plenty of time for questions. 500 00:17:24,600 --> 00:17:26,910 So, just thank you to our cooperators on this project; 501 00:17:26,910 --> 00:17:27,990 Vermont Fish and Wildlife, 502 00:17:27,990 --> 00:17:29,640 Green Mountain National Forest, 503 00:17:29,640 --> 00:17:31,350 received support from Ruffed Grouse Society 504 00:17:31,350 --> 00:17:35,760 and FEMC and also to the USGS Cloud Hosting Solutions, 505 00:17:35,760 --> 00:17:37,167 which provided a lot of the, 506 00:17:37,167 --> 00:17:39,130 the computing resources and technical support to, 507 00:17:39,130 --> 00:17:40,860 to get BirdNET up and running. 508 00:17:40,860 --> 00:17:44,040 And I'm sure I forgot somebody, but well, you all. 509 00:17:44,040 --> 00:17:44,873 -I forgot you all. -(audience laughing) 510 00:17:44,873 --> 00:17:45,706 Thank you all. 511 00:17:46,890 --> 00:17:47,903 Yeah. 512 00:17:47,903 --> 00:17:51,070 (audience applauding) 513 00:17:53,340 --> 00:17:54,173 Mm hm? 514 00:17:54,173 --> 00:17:55,470 [man in audience] What's recent? 515 00:17:55,470 --> 00:17:56,303 Excuse me? 516 00:17:56,303 --> 00:17:58,140 [man in audience] What is recent? 517 00:17:58,140 --> 00:17:59,370 Recent cuts? 518 00:17:59,370 --> 00:18:02,460 Oh, what, okay, what is defined as a recent cut? 519 00:18:02,460 --> 00:18:04,200 Yeah, so that's a good question. 520 00:18:04,200 --> 00:18:05,550 One of the harder parts 521 00:18:05,550 --> 00:18:07,320 of this was actually finding good places 522 00:18:07,320 --> 00:18:10,440 where we could place these devices. 523 00:18:10,440 --> 00:18:11,910 Some of the, you know, when we say recent, 524 00:18:11,910 --> 00:18:13,110 some of these just from looking 525 00:18:13,110 --> 00:18:16,080 at the size of the stems could be 10 years ago. 526 00:18:16,080 --> 00:18:16,913 [man in audience] Oh yeah. 527 00:18:16,913 --> 00:18:19,920 So, this is still part of what we need to work out, 528 00:18:19,920 --> 00:18:22,260 is, step one is what is actually in these recordings. 529 00:18:22,260 --> 00:18:24,903 Step two is how do we better define the different, 530 00:18:26,040 --> 00:18:28,380 the different experimental and control groups 531 00:18:28,380 --> 00:18:29,340 that we're trying to assess 532 00:18:29,340 --> 00:18:31,650 because it's not like, like recent, 533 00:18:31,650 --> 00:18:32,910 like unless we all, you're right, 534 00:18:32,910 --> 00:18:36,840 unless they're all cut at the same exact time, you know, 535 00:18:36,840 --> 00:18:38,197 it's apples to oranges. 536 00:18:38,197 --> 00:18:39,030 [man in audience #2] So different species, 537 00:18:39,030 --> 00:18:40,980 like different phases of their regrowth. 538 00:18:40,980 --> 00:18:42,240 Yeah. And in some cases they're, 539 00:18:42,240 --> 00:18:44,310 and actually the one of the other talks from, 540 00:18:44,310 --> 00:18:45,630 from this morning's session, 541 00:18:45,630 --> 00:18:48,450 looking at using tools to assess different kinds of cuts. 542 00:18:48,450 --> 00:18:50,880 Like, you know, did they leave a couple large trees 543 00:18:50,880 --> 00:18:52,740 and or or was it, was it a clear cut? 544 00:18:52,740 --> 00:18:54,690 So, these are all, these are all the different types 545 00:18:54,690 --> 00:18:57,930 of co-variants that we're gonna need to work into. 546 00:18:57,930 --> 00:18:59,130 We're gonna also, like Angela, 547 00:18:59,130 --> 00:19:00,600 we're gonna use an occupancy framework 548 00:19:00,600 --> 00:19:02,820 to try to measure what these impacts are. 549 00:19:02,820 --> 00:19:05,100 And so we need to, we need to find a way 550 00:19:05,100 --> 00:19:06,270 to incorporate these co-variants. 551 00:19:06,270 --> 00:19:08,356 It's still to be determined, but you know, 552 00:19:08,356 --> 00:19:10,530 the other thing is that when you look at it on the map, 553 00:19:10,530 --> 00:19:12,990 sometimes once you get down there in the field, 554 00:19:12,990 --> 00:19:14,250 it gives you a much better impression. 555 00:19:14,250 --> 00:19:15,690 And so, one of the informative things 556 00:19:15,690 --> 00:19:17,370 about setting all these microphones is 557 00:19:17,370 --> 00:19:19,980 that we might need to be more diligent 558 00:19:19,980 --> 00:19:21,720 when we go back out to these same locations 559 00:19:21,720 --> 00:19:23,910 to redeploy them next spring, 560 00:19:23,910 --> 00:19:27,000 to also be taking some sort of forest measurement 561 00:19:27,000 --> 00:19:28,800 so that we can accurately depict 562 00:19:28,800 --> 00:19:29,760 what those co-variants are. 563 00:19:29,760 --> 00:19:31,650 Because to your point, "recent"; 564 00:19:31,650 --> 00:19:33,030 it should have the air quotes on it. 565 00:19:33,030 --> 00:19:35,040 -Right. Sure. Right. -"Recent' cuts. 566 00:19:35,040 --> 00:19:35,873 Yeah. 567 00:19:36,810 --> 00:19:37,643 Yup? 568 00:19:37,643 --> 00:19:39,510 So, when you mentioned, like, 569 00:19:39,510 --> 00:19:42,930 with the wind being tagged as a ruffed grouse, 570 00:19:42,930 --> 00:19:45,120 if there's noise that wasn't birds, 571 00:19:45,120 --> 00:19:46,380 like that it's getting picked up by the microphones, 572 00:19:46,380 --> 00:19:49,110 is BirdNET going to give you a prediction no matter what? 573 00:19:49,110 --> 00:19:52,710 Like, will it refuse to say "No bird", or... 574 00:19:52,710 --> 00:19:55,620 So, it's, BirdNET itself is a, 575 00:19:55,620 --> 00:19:57,030 it's like a global model that's trained 576 00:19:57,030 --> 00:19:59,783 on, I think like, 3000 bird species 577 00:19:59,783 --> 00:20:01,830 and supposedly, I think, there might be 578 00:20:01,830 --> 00:20:04,262 some non-avian classes that it can do as well. 579 00:20:04,262 --> 00:20:08,010 I mean from our perspective, if there's nothing there, 580 00:20:08,010 --> 00:20:09,390 then we just get in all data. 581 00:20:09,390 --> 00:20:10,800 So like, we have these recorders going 582 00:20:10,800 --> 00:20:13,290 at night sometimes too, so that we could pick up, 583 00:20:13,290 --> 00:20:15,660 you know, things like woodcocks or frogs 584 00:20:15,660 --> 00:20:16,530 or other things like that. 585 00:20:16,530 --> 00:20:18,180 So, BirdNET's not gonna do the frogs; 586 00:20:18,180 --> 00:20:20,340 we have to find another tool to do that. 587 00:20:20,340 --> 00:20:21,600 But if it's just a, 588 00:20:21,600 --> 00:20:22,620 like, a minute minute of silence 589 00:20:22,620 --> 00:20:24,130 or a minute of heavy rain, 590 00:20:24,130 --> 00:20:27,270 it's, it's just gonna have no detections. 591 00:20:27,270 --> 00:20:29,300 Rain and wind happen to be pretty easy to, 592 00:20:29,300 --> 00:20:31,500 to figure out when it comes to recordings 593 00:20:31,500 --> 00:20:33,540 because rain just has a really distinct form 594 00:20:33,540 --> 00:20:35,820 on the spectrogram and wind, it just looks 595 00:20:35,820 --> 00:20:37,590 like a steady noise throughout. 596 00:20:37,590 --> 00:20:38,910 Those are actually gonna be pretty easy 597 00:20:38,910 --> 00:20:40,327 without sophisticated, 598 00:20:40,327 --> 00:20:42,120 you know, convolutional neural networks 599 00:20:42,120 --> 00:20:43,590 for us to, to tease out. 600 00:20:43,590 --> 00:20:46,050 But it remains to be done and I think 601 00:20:46,050 --> 00:20:47,580 that definitely has, has an impact. 602 00:20:47,580 --> 00:20:50,040 Like, we can't, we can't consider like, 603 00:20:50,040 --> 00:20:52,317 like, if it's a windy day, we're our, 604 00:20:52,317 --> 00:20:54,750 our detection probability is gonna be lower too, right? 605 00:20:54,750 --> 00:20:57,600 That might be what you were trying to get at there. 606 00:20:57,600 --> 00:21:01,020 So, that does need to be taken into account as well. 607 00:21:01,020 --> 00:21:02,160 Oh, a lot of questions. 608 00:21:02,160 --> 00:21:03,030 Yeah, go ahead. 609 00:21:03,030 --> 00:21:04,140 [man in audience #3] Were there any species 610 00:21:04,140 --> 00:21:06,900 that you didn't expect to see that were detected? 611 00:21:06,900 --> 00:21:07,947 -Ruffed grouse. -(audience laughs) 612 00:21:07,947 --> 00:21:08,780 [man in audience #3] Okay. 613 00:21:08,780 --> 00:21:10,020 I mean, I knew that they were there. 614 00:21:10,020 --> 00:21:12,360 I just was skeptical whether BirdNET would be able 615 00:21:12,360 --> 00:21:14,490 to pick one out, and I was especially surprised 616 00:21:14,490 --> 00:21:16,800 that it new the bark vocalization; 617 00:21:16,800 --> 00:21:19,170 just 'cuz, you know, these algorithms tend to do better 618 00:21:19,170 --> 00:21:22,080 on species that they have a lot of training data 619 00:21:22,080 --> 00:21:22,950 that they are exposed to. 620 00:21:22,950 --> 00:21:25,440 This is not, it's not something people record a lot. 621 00:21:25,440 --> 00:21:27,660 So, I was really pleasantly surprised with that. 622 00:21:27,660 --> 00:21:29,430 But I've only just scratched the surface. 623 00:21:29,430 --> 00:21:31,740 We haven't run it on the full gamut of species 624 00:21:31,740 --> 00:21:34,110 that we expect and I'm anticipating 625 00:21:34,110 --> 00:21:36,660 that it'll do better on some species than others. 626 00:21:36,660 --> 00:21:37,777 The easy, the easy things are, 627 00:21:37,777 --> 00:21:40,350 you know, it's a true positive, that's easy. 628 00:21:40,350 --> 00:21:41,880 It detected a species, 629 00:21:41,880 --> 00:21:43,830 we confirm that it was right about that species. 630 00:21:43,830 --> 00:21:44,790 That's easy. 631 00:21:44,790 --> 00:21:46,320 False positives are easy too. 632 00:21:46,320 --> 00:21:48,987 We can know that it said the species was there, 633 00:21:48,987 --> 00:21:51,720 but then we listened to it and, and it wasn't accurate. 634 00:21:51,720 --> 00:21:54,510 But yeah, the false negatives are harder. 635 00:21:54,510 --> 00:21:57,780 Knowing what it missed means going through audio 636 00:21:57,780 --> 00:22:00,000 and just listening to it, where we don't know anything, 637 00:22:00,000 --> 00:22:01,110 we don't have any preconception 638 00:22:01,110 --> 00:22:02,970 about what BirdNET detected or not 639 00:22:02,970 --> 00:22:06,090 and how, especially for species that may be infrequent 640 00:22:06,090 --> 00:22:08,850 that are expected to be infrequent, you know, 641 00:22:08,850 --> 00:22:09,870 that is a real challenge. 642 00:22:09,870 --> 00:22:11,340 How, how do you know, 643 00:22:11,340 --> 00:22:12,420 how do you get an impression 644 00:22:12,420 --> 00:22:13,950 for how many false negatives there are 645 00:22:13,950 --> 00:22:14,783 that are getting through, 646 00:22:14,783 --> 00:22:16,320 which is important in your modeling framework 647 00:22:16,320 --> 00:22:17,760 if you wanna be able to take that into account. 648 00:22:17,760 --> 00:22:19,590 So, there's still a lot of work to be done. 649 00:22:19,590 --> 00:22:21,179 This is very pretty preliminary, 650 00:22:21,179 --> 00:22:24,417 but the early results seem promising. 651 00:22:25,336 --> 00:22:26,169 Yup? 652 00:22:27,059 --> 00:22:29,040 [man in audience #4] It looks like you're, 653 00:22:29,040 --> 00:22:32,220 maybe, for like a species and you're monitoring, 654 00:22:32,220 --> 00:22:34,140 like how are you dealing 655 00:22:34,140 --> 00:22:37,000 with different preference towards, 656 00:22:38,864 --> 00:22:41,814 like, distance from that cut... 657 00:22:41,814 --> 00:22:44,606 you might find a couple of grouse 658 00:22:44,606 --> 00:22:45,939 to, compared to, 659 00:22:46,862 --> 00:22:48,417 are you just putting a seat down or up 660 00:22:48,417 --> 00:22:49,880 and seeing what you get or... 661 00:22:51,564 --> 00:22:52,770 Yeah, so I think that, you know, 662 00:22:52,770 --> 00:22:54,420 each species is going to, basically, 663 00:22:54,420 --> 00:22:56,520 have its own occupancy model. 664 00:22:56,520 --> 00:22:58,980 So, I think right now we're, we're looking 665 00:22:58,980 --> 00:23:01,620 at that in that framework where, again, we're, 666 00:23:01,620 --> 00:23:03,853 I'm a little bit ahead of it here because, 667 00:23:03,853 --> 00:23:06,587 and my role is more so in the machine, you know, 668 00:23:06,587 --> 00:23:07,917 in the machine learning and, 669 00:23:07,917 --> 00:23:10,560 and work, data workflow pieces working. 670 00:23:10,560 --> 00:23:11,910 But I mean, to try to anticipate 671 00:23:11,910 --> 00:23:14,670 what we might do going forward is that, you know, 672 00:23:14,670 --> 00:23:16,620 each species could be looked at totally independently. 673 00:23:16,620 --> 00:23:18,510 So, we have one model for woodcock 674 00:23:18,510 --> 00:23:20,400 where we might find their responses 675 00:23:20,400 --> 00:23:22,680 to different co-variants differ from something 676 00:23:22,680 --> 00:23:25,650 like a ruffed grouse, which is sort of what we expect. 677 00:23:25,650 --> 00:23:27,090 I don't know if, looking at, 678 00:23:27,090 --> 00:23:29,037 it doesn't sound like I answered your question all the way. 679 00:23:29,037 --> 00:23:30,340 [man in audience #4] Well just... 680 00:23:30,340 --> 00:23:33,032 Like, the monitor says the location 681 00:23:33,032 --> 00:23:34,731 -of a single site, -Yes. 682 00:23:34,731 --> 00:23:35,570 [man in audience #4] But each species can vary 683 00:23:35,570 --> 00:23:40,570 and you can use that depending on how far or close they are. 684 00:23:41,250 --> 00:23:42,600 Yes. Yeah. 685 00:23:42,600 --> 00:23:44,460 So I, I mean I think the initial assumption, 686 00:23:44,460 --> 00:23:45,300 the simple assumption is 687 00:23:45,300 --> 00:23:47,370 that each species is acting independent 688 00:23:47,370 --> 00:23:48,240 from other species. 689 00:23:48,240 --> 00:23:50,790 So yeah, we would expect those behaviors to vary. 690 00:23:50,790 --> 00:23:53,490 It's not gonna be, I would be very surprised 691 00:23:53,490 --> 00:23:55,200 if our outcome or if our, yeah, 692 00:23:55,200 --> 00:23:57,060 if the outcomes from this, were that, you know, 693 00:23:57,060 --> 00:23:58,770 ruffed grouse behaved the same way 694 00:23:58,770 --> 00:24:01,113 to different forest disturbances as woodcock. 695 00:24:03,330 --> 00:24:04,163 [host] Thank you Larry. 696 00:24:04,163 --> 00:24:05,220 I think that's all we have time for. 697 00:24:05,220 --> 00:24:06,053 Okay. We can... 698 00:24:06,053 --> 00:24:07,800 I'm the last, so we can catch ourselves afterwards. 699 00:24:07,800 --> 00:24:08,915 Thank, thanks again everyone. 700 00:24:08,915 --> 00:24:12,082 (audience applauding)