1
00:00:00,000 --> 00:00:00,820
 

2
00:00:00,820 --> 00:00:03,973
 

3
00:00:07,000 --> 00:00:09,460
Hello everyone, I'm super
excited to be here today.

4
00:00:09,460 --> 00:00:10,660
I live in New York now

5
00:00:10,660 --> 00:00:12,520
but I grew up in Vermont,
I grew up in St. Johnsbury,

6
00:00:12,520 --> 00:00:15,820
so getting to come to Burlington
feels like a homecoming.

7
00:00:15,820 --> 00:00:17,530
This work that I'm gonna
talk to you about today

8
00:00:17,530 --> 00:00:19,540
was actually part of my undergrad thesis,

9
00:00:19,540 --> 00:00:22,000
but I'm continuing it as part
of my master's thesis now

10
00:00:22,000 --> 00:00:23,560
at SUNY ESF.

11
00:00:23,560 --> 00:00:24,861
So I'm just gonna get started and get

12
00:00:24,861 --> 00:00:26,500
the important thing out of the way.

13
00:00:26,500 --> 00:00:27,333
We're gonna talk about

14
00:00:27,333 --> 00:00:29,680
What is a Disturbance Detection Algorithm.

15
00:00:29,680 --> 00:00:30,640
So there's a whole bunch of

16
00:00:30,640 --> 00:00:32,590
disturbance detection
algorithms, or sometimes

17
00:00:32,590 --> 00:00:34,720
they're called change
detection algorithms, out there

18
00:00:34,720 --> 00:00:37,030
and they all operate on
the same basic principle,

19
00:00:37,030 --> 00:00:39,490
which is you take a big
time series of imagery,

20
00:00:39,490 --> 00:00:41,040
could be imagery from anywhere.

21
00:00:41,920 --> 00:00:44,410
You find whatever the baseline
condition of that imagery is

22
00:00:44,410 --> 00:00:46,900
and then you look for
deviations from that condition,

23
00:00:46,900 --> 00:00:49,840
and we mark those as
disturbances or changes.

24
00:00:49,840 --> 00:00:51,580
So all the algorithms that I work with

25
00:00:51,580 --> 00:00:52,840
use satellite imagery.

26
00:00:52,840 --> 00:00:54,760
Satellite imagery is the common input

27
00:00:54,760 --> 00:00:55,780
into these kinds of things.

28
00:00:55,780 --> 00:00:58,600
And specifically a lot of
them use Landsat imagery.

29
00:00:58,600 --> 00:01:00,580
So if you're not familiar
with the Landsat program,

30
00:01:00,580 --> 00:01:03,730
it's a satellite platform
that's been around

31
00:01:03,730 --> 00:01:05,380
since the early eighties,

32
00:01:05,380 --> 00:01:07,960
and it takes images on
a 16 day return interval

33
00:01:07,960 --> 00:01:09,370
of the entire globe, which means

34
00:01:09,370 --> 00:01:13,030
there's this absolute wealth
of publicly available images

35
00:01:13,030 --> 00:01:15,010
and that's what makes it so popular.

36
00:01:15,010 --> 00:01:18,100
So this here is an
example of a Landsat scene

37
00:01:18,100 --> 00:01:19,120
from the Pacific Northwest,

38
00:01:19,120 --> 00:01:21,400
and we're looking at the
central pixel of that image

39
00:01:21,400 --> 00:01:23,470
and that row there, that's,

40
00:01:23,470 --> 00:01:26,260
this is our time series
of values from that pixel.

41
00:01:26,260 --> 00:01:27,820
And you can see there's
a baseline condition.

42
00:01:27,820 --> 00:01:29,320
We've got a pretty steady state.

43
00:01:29,320 --> 00:01:30,850
There's a deviation from that condition.

44
00:01:30,850 --> 00:01:32,530
That's our change that we've marked.

45
00:01:32,530 --> 00:01:35,280
And you can see the recovery
back to this steady state.

46
00:01:36,640 --> 00:01:38,050
And you might notice that the,

47
00:01:38,050 --> 00:01:39,580
this is kind of a weird looking image.

48
00:01:39,580 --> 00:01:41,380
This isn't, doesn't
look like a photograph.

49
00:01:41,380 --> 00:01:43,300
It doesn't look like what
you would expect to see

50
00:01:43,300 --> 00:01:45,640
if you're looking at
the forest from space.

51
00:01:45,640 --> 00:01:47,320
And that's because these
are composite imagery.

52
00:01:47,320 --> 00:01:49,690
These satellites record
information not only

53
00:01:49,690 --> 00:01:52,210
about visual reflectance,
visual light reflectance

54
00:01:52,210 --> 00:01:54,850
but also infrared
reflectance and vegetation's

55
00:01:54,850 --> 00:01:57,040
a really strong reflector
of infrared light.

56
00:01:57,040 --> 00:01:58,930
And we can use that information

57
00:01:58,930 --> 00:02:01,630
to help us differentiate forests

58
00:02:01,630 --> 00:02:02,980
from other things that are green.

59
00:02:02,980 --> 00:02:07,150
And so all of the work that
you'll see me talk about today,

60
00:02:07,150 --> 00:02:08,470
or if you see other
people talking about this,

61
00:02:08,470 --> 00:02:10,690
they're probably using these composites

62
00:02:10,690 --> 00:02:12,460
that include that imagery.

63
00:02:12,460 --> 00:02:15,610
And we can use this
reflectance information

64
00:02:15,610 --> 00:02:17,710
and we can track changes
in forest canopy cover

65
00:02:17,710 --> 00:02:19,420
remotely out of wide scale.

66
00:02:19,420 --> 00:02:20,890
But then we can also look historically

67
00:02:20,890 --> 00:02:24,134
because we have this wealth
of images in the archive.

68
00:02:24,134 --> 00:02:26,500
And so this becomes a very powerful tool

69
00:02:26,500 --> 00:02:27,800
that we have in our hands.

70
00:02:30,040 --> 00:02:30,890
I have a clicker.

71
00:02:33,850 --> 00:02:35,620
Interest in this remote
monitoring technology

72
00:02:35,620 --> 00:02:37,810
is really growing in the forestry sector

73
00:02:37,810 --> 00:02:39,040
in the last couple of years.

74
00:02:39,040 --> 00:02:40,450
We're seeing like a skyrocketing

75
00:02:40,450 --> 00:02:42,200
and exponential growth in interest.

76
00:02:43,330 --> 00:02:45,460
And that's because the
demand for the benefits

77
00:02:45,460 --> 00:02:48,640
that forests provide is also growing.

78
00:02:48,640 --> 00:02:49,900
There's regulations being passed

79
00:02:49,900 --> 00:02:52,420
all across the United States
and all across the world

80
00:02:52,420 --> 00:02:54,340
that emphasize carbon management

81
00:02:54,340 --> 00:02:55,810
or set net zero emissions goals.

82
00:02:55,810 --> 00:02:57,340
So in New York State we have

83
00:02:57,340 --> 00:03:00,280
the Climate Leadership and
Community Protection Act

84
00:03:00,280 --> 00:03:02,860
that requires net zero emissions by 2050.

85
00:03:02,860 --> 00:03:04,270
There's a very similar legislation

86
00:03:04,270 --> 00:03:06,430
with very similar goals here in Vermont.

87
00:03:06,430 --> 00:03:10,330
And these legislations are
putting a lot of pressure

88
00:03:10,330 --> 00:03:11,920
on landowners

89
00:03:11,920 --> 00:03:16,330
to conserve their forest
lands and sequester

90
00:03:16,330 --> 00:03:18,830
increased amounts of carbon
on their forest lands.

91
00:03:21,640 --> 00:03:23,590
And these programs, just
like the longer standing

92
00:03:23,590 --> 00:03:25,150
forest certification programs

93
00:03:25,150 --> 00:03:26,860
that have been around for a long time

94
00:03:26,860 --> 00:03:29,200
require some kind of
monitoring and verification

95
00:03:29,200 --> 00:03:31,750
to ensure compliance with their goals.

96
00:03:31,750 --> 00:03:33,902
And the monitoring and
verification processes

97
00:03:33,902 --> 00:03:36,280
all of you probably know is expensive

98
00:03:36,280 --> 00:03:38,050
and it's very time consuming.

99
00:03:38,050 --> 00:03:41,200
And so remote monitoring becomes a really

100
00:03:41,200 --> 00:03:44,443
appealing alternative in that situation.

101
00:03:45,310 --> 00:03:47,500
And I just wanna give the caveat that like

102
00:03:47,500 --> 00:03:49,780
I understand that remote
monitoring technology

103
00:03:49,780 --> 00:03:52,210
is not everybody's favorite thing.

104
00:03:52,210 --> 00:03:53,710
It can be kind of a hot button issue

105
00:03:53,710 --> 00:03:55,630
when you start talking about
watching people from space.

106
00:03:55,630 --> 00:03:57,280
Some people don't like
that and I understand it.

107
00:03:57,280 --> 00:03:59,590
It's kind of like big brother-ish.

108
00:03:59,590 --> 00:04:02,800
But the point is, is
that technology exists,

109
00:04:02,800 --> 00:04:05,350
it's available online, anyone can use it,

110
00:04:05,350 --> 00:04:07,990
and it's gonna be a part
of the future of forestry.

111
00:04:07,990 --> 00:04:10,150
It's gonna be a part of
your SFI or your FSC audits,

112
00:04:10,150 --> 00:04:11,200
if it's not already.

113
00:04:11,200 --> 00:04:13,210
It's gonna be a part of
your conservation easements,

114
00:04:13,210 --> 00:04:14,170
if it's not already.

115
00:04:14,170 --> 00:04:15,880
It's gonna be a part of
carbon credit programs

116
00:04:15,880 --> 00:04:17,170
if it isn't already.

117
00:04:17,170 --> 00:04:19,930
And so if these tools
are gonna be out there,

118
00:04:19,930 --> 00:04:22,660
if these tools are being
used to assess our forests,

119
00:04:22,660 --> 00:04:25,300
it's important and it's
essential, I think,

120
00:04:25,300 --> 00:04:27,490
that they're designed from
a forestry perspective

121
00:04:27,490 --> 00:04:28,813
with landowners in mind.

122
00:04:30,190 --> 00:04:32,530
And if that alone isn't
enough to convince you

123
00:04:32,530 --> 00:04:33,880
that the remote monitoring techniques

124
00:04:33,880 --> 00:04:36,400
and these remote monitoring
tools are important,

125
00:04:36,400 --> 00:04:39,250
like Colby mentioned,
had a nice slide about,

126
00:04:39,250 --> 00:04:42,610
studies are proving that
these disturbance regimes

127
00:04:42,610 --> 00:04:44,200
in our forests are shifting.

128
00:04:44,200 --> 00:04:45,033
They're shifting now.

129
00:04:45,033 --> 00:04:46,330
They're expected to continue to shift.

130
00:04:46,330 --> 00:04:48,730
We're expecting increased
insect pest outbreaks,

131
00:04:48,730 --> 00:04:51,130
increased extreme weather,

132
00:04:51,130 --> 00:04:53,800
and that's gonna change the
disturbance regimes that we see.

133
00:04:53,800 --> 00:04:56,110
And we need a way to keep
track of those things

134
00:04:56,110 --> 00:04:58,090
quickly and accurately,

135
00:04:58,090 --> 00:05:01,510
so we can see how they're
affecting our forests.

136
00:05:01,510 --> 00:05:03,520
So putting all of that together

137
00:05:03,520 --> 00:05:06,340
it's clear that we
urgently need these tools

138
00:05:06,340 --> 00:05:08,650
and they need to be
accurate and efficient.

139
00:05:08,650 --> 00:05:10,420
And so we can monitor
how and when disturbance

140
00:05:10,420 --> 00:05:13,090
is taking place in our
forests and understand

141
00:05:13,090 --> 00:05:15,610
how it's affecting ecosystems
structure and functions

142
00:05:15,610 --> 00:05:18,820
and to inform our stewardship
actions in response.

143
00:05:18,820 --> 00:05:22,240
So like I said, oh, okay, there we go.

144
00:05:22,240 --> 00:05:25,420
We do have these monitoring
tools and they are powerful,

145
00:05:25,420 --> 00:05:27,460
but one of the things that
we're concerned about with them

146
00:05:27,460 --> 00:05:28,690
is they're they're not developed

147
00:05:28,690 --> 00:05:30,370
in a regionally specific way.

148
00:05:30,370 --> 00:05:31,540
So they're often developed

149
00:05:31,540 --> 00:05:35,110
with a national or global data set.

150
00:05:35,110 --> 00:05:37,390
And as we all know, there's
tremendous variation

151
00:05:37,390 --> 00:05:42,160
in forest cover types, vegetation species,

152
00:05:42,160 --> 00:05:44,320
and the disturbance
regimes in these areas.

153
00:05:44,320 --> 00:05:49,263
And so we don't wanna apply
national models to them.

154
00:05:50,980 --> 00:05:52,480
So just to demonstrate that point,

155
00:05:52,480 --> 00:05:54,253
I have two sets of images here.

156
00:05:55,480 --> 00:05:58,750
Perfect, so these images here
are from the Adirondack Park,

157
00:05:58,750 --> 00:06:00,190
which is outside of Newcomb, New York.

158
00:06:00,190 --> 00:06:01,810
And we're looking at a
shelter wood harvest.

159
00:06:01,810 --> 00:06:03,550
And these images over
here are from Oregon,

160
00:06:03,550 --> 00:06:05,050
just outside the Willamette Valley.

161
00:06:05,050 --> 00:06:07,570
And we're looking at a
variety of clearcut harvest.

162
00:06:07,570 --> 00:06:10,180
Each set of images are six years apart.

163
00:06:10,180 --> 00:06:11,950
And you could really see
that this shelter wood

164
00:06:11,950 --> 00:06:15,460
this signature of the shelter
wood disappears very quickly

165
00:06:15,460 --> 00:06:17,380
and it fades back into the landscape.

166
00:06:17,380 --> 00:06:19,774
Whereas the clear cuts,
so look specifically

167
00:06:19,774 --> 00:06:23,920
at this clear cut, six
years later looks like that.

168
00:06:23,920 --> 00:06:26,170
So they're much more persistent

169
00:06:26,170 --> 00:06:28,420
and they're much more
prominent in the landscape.

170
00:06:28,420 --> 00:06:29,253
And remember if we're looking

171
00:06:29,253 --> 00:06:30,460
at this from a satellite perspective,

172
00:06:30,460 --> 00:06:32,620
we're looking at 30 meter pixels.

173
00:06:32,620 --> 00:06:34,900
So this is a higher resolution image

174
00:06:34,900 --> 00:06:35,830
than the satellite sees.

175
00:06:35,830 --> 00:06:38,050
So you're gonna see this variation here

176
00:06:38,050 --> 00:06:40,600
is gonna be less apparent
to the satellite.

177
00:06:40,600 --> 00:06:43,960
And so what we wanted to see,
what we wanted to look at is

178
00:06:43,960 --> 00:06:48,700
due to these differences
in harvesting regimes,

179
00:06:48,700 --> 00:06:52,210
are they affecting the
accuracy of these algorithms?

180
00:06:52,210 --> 00:06:55,240
So we looked at three algorithms
that are commonly used.

181
00:06:55,240 --> 00:06:56,590
The continuous change detection

182
00:06:56,590 --> 00:06:59,290
and classification algorithm,
which is called CCDC,

183
00:06:59,290 --> 00:07:00,700
the Land Trendr algorithm.

184
00:07:00,700 --> 00:07:02,680
And then the landscape
change monitoring system

185
00:07:02,680 --> 00:07:04,900
is actually an ensemble data product

186
00:07:04,900 --> 00:07:08,317
of the outputs of CCDC and Land Trendr.

187
00:07:08,317 --> 00:07:10,390
And here's an example of
what the outputs look like

188
00:07:10,390 --> 00:07:11,740
from these algorithms.

189
00:07:11,740 --> 00:07:14,710
So each of those red pixels is
a 30 meter by 30 meter pixel

190
00:07:14,710 --> 00:07:16,330
where disturbance was detected

191
00:07:16,330 --> 00:07:18,550
over the 30 year time span of our study.

192
00:07:18,550 --> 00:07:20,500
And you can see that there's
a significant variation

193
00:07:20,500 --> 00:07:22,420
in the amount of
disturbance that's detected

194
00:07:22,420 --> 00:07:24,490
from each algorithm, but
also where the algorithms

195
00:07:24,490 --> 00:07:26,830
are detecting the disturbances.

196
00:07:26,830 --> 00:07:31,600
And we compared these, these
outputs from these algorithms

197
00:07:31,600 --> 00:07:35,170
to harvest records that were
provided very generously

198
00:07:35,170 --> 00:07:36,490
through our landowner partners.

199
00:07:36,490 --> 00:07:40,000
We have a lovely partnership
with the people at F&W Forestry

200
00:07:40,000 --> 00:07:42,190
and they gave us all of
their harvest records

201
00:07:42,190 --> 00:07:44,860
for 43,000 hectares of forest
land in the Adirondack Parks.

202
00:07:44,860 --> 00:07:48,040
That's a huge chunk of land
that we were able to look at.

203
00:07:48,040 --> 00:07:50,440
And we had the location
and timing of their harvest

204
00:07:50,440 --> 00:07:53,320
as well as the prescription
associated with the harvest.

205
00:07:53,320 --> 00:07:55,870
And we were able to compare
those to the outputs.

206
00:07:55,870 --> 00:07:57,820
And when we did that, we
saw something like this.

207
00:07:57,820 --> 00:08:01,660
So across the board, we
saw a higher level of,

208
00:08:01,660 --> 00:08:03,340
with the higher levels
of disturbance detection,

209
00:08:03,340 --> 00:08:05,650
we saw higher levels of accuracy.

210
00:08:05,650 --> 00:08:07,300
So I'm just gonna take a second to explain

211
00:08:07,300 --> 00:08:08,950
what you're seeing on the maps here.

212
00:08:08,950 --> 00:08:12,640
So all of this dark
blue color on the maps,

213
00:08:12,640 --> 00:08:13,690
those are true positives.

214
00:08:13,690 --> 00:08:16,090
That's where our maps
matched our harvest records.

215
00:08:16,090 --> 00:08:17,290
That's really good.

216
00:08:17,290 --> 00:08:19,237
And the rest of the blue, the

217
00:08:19,237 --> 00:08:22,390
the lighter blue color,
that's also a match.

218
00:08:22,390 --> 00:08:24,220
Those are true negative
detections we call them.

219
00:08:24,220 --> 00:08:25,990
So that's places where
there was no harvest

220
00:08:25,990 --> 00:08:27,970
and the algorithm wasn't
detecting anything.

221
00:08:27,970 --> 00:08:29,920
But then the things that
we really are focusing on

222
00:08:29,920 --> 00:08:31,810
are the errors, the things where we have

223
00:08:31,810 --> 00:08:33,910
a mismatch in the record.

224
00:08:33,910 --> 00:08:37,300
So the orangey color,
the darker orange color,

225
00:08:37,300 --> 00:08:39,910
those are areas inside
of the harvest polygons

226
00:08:39,910 --> 00:08:41,980
that were not marked by the algorithm.

227
00:08:41,980 --> 00:08:44,380
So we call those false
negatives or errors of omission.

228
00:08:44,380 --> 00:08:46,630
That may be a word that some
of you are more familiar with.

229
00:08:46,630 --> 00:08:49,750
And then the final one
are the false positives.

230
00:08:49,750 --> 00:08:51,340
So that's this yellow color.

231
00:08:51,340 --> 00:08:53,260
And so those are areas that are either

232
00:08:53,260 --> 00:08:56,020
outside of the polygons, we
don't have records for that

233
00:08:56,020 --> 00:08:58,273
or the timing doesn't match up.

234
00:08:59,200 --> 00:09:02,530
And so these two error classes,
these mismatch classes,

235
00:09:02,530 --> 00:09:04,840
are pretty interesting
and I think do a good job

236
00:09:04,840 --> 00:09:06,310
of illustrating the difficulties

237
00:09:06,310 --> 00:09:08,110
with using these programs in our region

238
00:09:08,110 --> 00:09:10,420
and also kind of using them in general.

239
00:09:10,420 --> 00:09:13,180
So with selection harvest and

240
00:09:13,180 --> 00:09:14,830
which is most of what we have here,

241
00:09:14,830 --> 00:09:16,480
there's a lot of variation
of what's happening

242
00:09:16,480 --> 00:09:18,820
inside the boundaries
of your harvest unit.

243
00:09:18,820 --> 00:09:19,720
Depending on the prescription,

244
00:09:19,720 --> 00:09:21,880
you might have different
proportions of leave trees.

245
00:09:21,880 --> 00:09:23,740
You have stream buffers,
you have areas that maybe

246
00:09:23,740 --> 00:09:25,990
are inaccessible to your
harvesting equipment.

247
00:09:25,990 --> 00:09:27,850
So you really can't
expect the algorithms to

248
00:09:27,850 --> 00:09:29,983
produce perfect polygon outlines.

249
00:09:31,090 --> 00:09:32,530
That's just not reasonable.

250
00:09:32,530 --> 00:09:34,270
And then on the other hand,

251
00:09:34,270 --> 00:09:36,040
we've got non harvests
related disturbances,

252
00:09:36,040 --> 00:09:38,830
natural disturbances, that
happen all of the time

253
00:09:38,830 --> 00:09:40,810
in and outside of our harvest units.

254
00:09:40,810 --> 00:09:42,940
And with this kind of reference data,

255
00:09:42,940 --> 00:09:46,210
you can't assess the
accuracy of those things.

256
00:09:46,210 --> 00:09:49,510
So they're false positives,
but that doesn't mean

257
00:09:49,510 --> 00:09:50,343
that they don't represent

258
00:09:50,343 --> 00:09:53,650
a real disturbance that
happened on the ground.

259
00:09:53,650 --> 00:09:55,180
And that's something that
we need to keep in mind

260
00:09:55,180 --> 00:09:56,320
when you're assessing these kind

261
00:09:56,320 --> 00:09:57,520
of products in this way.

262
00:09:58,600 --> 00:10:00,400
So of the three products
that we looked at here,

263
00:10:00,400 --> 00:10:03,250
the CCDC algorithm is the
one that kind of struggles

264
00:10:03,250 --> 00:10:05,050
in this area because it doesn't detect

265
00:10:05,050 --> 00:10:07,120
a very high proportion of disturbance.

266
00:10:07,120 --> 00:10:09,170
It doesn't fill these polygons very well,

267
00:10:10,630 --> 00:10:13,240
which sort of leaves it
behind the other two.

268
00:10:13,240 --> 00:10:15,430
The Land Trendr algorithm
especially is the one

269
00:10:15,430 --> 00:10:18,010
that does the best job
of filling the polygons.

270
00:10:18,010 --> 00:10:20,560
But you can see that it
also has the highest level

271
00:10:20,560 --> 00:10:23,110
of commission error, these
false positive errors.

272
00:10:23,110 --> 00:10:24,520
But again, I just wanna emphasize

273
00:10:24,520 --> 00:10:27,070
that being sensitive to disturbance means

274
00:10:27,070 --> 00:10:29,380
that it could be picking up
real things that are happening.

275
00:10:29,380 --> 00:10:31,680
We just don't have the
records to assess that.

276
00:10:32,620 --> 00:10:34,987
We also looked at detection
by harvest prescription

277
00:10:34,987 --> 00:10:38,620
because our, the shelter would harvest,

278
00:10:38,620 --> 00:10:39,850
selective harvest were the things

279
00:10:39,850 --> 00:10:41,860
that were represented
the most in our data set.

280
00:10:41,860 --> 00:10:44,920
And we hypothesized that they
were likely not available

281
00:10:44,920 --> 00:10:48,262
at that concentration in
the data that was used

282
00:10:48,262 --> 00:10:51,580
when validating this
algorithm in its development.

283
00:10:51,580 --> 00:10:52,810
And we thought that they would probably

284
00:10:52,810 --> 00:10:53,980
struggle to detect them.

285
00:10:53,980 --> 00:10:55,210
And that is what we saw.

286
00:10:55,210 --> 00:10:59,290
So on the X axis here, you
have our harvest categories

287
00:10:59,290 --> 00:11:02,110
and then the Y axis is the mean percent

288
00:11:02,110 --> 00:11:04,990
of the polygon area detected
in those categories.

289
00:11:04,990 --> 00:11:06,550
And you can see that across the board,

290
00:11:06,550 --> 00:11:10,540
we're detecting a fairly
low level of the harvest.

291
00:11:10,540 --> 00:11:13,570
We're like at the 30% level
at best for most things.

292
00:11:13,570 --> 00:11:15,880
The clear cut category we're
performing pretty well in

293
00:11:15,880 --> 00:11:17,080
but we expected that

294
00:11:17,080 --> 00:11:19,660
because those were likely
represented in the training data,

295
00:11:19,660 --> 00:11:23,200
they're persistent and
pervasive, easy to see.

296
00:11:23,200 --> 00:11:26,890
But with these other harvests,
we're just not performing up

297
00:11:26,890 --> 00:11:28,660
to the level that we
really need to be able to

298
00:11:28,660 --> 00:11:29,890
use them in this region.

299
00:11:29,890 --> 00:11:32,830
The Land Trendr algorithm,
which is the green bars

300
00:11:32,830 --> 00:11:34,540
is giving us the best result.

301
00:11:34,540 --> 00:11:36,280
And so that's something
we're taking forward

302
00:11:36,280 --> 00:11:37,603
into our future work.

303
00:11:38,500 --> 00:11:41,920
And then the final thing
that we looked at is the CCDC

304
00:11:41,920 --> 00:11:43,750
and Land Trendr algorithms
give you estimates

305
00:11:43,750 --> 00:11:47,170
of disturbance magnitude
with their outputs.

306
00:11:47,170 --> 00:11:50,200
And we were curious to see if
you could use that information

307
00:11:50,200 --> 00:11:53,440
to model the amount of
timber volume removed

308
00:11:53,440 --> 00:11:55,540
from the harvest.

309
00:11:55,540 --> 00:11:57,700
And we compared, so we compared
it to pulpwood removals

310
00:11:57,700 --> 00:11:59,560
and we found that the Land
Trendr algorithm actually

311
00:11:59,560 --> 00:12:02,050
does a fairly good job of
modeling that relationship

312
00:12:02,050 --> 00:12:04,030
which is something that is useful

313
00:12:04,030 --> 00:12:05,320
in making management decisions,

314
00:12:05,320 --> 00:12:09,220
knowing that you can use
those values to look at

315
00:12:09,220 --> 00:12:10,670
the intensity of the harvest.

316
00:12:12,880 --> 00:12:15,140
So just to recap everything that I've said

317
00:12:16,600 --> 00:12:18,130
we need the regional monitoring tool.

318
00:12:18,130 --> 00:12:20,230
Remote monitoring is gonna
be playing an important role

319
00:12:20,230 --> 00:12:21,520
in the future of forest management.

320
00:12:21,520 --> 00:12:23,740
So the tools need to suit the region

321
00:12:23,740 --> 00:12:25,900
and of the tools that we looked at

322
00:12:25,900 --> 00:12:27,280
there's significant room for improvement.

323
00:12:27,280 --> 00:12:29,740
But the Land Trendr
algorithm performed the best

324
00:12:29,740 --> 00:12:32,863
at detecting partial harvest
and estimating magnitudes.

325
00:12:34,210 --> 00:12:35,860
So finally, I would just like to thank

326
00:12:35,860 --> 00:12:37,660
the Climate Applied
Forest Research Institute

327
00:12:37,660 --> 00:12:39,610
that provided the funding for this work

328
00:12:39,610 --> 00:12:42,073
and also our landowner
partners, F&W Forestry.

329
00:12:43,652 --> 00:12:46,652
(audience applauds)

330
00:12:51,464 --> 00:12:52,797
Yeah, go ahead.

331
00:12:54,229 --> 00:12:57,372
[Audience Member] You mentioned
about 30% of the area of,

332
00:12:57,372 --> 00:13:01,333
30% of the harvest area was impacted with.

333
00:13:03,490 --> 00:13:07,570
Did you look at the, which the
number of harvests detected?

334
00:13:08,563 --> 00:13:12,670
Yeah, so we are looking
polygon to polygon and so

335
00:13:12,670 --> 00:13:14,500
I I don't have that data in these slides

336
00:13:14,500 --> 00:13:18,220
but we did look at the
number of like total misses,

337
00:13:18,220 --> 00:13:21,460
like how many polygons were
totally missed by the algorithm.

338
00:13:21,460 --> 00:13:26,460
And the Land Trendr ones
on the order of 10 or so.

339
00:13:26,710 --> 00:13:30,520
We're looking at like 170 harvests.

340
00:13:30,520 --> 00:13:35,520
The CCDC algorithm were more
in the 60 or 70 harvests

341
00:13:35,620 --> 00:13:36,670
totally missed.

342
00:13:36,670 --> 00:13:38,860
And for the purposes of what we looked at,

343
00:13:38,860 --> 00:13:42,670
we were talking about one
pixel in the harvest unit.

344
00:13:42,670 --> 00:13:44,260
If there's one pixel in the harvest unit,

345
00:13:44,260 --> 00:13:45,310
that's not a total miss,

346
00:13:45,310 --> 00:13:49,060
but from an operational
perspective, one 30 meter pixel

347
00:13:49,060 --> 00:13:51,855
in a harvest unit, that's kind of a miss.

348
00:13:51,855 --> 00:13:54,310
[Audience Member] And how easy would it be

349
00:13:54,310 --> 00:13:58,660
to confuse the false positives
that were the false positives

350
00:13:58,660 --> 00:14:01,660
with the kinda incomplete positives?

351
00:14:05,280 --> 00:14:09,460
Yeah, so you're, you're
saying how easy would it be

352
00:14:09,460 --> 00:14:11,770
if you're looking at a map like this

353
00:14:11,770 --> 00:14:15,010
to confuse pixels in an
area that there might

354
00:14:15,010 --> 00:14:18,943
have been a harvest with the detection?

355
00:14:19,877 --> 00:14:21,028
[Audience Member] that occurred.

356
00:14:21,028 --> 00:14:24,457
There's a, there's a single
that doesn't represent

357
00:14:24,457 --> 00:14:29,140
the area accurately, but it, we know cause

358
00:14:29,140 --> 00:14:31,180
of additional information
that it's accurate.

359
00:14:31,180 --> 00:14:36,180
But outside of an experiment
we, we look at the positive

360
00:14:36,183 --> 00:14:39,490
and we don't, we're asking
ourselves was there a disturbance

361
00:14:39,490 --> 00:14:40,447
or treatment here or not?

362
00:14:40,447 --> 00:14:43,940
And is there other information
that you can use to

363
00:14:43,940 --> 00:14:45,107
evaluate that?

364
00:14:45,960 --> 00:14:48,314
The one that is in fact positive is more

365
00:14:48,314 --> 00:14:50,164
like the be positive words

366
00:14:50,164 --> 00:14:51,507
Positive

367
00:14:51,507 --> 00:14:52,480
Or is that just

368
00:14:52,480 --> 00:14:55,690
Some of it has to do with user ability,

369
00:14:55,690 --> 00:14:57,190
how much time you spent
looking at the maps?

370
00:14:57,190 --> 00:14:58,810
So I spent a lot of time
looking at these maps

371
00:14:58,810 --> 00:15:02,290
and I can look at a map
that's from not my study area

372
00:15:02,290 --> 00:15:04,270
and I can make pretty educated guesses

373
00:15:04,270 --> 00:15:07,150
about what's a harvest
and what's not a harvest

374
00:15:07,150 --> 00:15:10,240
based on the spacial pattern
and the way it shows up.

375
00:15:10,240 --> 00:15:12,550
And that's a big piece of my master's work

376
00:15:12,550 --> 00:15:15,400
is doing spatial analysis
and pattern analysis

377
00:15:15,400 --> 00:15:17,800
of these things to try to figure out

378
00:15:17,800 --> 00:15:20,353
if we can attribute
harvest or non harvest.

379
00:15:21,490 --> 00:15:26,490
But for example, this,
so this area right here,

380
00:15:26,500 --> 00:15:29,230
if I flipped back over to this map,

381
00:15:29,230 --> 00:15:31,930
you'll see that it shows
up as a false positive.

382
00:15:31,930 --> 00:15:35,410
But because of the density
and the way it's grouped,

383
00:15:35,410 --> 00:15:36,243
it's a harvest.

384
00:15:36,243 --> 00:15:37,076
I've been there in person.

385
00:15:37,076 --> 00:15:39,040
The reason it's a false
positive on this map

386
00:15:39,040 --> 00:15:41,920
is because it was harvested
right before the land was sold.

387
00:15:41,920 --> 00:15:45,673
So it's not in our harvest
records, but it is a harvest.

388
00:15:49,480 --> 00:15:50,313
[Audience Member] Could you take a look

389
00:15:50,313 --> 00:15:52,630
at the accuracy in terms
of the timing when it,

390
00:15:52,630 --> 00:15:53,463
Yes.

391
00:15:53,463 --> 00:15:56,590
Yep. Yep, so it has to identify
it within the harvest window

392
00:15:56,590 --> 00:16:00,970
and we use a one year leniency
period on either side.

393
00:16:00,970 --> 00:16:01,960
If it's a three year harvest,

394
00:16:01,960 --> 00:16:04,000
any detection in the three years is good

395
00:16:04,000 --> 00:16:06,040
and then one on either side is good,

396
00:16:06,040 --> 00:16:08,320
but if it, the harvest was in 2017

397
00:16:08,320 --> 00:16:09,910
and the detection is in 2020,

398
00:16:09,910 --> 00:16:12,510
that's outside the window
and it's a false positive.

399
00:16:15,580 --> 00:16:16,600
[Audience Member] So I have two questions,

400
00:16:16,600 --> 00:16:19,330
the first is that there
seems to be that blue area

401
00:16:19,330 --> 00:16:21,055
in the upper right that
all of the different

402
00:16:21,055 --> 00:16:22,613
detection operations found.

403
00:16:22,613 --> 00:16:25,630
Is there something special
about that harvest area

404
00:16:25,630 --> 00:16:27,733
that you think made it easier for the,

405
00:16:27,733 --> 00:16:30,206
[Madeline] Are you talking
about this one or this one?

406
00:16:30,206 --> 00:16:31,460
[Audience Member] Oh, I
didn't even see that top one.

407
00:16:31,460 --> 00:16:32,735
I was talking about the lower one.

408
00:16:32,735 --> 00:16:34,658
[Madeline] This one?

409
00:16:34,658 --> 00:16:36,354
[Audience Member] on the right.

410
00:16:36,354 --> 00:16:37,299
That one.

411
00:16:37,299 --> 00:16:38,612
They all get that.

412
00:16:38,612 --> 00:16:41,780
Is there anything else
special about that one?

413
00:16:41,780 --> 00:16:43,600
It's technically a shelter wood.

414
00:16:43,600 --> 00:16:45,190
It's a very heavy shelter wood.

415
00:16:45,190 --> 00:16:46,690
This is on ESF property.

416
00:16:46,690 --> 00:16:48,100
This is an ESF research cut.

417
00:16:48,100 --> 00:16:50,290
So this is another place I've
spent a fair amount of time

418
00:16:50,290 --> 00:16:55,090
and I think that my guess for
why they pick it up better

419
00:16:55,090 --> 00:16:57,400
is because it's pretty heavy,

420
00:16:57,400 --> 00:17:01,780
but also because of the
nature of the property.

421
00:17:01,780 --> 00:17:04,870
Everything around it is
left like more alone,

422
00:17:04,870 --> 00:17:06,850
it sees a lot less traffic.

423
00:17:06,850 --> 00:17:09,760
The Huntington ESF property here

424
00:17:09,760 --> 00:17:10,990
is a working research forest,

425
00:17:10,990 --> 00:17:13,960
but the research activities
are fairly concentrated.

426
00:17:13,960 --> 00:17:15,390
That's my guess.

427
00:17:15,390 --> 00:17:17,110
[Audience Member] And I
guess the second question

428
00:17:17,110 --> 00:17:19,995
is how viable do you
think it is to train model

429
00:17:19,995 --> 00:17:22,114
that's specific to data from our region

430
00:17:22,114 --> 00:17:25,020
that might have better
prospects of of performance?

431
00:17:25,020 --> 00:17:26,650
So that's what I'm in in
the thick of right now,

432
00:17:26,650 --> 00:17:27,493
working on it.

433
00:17:29,353 --> 00:17:30,880
And the answer is,

434
00:17:30,880 --> 00:17:33,640
I think the answer depends
on what you consider

435
00:17:33,640 --> 00:17:37,840
to be an improvement and how
incremental an improvement is.

436
00:17:37,840 --> 00:17:41,170
One of the things that we're
working on is classification.

437
00:17:41,170 --> 00:17:43,390
And the classification
stuff I think is gonna be

438
00:17:43,390 --> 00:17:47,860
more regionally specific and
more effective regionally

439
00:17:47,860 --> 00:17:49,420
because that's kind of where we're,

440
00:17:49,420 --> 00:17:51,720
I think we're gonna be
able to make the gains.

441
00:17:52,840 --> 00:17:55,660
There's this like really
careful line you have to walk

442
00:17:55,660 --> 00:17:58,190
when you're doing
something like this between

443
00:17:59,170 --> 00:18:03,910
sensitivity and noise and
what side of that line

444
00:18:03,910 --> 00:18:07,060
do you wanna end up on? And
some of that's application.

445
00:18:07,060 --> 00:18:12,060
So if you are somebody
who's doing an audit,

446
00:18:12,070 --> 00:18:15,130
like a carbon credit audit,
maybe you have more time

447
00:18:15,130 --> 00:18:17,950
and more resources to chase down and

448
00:18:17,950 --> 00:18:20,590
human validate some of these
things on the other end,

449
00:18:20,590 --> 00:18:22,180
like you're using it to pick your spots

450
00:18:22,180 --> 00:18:23,680
where you wanna go out.

451
00:18:23,680 --> 00:18:25,030
You might have more time than somebody

452
00:18:25,030 --> 00:18:29,440
who's a consulting forester
working for a private landowner.

453
00:18:29,440 --> 00:18:30,490
Does that kind of thing make sense?

454
00:18:30,490 --> 00:18:33,010
So I think it depends on your application

455
00:18:33,010 --> 00:18:34,870
and what you hope to do with it.

456
00:18:34,870 --> 00:18:35,913
[Audience Member] Cool.

457
00:18:39,070 --> 00:18:40,695
Yeah. Aiden.

458
00:18:40,695 --> 00:18:44,278
(audience member question)

459
00:18:54,420 --> 00:18:55,720
I, I hope so.

460
00:18:55,720 --> 00:18:59,620
So I actually, with FEMC gave me a grant

461
00:18:59,620 --> 00:19:02,620
and I spent some time this
summer doing field work

462
00:19:02,620 --> 00:19:05,140
going to places we didn't
have harvest records for,

463
00:19:05,140 --> 00:19:05,980
going to the pixels

464
00:19:05,980 --> 00:19:08,080
and trying to identify
what happened there.

465
00:19:08,080 --> 00:19:10,330
And I'm hoping to use that information to

466
00:19:10,330 --> 00:19:13,813
help train another
tuning of the algorithm.

467
00:19:16,435 --> 00:19:18,860
[Audience Member] I love
the idea of going pixel.

468
00:19:21,790 --> 00:19:25,405
just practically, how
does somebody even begin

469
00:19:25,405 --> 00:19:27,988
to access the data and explore?

470
00:19:31,049 --> 00:19:34,750
So, the easiest thing for
anybody to do is to go

471
00:19:34,750 --> 00:19:38,140
to the LCMS website where
they have a data explorer.

472
00:19:38,140 --> 00:19:40,480
The product is created for
you and then you can look

473
00:19:40,480 --> 00:19:43,630
at wherever you like for
like the northern region,

474
00:19:43,630 --> 00:19:47,230
look at the data and you can
compare it to your stuff.

475
00:19:47,230 --> 00:19:50,380
If you want the data for yourself to use,

476
00:19:50,380 --> 00:19:53,470
The Land Trendr program
runs on Google Earth Engine.

477
00:19:53,470 --> 00:19:58,470
And so it, you can pretty
much run their default script

478
00:19:58,690 --> 00:20:00,850
without any coding knowledge.

479
00:20:00,850 --> 00:20:01,810
That's basically what I do.

480
00:20:01,810 --> 00:20:04,453
I like, my knowledge
of Python is minuscule,

481
00:20:05,830 --> 00:20:08,950
but they've got a very helpful user guide

482
00:20:08,950 --> 00:20:10,870
and then you can just run it
through Google Earth engine

483
00:20:10,870 --> 00:20:13,760
and it's a roster output
that, that you can have

484
00:20:14,860 --> 00:20:17,230
on your computer and do
whatever you want with it.

485
00:20:17,230 --> 00:20:20,560
So like it is a publicly
available open access tool

486
00:20:20,560 --> 00:20:21,610
for anybody to use.

487
00:20:21,610 --> 00:20:23,320
Anybody who wants to can make these maps,

488
00:20:23,320 --> 00:20:25,743
as long as you're willing
to read the instructions.

489
00:20:40,799 --> 00:20:43,799
(audience applauds)