1
00:00:04,190 --> 00:00:06,630
Hi everybody, Welcome to

2
00:00:06,630 --> 00:00:08,099
the System Science Friday

3
00:00:08,099 --> 00:00:09,359
noon seminar series.

4
00:00:09,359 --> 00:00:11,129
And today we're glad to

5
00:00:11,129 --> 00:00:14,265
have one of our graduates.

6
00:00:14,265 --> 00:00:17,339
David Percy
goes by Percy,

7
00:00:17,339 --> 00:00:19,650
and he's been

8
00:00:19,650 --> 00:00:22,049
an instructor or
assistant professor,

9
00:00:22,049 --> 00:00:23,925
whatever they call
you nowadays.

10
00:00:23,925 --> 00:00:26,759
Over in the
geology for many,

11
00:00:26,759 --> 00:00:29,535
many years with a focus on

12
00:00:29,535 --> 00:00:33,334
GIS instruction
and research.

13
00:00:33,334 --> 00:00:35,569
And I'll let him complete
his introduction.

14
00:00:35,569 --> 00:00:36,649
He can tell you lots more

15
00:00:36,649 --> 00:00:38,254
interesting things
about themselves.

16
00:00:38,254 --> 00:00:39,079
Take it away.

17
00:00:39,079 --> 00:00:40,655
Thank you, Wayne.

18
00:00:40,655 --> 00:00:43,279
I just want to
say how happy

19
00:00:43,279 --> 00:00:44,030
I am to be here in

20
00:00:44,030 --> 00:00:45,619
the System
Science Seminar.

21
00:00:45,619 --> 00:00:48,049
This is my true home.

22
00:00:48,049 --> 00:00:50,435
I'm in a lot of
ways at PSU,

23
00:00:50,435 --> 00:00:51,889
even though I have

24
00:00:51,889 --> 00:00:53,269
been in the geology
department for

25
00:00:53,269 --> 00:00:54,589
over 20 years and got

26
00:00:54,589 --> 00:00:57,319
my undergraduate
degree there as well.

27
00:00:57,319 --> 00:00:59,330
As pointed out,

28
00:00:59,330 --> 00:01:00,829
I've done a lot
of graduate work

29
00:01:00,829 --> 00:01:05,165
in systems science and
did a master's here,

30
00:01:05,165 --> 00:01:08,659
some just based
on coursework.

31
00:01:08,659 --> 00:01:11,165
And so I've been,

32
00:01:11,165 --> 00:01:13,800
I've been interested in

33
00:01:14,130 --> 00:01:17,200
all sorts of exotic

34
00:01:17,200 --> 00:01:18,925
computing as I
like to put it,

35
00:01:18,925 --> 00:01:21,639
with regards to a
cellular automata and

36
00:01:21,639 --> 00:01:22,870
fractals and chaos and

37
00:01:22,870 --> 00:01:24,850
artificial
intelligence course.

38
00:01:24,850 --> 00:01:26,350
And I did do

39
00:01:26,350 --> 00:01:29,169
my first neural
networks class in

40
00:01:29,169 --> 00:01:31,149
the system science
department right here

41
00:01:31,149 --> 00:01:34,389
with Georgia and
Doris back in 1991.

42
00:01:34,389 --> 00:01:36,895
And I ended up

43
00:01:36,895 --> 00:01:40,240
retaking that class
around 2015 or 2016,

44
00:01:40,240 --> 00:01:43,070
and it was
remarkably similar.

45
00:01:45,390 --> 00:01:50,019
So that's the cool thing

46
00:01:50,019 --> 00:01:51,879
is that we've had this,

47
00:01:51,879 --> 00:01:56,540
this AI and machine
learning stuff ANR in

48
00:01:56,540 --> 00:01:58,880
our toolkit for

49
00:01:58,880 --> 00:02:01,369
a long time in
system science.

50
00:02:01,369 --> 00:02:03,575
And it's just, it's fun.

51
00:02:03,575 --> 00:02:04,909
It's good to be able to

52
00:02:04,909 --> 00:02:06,605
apply it to
different things.

53
00:02:06,605 --> 00:02:08,810
And so yeah, so

54
00:02:08,810 --> 00:02:12,350
let's what I'm
doing with it with

55
00:02:12,350 --> 00:02:16,325
my work today is sort of

56
00:02:16,325 --> 00:02:18,200
piggybacking off of
something that I

57
00:02:18,200 --> 00:02:20,929
do in the geology
department.

58
00:02:20,929 --> 00:02:23,480
I'm a database guy and

59
00:02:23,480 --> 00:02:26,314
I've been dealing with
large databases for

60
00:02:26,314 --> 00:02:28,550
decades And I was doing

61
00:02:28,550 --> 00:02:30,605
medical research
for a long time

62
00:02:30,605 --> 00:02:31,954
and then I got
into geology.

63
00:02:31,954 --> 00:02:34,369
And so the
geologic database,

64
00:02:34,369 --> 00:02:36,290
of course, was one of

65
00:02:36,290 --> 00:02:37,160
the obvious things for

66
00:02:37,160 --> 00:02:38,675
me to get involved with.

67
00:02:38,675 --> 00:02:40,940
And I was actually
involved back in

68
00:02:40,940 --> 00:02:42,470
the early 2000s in

69
00:02:42,470 --> 00:02:43,610
getting the standard that

70
00:02:43,610 --> 00:02:44,585
we're talking about today.

71
00:02:44,585 --> 00:02:47,359
This particular
standard adopted

72
00:02:47,359 --> 00:02:49,594
by the state then.

73
00:02:49,594 --> 00:02:52,670
And then I've also
taken over as

74
00:02:52,670 --> 00:02:54,634
the framework
implementation lead

75
00:02:54,634 --> 00:02:56,975
for this dataset for
the state of Oregon.

76
00:02:56,975 --> 00:02:59,600
I'm really involved and

77
00:02:59,600 --> 00:03:01,759
they organs geologic map.

78
00:03:01,759 --> 00:03:03,740
And that's kinda,
That's actually how

79
00:03:03,740 --> 00:03:06,810
I got involved
with this stuff.

80
00:03:06,940 --> 00:03:10,220
So the takeaway from

81
00:03:10,220 --> 00:03:12,320
this whole presentation is

82
00:03:12,320 --> 00:03:13,955
right here on this slide.

83
00:03:13,955 --> 00:03:17,179
Oregon geology is mapped
at different scales

84
00:03:17,179 --> 00:03:20,555
in similar terrains.
Let's unpack that.

85
00:03:20,555 --> 00:03:24,904
Scales and mapping are
like one to 24,000,

86
00:03:24,904 --> 00:03:26,945
where one unit on your map

87
00:03:26,945 --> 00:03:28,280
represents 24,000

88
00:03:28,280 --> 00:03:29,750
of those units out
in real space,

89
00:03:29,750 --> 00:03:32,630
all the way out to serve
like 1.2000001 to 5

90
00:03:32,630 --> 00:03:34,580
million would be
at the scale of

91
00:03:34,580 --> 00:03:37,414
the Nash of the
North America.

92
00:03:37,414 --> 00:03:40,909
So different scales are

93
00:03:40,909 --> 00:03:42,349
useful for
different purposes

94
00:03:42,349 --> 00:03:43,790
and I talk about
that more later.

95
00:03:43,790 --> 00:03:46,580
But this part about
different scales in

96
00:03:46,580 --> 00:03:48,440
similar terrains
is the key that's

97
00:03:48,440 --> 00:03:50,704
actually going to help
us solve this problem.

98
00:03:50,704 --> 00:03:54,259
Because terrains or
different kinds of

99
00:03:54,259 --> 00:03:58,489
packages of
environmental variables.

100
00:03:58,489 --> 00:04:00,950
Mostly we're talking
here about geology,

101
00:04:00,950 --> 00:04:03,050
but it also has
to do with,

102
00:04:03,050 --> 00:04:05,570
well, the elevation and

103
00:04:05,570 --> 00:04:06,830
the slope and all
of that sort of

104
00:04:06,830 --> 00:04:08,165
stuff can come
into the what,

105
00:04:08,165 --> 00:04:10,024
what makes up a terrain.

106
00:04:10,024 --> 00:04:12,529
Sentinel data is
satellite data.

107
00:04:12,529 --> 00:04:13,880
As I've got a picture up

108
00:04:13,880 --> 00:04:16,174
here that are
multispectral.

109
00:04:16,174 --> 00:04:17,510
Now that's
important because

110
00:04:17,510 --> 00:04:19,609
most satellite
data actually

111
00:04:19,609 --> 00:04:21,440
only have four
bands of data,

112
00:04:21,440 --> 00:04:22,639
red, green, blue,

113
00:04:22,639 --> 00:04:25,114
and infrared, or
near infrared.

114
00:04:25,114 --> 00:04:29,060
And so this Sentinel data

115
00:04:29,060 --> 00:04:30,769
here is important
because it's

116
00:04:30,769 --> 00:04:33,350
actually 12 bands
to start with,

117
00:04:33,350 --> 00:04:36,124
and we actually, I end
up using ten later.

118
00:04:36,124 --> 00:04:39,499
But these extra bands

119
00:04:39,499 --> 00:04:40,789
of information are

120
00:04:40,789 --> 00:04:42,965
the key to this
whole thing.

121
00:04:42,965 --> 00:04:44,929
They're freely
available at a scale

122
00:04:44,929 --> 00:04:47,105
of 20 m. Now,

123
00:04:47,105 --> 00:04:48,919
a lot of you may
not be able to

124
00:04:48,919 --> 00:04:50,659
picture how big 20 m is.

125
00:04:50,659 --> 00:04:53,794
I just want to sort
of think about that.

126
00:04:53,794 --> 00:04:55,039
I don't really
have something

127
00:04:55,039 --> 00:04:57,215
to give you as a scale.

128
00:04:57,215 --> 00:05:00,770
But like a football
field is 100 m long,

129
00:05:00,770 --> 00:05:02,660
so kinda scale it to that.

130
00:05:02,660 --> 00:05:04,789
Maybe. That gives you

131
00:05:04,789 --> 00:05:06,679
a sense that the
resolution that we have is

132
00:05:06,679 --> 00:05:12,889
not down at really
high resolution

133
00:05:12,889 --> 00:05:15,364
that we can make really
detailed maps with.

134
00:05:15,364 --> 00:05:16,969
But it's good to
get a good cut

135
00:05:16,969 --> 00:05:18,995
at stuff and 20 m
is pretty good.

136
00:05:18,995 --> 00:05:20,270
So using the
areas that are

137
00:05:20,270 --> 00:05:21,980
mapped at high resolution.

138
00:05:21,980 --> 00:05:24,410
That are adjacent
to areas mapped at

139
00:05:24,410 --> 00:05:26,420
low resolution was
the insight that

140
00:05:26,420 --> 00:05:28,805
I've had to do this.

141
00:05:28,805 --> 00:05:32,539
I'm using adjacent
areas that are mapped

142
00:05:32,539 --> 00:05:36,184
at high resolution
as training areas

143
00:05:36,184 --> 00:05:39,410
to then predict what

144
00:05:39,410 --> 00:05:42,080
the geologic units would

145
00:05:42,080 --> 00:05:43,310
be in areas that are

146
00:05:43,310 --> 00:05:44,959
mapped at lower
resolution.

147
00:05:44,959 --> 00:05:46,369
And I'm using
machine learning

148
00:05:46,369 --> 00:05:48,845
to do the classification.

149
00:05:48,845 --> 00:05:50,990
And my systems approach

150
00:05:50,990 --> 00:05:52,714
to this taken ties
it all together.

151
00:05:52,714 --> 00:05:56,720
We were talking about
how is this system Z?

152
00:05:56,720 --> 00:05:58,430
And I almost
put in a slide,

153
00:05:58,430 --> 00:06:00,545
how is this system Z?

154
00:06:00,545 --> 00:06:02,359
To try to encourage all of

155
00:06:02,359 --> 00:06:03,875
us that are
giving toxin in

156
00:06:03,875 --> 00:06:05,479
the system side of
seminar and sort of

157
00:06:05,479 --> 00:06:07,549
point out the system
is Ynez of it.

158
00:06:07,549 --> 00:06:09,890
And the way I kinda feel

159
00:06:09,890 --> 00:06:11,989
like it just
me doing this,

160
00:06:11,989 --> 00:06:13,759
it makes it systemsy
because I am

161
00:06:13,759 --> 00:06:15,670
so like I just said,

162
00:06:15,670 --> 00:06:18,455
attack everything
as a systems lens.

163
00:06:18,455 --> 00:06:20,059
And so I'm working at this

164
00:06:20,059 --> 00:06:22,040
and multiple scales and

165
00:06:22,040 --> 00:06:27,709
multiple lines of domains.

166
00:06:27,709 --> 00:06:29,570
So we're working with
satellite data and

167
00:06:29,570 --> 00:06:31,519
geologic data and all
of that sort of stuff.

168
00:06:31,519 --> 00:06:34,789
So I think that's what
makes IT system Z is

169
00:06:34,789 --> 00:06:36,170
this working

170
00:06:36,170 --> 00:06:39,245
across all the
different boundaries.

171
00:06:39,245 --> 00:06:42,169
So there's another copy

172
00:06:42,169 --> 00:06:43,910
of the geologic
map of Oregon.

173
00:06:43,910 --> 00:06:47,029
And you will see

174
00:06:47,029 --> 00:06:49,280
shortly that
these areas over

175
00:06:49,280 --> 00:06:51,169
here in Western
Oregon are mapped in

176
00:06:51,169 --> 00:06:52,880
much greater detail than

177
00:06:52,880 --> 00:06:55,460
these areas out here
in Eastern Oregon.

178
00:06:55,460 --> 00:06:58,099
And you'll actually
see this on

179
00:06:58,099 --> 00:07:00,530
the map where there's
actual places where

180
00:07:00,530 --> 00:07:02,449
there's solid
straight lines

181
00:07:02,449 --> 00:07:04,235
which just don't look like

182
00:07:04,235 --> 00:07:07,399
the way that geology

183
00:07:07,399 --> 00:07:09,485
actually maps
out in reality.

184
00:07:09,485 --> 00:07:11,390
So here's a scale,

185
00:07:11,390 --> 00:07:12,980
here's a map of the scale

186
00:07:12,980 --> 00:07:14,389
of the existing mapping.

187
00:07:14,389 --> 00:07:16,009
So what I've
done here is I

188
00:07:16,009 --> 00:07:17,779
took the geologic map and

189
00:07:17,779 --> 00:07:19,969
every actual unit in

190
00:07:19,969 --> 00:07:22,270
the geologic map is

191
00:07:22,270 --> 00:07:24,654
tagged by what scale
it was mapped out.

192
00:07:24,654 --> 00:07:26,050
And I should have actually

193
00:07:26,050 --> 00:07:27,490
had a reference
mapping here that

194
00:07:27,490 --> 00:07:30,279
shows you where all of
these data come from.

195
00:07:30,279 --> 00:07:31,450
But the idea behind

196
00:07:31,450 --> 00:07:33,160
this map is that
we were using

197
00:07:33,160 --> 00:07:36,520
the best available data
for every location.

198
00:07:36,520 --> 00:07:39,520
So if we've got a one
in 2014 to scale map

199
00:07:39,520 --> 00:07:41,020
at this area here where

200
00:07:41,020 --> 00:07:42,775
I've got green,
we'll use that.

201
00:07:42,775 --> 00:07:47,514
If we've got a
one to 250 K map

202
00:07:47,514 --> 00:07:50,230
just adjacent to it,
then we'll use that.

203
00:07:50,230 --> 00:07:51,339
So you can see
that there's

204
00:07:51,339 --> 00:07:52,809
areas right next to

205
00:07:52,809 --> 00:07:54,040
each other where
there's stuff

206
00:07:54,040 --> 00:07:55,420
mapped in high resolution,

207
00:07:55,420 --> 00:07:56,739
right next to stuff mapped

208
00:07:56,739 --> 00:07:58,779
in low, lower resolution.

209
00:07:58,779 --> 00:08:00,489
Here I've put
the break for

210
00:08:00,489 --> 00:08:01,974
lower resolution at

211
00:08:01,974 --> 00:08:06,305
greater than 125
Thousands scale.

212
00:08:06,305 --> 00:08:10,084
And I've said 124 k
is our gold standard.

213
00:08:10,084 --> 00:08:12,590
And anything in the 3,125

214
00:08:12,590 --> 00:08:15,109
K is good, usable stuff.

215
00:08:15,109 --> 00:08:16,819
It's not as good
as the 24 k,

216
00:08:16,819 --> 00:08:17,990
but it could work for us

217
00:08:17,990 --> 00:08:20,029
if that's what we
need to fall back on.

218
00:08:20,029 --> 00:08:22,430
And what we'll see
here in these parts of

219
00:08:22,430 --> 00:08:23,780
Eastern Oregon
where I'm wiggling

220
00:08:23,780 --> 00:08:25,385
my mouse over right now,

221
00:08:25,385 --> 00:08:27,424
there's a lot of red.

222
00:08:27,424 --> 00:08:29,510
You can see all
these red areas

223
00:08:29,510 --> 00:08:31,220
are mapped in
low resolution.

224
00:08:31,220 --> 00:08:33,139
And there's a couple of
little postage stamps

225
00:08:33,139 --> 00:08:34,580
in here, green stuff.

226
00:08:34,580 --> 00:08:36,560
And then some nice swaths

227
00:08:36,560 --> 00:08:39,964
of yellow that we can,
we'll end up using.

228
00:08:39,964 --> 00:08:44,659
So this is the
motivation for

229
00:08:44,659 --> 00:08:46,940
this was actually Part of

230
00:08:46,940 --> 00:08:49,625
my framework
implementation role.

231
00:08:49,625 --> 00:08:52,159
I was talking to
the director of

232
00:08:52,159 --> 00:08:54,319
the Institute for Natural

233
00:08:54,319 --> 00:08:55,699
Resources and he was

234
00:08:55,699 --> 00:08:58,610
bemoaning the fact that
we don't have a map

235
00:08:58,610 --> 00:09:00,260
at the scale of
one to 100,000

236
00:09:00,260 --> 00:09:01,849
K for the entire state.

237
00:09:01,849 --> 00:09:04,099
And that we have
that we can

238
00:09:04,099 --> 00:09:06,485
do all sorts of better
climate modeling

239
00:09:06,485 --> 00:09:10,340
and biodiversity modelling

240
00:09:10,340 --> 00:09:11,539
and all of that
sort of stuff.

241
00:09:11,539 --> 00:09:14,329
So that was, the
motivation was,

242
00:09:14,329 --> 00:09:15,590
well, if we can
at least get to

243
00:09:15,590 --> 00:09:17,420
a scale of one to 100 K,

244
00:09:17,420 --> 00:09:20,269
then we could actually
satisfy the needs of

245
00:09:20,269 --> 00:09:22,249
the people that are
actually trying to use

246
00:09:22,249 --> 00:09:25,639
this map to do
modeling and stuff.

247
00:09:25,639 --> 00:09:28,445
I was originally, when I,

248
00:09:28,445 --> 00:09:29,989
when I first started
looking into this,

249
00:09:29,989 --> 00:09:32,570
I ran into this paper
by cracked null,

250
00:09:32,570 --> 00:09:34,145
where he basically did

251
00:09:34,145 --> 00:09:35,600
the exact thing that
we're talking about.

252
00:09:35,600 --> 00:09:38,269
He used satellite data

253
00:09:38,269 --> 00:09:40,415
and machine
learning algorithms

254
00:09:40,415 --> 00:09:46,145
to produce geologic
maps in Tasmania.

255
00:09:46,145 --> 00:09:48,079
And his conclusion,

256
00:09:48,079 --> 00:09:50,210
after looking at all
of these methods,

257
00:09:50,210 --> 00:09:52,249
was the random
forest performed the

258
00:09:52,249 --> 00:09:54,605
best and was the
easiest to understand.

259
00:09:54,605 --> 00:09:56,600
As we know from

260
00:09:56,600 --> 00:09:59,359
trying to build
models of things.

261
00:09:59,359 --> 00:10:01,069
Some of these
things turn into

262
00:10:01,069 --> 00:10:03,245
black boxes that are
hard to understand.

263
00:10:03,245 --> 00:10:06,530
And so random
forest is nice.

264
00:10:06,530 --> 00:10:09,530
So I went ahead and
started there with

265
00:10:09,530 --> 00:10:11,419
the random forest
will just piggyback

266
00:10:11,419 --> 00:10:14,134
write-off of
practicals work here.

267
00:10:14,134 --> 00:10:16,564
And first we'll
take a look

268
00:10:16,564 --> 00:10:18,860
at the Sentinel data
that we're going to use.

269
00:10:18,860 --> 00:10:21,005
So these are the 12 bands

270
00:10:21,005 --> 00:10:23,359
of Sentinel data that
are coming down.

271
00:10:23,359 --> 00:10:28,264
It's it occupies each
place on the Earth,

272
00:10:28,264 --> 00:10:29,480
I think something
like every ten

273
00:10:29,480 --> 00:10:31,009
days or something
like that.

274
00:10:31,009 --> 00:10:32,929
So there's data at

275
00:10:32,929 --> 00:10:35,059
different times
of the year

276
00:10:35,059 --> 00:10:37,010
for particular
places which we'll

277
00:10:37,010 --> 00:10:39,754
see as important shortly.

278
00:10:39,754 --> 00:10:41,900
So the blue, green,

279
00:10:41,900 --> 00:10:43,489
and red are the
ones that we're all

280
00:10:43,489 --> 00:10:45,320
familiar with, red,
green, and blue.

281
00:10:45,320 --> 00:10:48,679
And then it's also
got this nice band

282
00:10:48,679 --> 00:10:50,119
called ultra blue for

283
00:10:50,119 --> 00:10:52,250
coastal and aerosol,
but that's a low,

284
00:10:52,250 --> 00:10:56,075
super low resolution
of only 60 m. We get,

285
00:10:56,075 --> 00:10:57,560
we pick up some
visible and near

286
00:10:57,560 --> 00:10:59,254
infrared bands in here.

287
00:10:59,254 --> 00:11:01,940
And then we pick
up a couple of

288
00:11:01,940 --> 00:11:03,364
these 60 meter bands that

289
00:11:03,364 --> 00:11:06,169
I end up dropping later.

290
00:11:06,169 --> 00:11:07,940
Then these two bands are

291
00:11:07,940 --> 00:11:10,124
the interesting ones
because these are bands

292
00:11:10,124 --> 00:11:11,750
11.12 are known to be

293
00:11:11,750 --> 00:11:14,569
useful for geologic
applications.

294
00:11:14,569 --> 00:11:17,149
So I know that those are
probably going to be

295
00:11:17,149 --> 00:11:19,699
important in my
model in trying to

296
00:11:19,699 --> 00:11:24,905
predict Geology from
satellite data.

297
00:11:24,905 --> 00:11:27,559
And the idea here

298
00:11:27,559 --> 00:11:29,974
is that with only 20
meter resolution,

299
00:11:29,974 --> 00:11:31,474
I'm not gonna be able
to actually make

300
00:11:31,474 --> 00:11:33,470
a really beautiful
geologic map

301
00:11:33,470 --> 00:11:34,744
is all detailed and stuff.

302
00:11:34,744 --> 00:11:36,169
But what I can

303
00:11:36,169 --> 00:11:37,940
do is shorten the
time that it takes to

304
00:11:37,940 --> 00:11:39,305
actually make the map by

305
00:11:39,305 --> 00:11:41,660
directing us to
outcrops where we're

306
00:11:41,660 --> 00:11:43,654
going to find exposures

307
00:11:43,654 --> 00:11:47,029
that will help us
with the mapping.

308
00:11:47,029 --> 00:11:51,200
So looking for a
pilot project,

309
00:11:51,200 --> 00:11:53,989
I wanted a place that
was not too far from

310
00:11:53,989 --> 00:11:55,970
roads with adjacent quads

311
00:11:55,970 --> 00:11:58,024
mapped at very
different scales.

312
00:11:58,024 --> 00:12:00,020
For reference, we
can see there's

313
00:12:00,020 --> 00:12:02,449
Eugene right there
and there's Ben.

314
00:12:02,449 --> 00:12:03,350
And so we know
that there's

315
00:12:03,350 --> 00:12:04,519
a nice road that goes

316
00:12:04,519 --> 00:12:06,830
through here from
Ben to Burns.

317
00:12:06,830 --> 00:12:10,355
And so if I needed
to go down here,

318
00:12:10,355 --> 00:12:12,424
I could do this drive.

319
00:12:12,424 --> 00:12:13,850
It takes about maybe

320
00:12:13,850 --> 00:12:15,739
two-and-a-half 3
h to get here.

321
00:12:15,739 --> 00:12:17,315
So this is doable.

322
00:12:17,315 --> 00:12:20,254
So now that being said,

323
00:12:20,254 --> 00:12:21,799
I haven't actually
gone out and

324
00:12:21,799 --> 00:12:23,179
collected the samples

325
00:12:23,179 --> 00:12:24,514
that I thought
I was going to.

326
00:12:24,514 --> 00:12:26,329
But this is this is

327
00:12:26,329 --> 00:12:28,774
the area that I
decided to focus on.

328
00:12:28,774 --> 00:12:31,054
And because
I've got these,

329
00:12:31,054 --> 00:12:32,554
I got a whole bunch of red

330
00:12:32,554 --> 00:12:34,099
around here and then
I've got a couple

331
00:12:34,099 --> 00:12:35,509
of little sections of

332
00:12:35,509 --> 00:12:36,799
green that

333
00:12:36,799 --> 00:12:38,810
are yellow that are
mapped in here,

334
00:12:38,810 --> 00:12:40,889
so we'll use those.

335
00:12:40,889 --> 00:12:43,329
So zooming in here,

336
00:12:43,329 --> 00:12:45,310
we've got Dickerson flat.

337
00:12:45,310 --> 00:12:47,140
So I don't actually have

338
00:12:47,140 --> 00:12:50,005
a quad map overlaid on
this particular thing.

339
00:12:50,005 --> 00:12:52,030
We can see the
Dickerson and soldiers

340
00:12:52,030 --> 00:12:53,349
quads are down here

341
00:12:53,349 --> 00:12:54,955
in the middle
of the state.

342
00:12:54,955 --> 00:12:59,349
And Dickerson is
mapped right now,

343
00:12:59,349 --> 00:13:02,125
clearly at as one unit,

344
00:13:02,125 --> 00:13:04,270
one geologic map map

345
00:13:04,270 --> 00:13:06,220
unit for that entire area.

346
00:13:06,220 --> 00:13:07,705
Whereas right next door

347
00:13:07,705 --> 00:13:09,774
and soldiers gap quad.

348
00:13:09,774 --> 00:13:12,700
And this is what
we do in mapping,

349
00:13:12,700 --> 00:13:13,959
is we divide
things up into

350
00:13:13,959 --> 00:13:16,524
quadrangles and name
them and stuff.

351
00:13:16,524 --> 00:13:19,089
And in the soldiers gap

352
00:13:19,089 --> 00:13:20,860
quadrangle right
next door to it,

353
00:13:20,860 --> 00:13:23,420
we've got a bunch of

354
00:13:23,420 --> 00:13:25,640
different geologic
units mapped

355
00:13:25,640 --> 00:13:27,650
and we can see over
here to the right,

356
00:13:27,650 --> 00:13:30,050
this level of detail
continues on over there.

357
00:13:30,050 --> 00:13:31,430
So we can see
that we've got

358
00:13:31,430 --> 00:13:33,200
quite a few units
mapped here.

359
00:13:33,200 --> 00:13:36,094
What I've done is a
stratified sampling

360
00:13:36,094 --> 00:13:39,004
with 50 points
per polygon.

361
00:13:39,004 --> 00:13:41,150
And then I'm using that

362
00:13:41,150 --> 00:13:44,570
to extract the
satellite data.

363
00:13:44,570 --> 00:13:47,510
The map units from
soldiers gap turned out to

364
00:13:47,510 --> 00:13:51,019
be this list that
we see here.

365
00:13:51,019 --> 00:13:55,639
Basalts, day sites,
more volcanic rocks,

366
00:13:55,639 --> 00:13:58,039
and these, these are of

367
00:13:58,039 --> 00:14:00,590
interest to any
geologists.

368
00:14:00,590 --> 00:14:01,819
I see Nick is
here, so he'd

369
00:14:01,819 --> 00:14:04,770
probably is
appreciating this part.

370
00:14:04,770 --> 00:14:08,800
So what I'm trying to do

371
00:14:08,800 --> 00:14:10,465
then is find these units

372
00:14:10,465 --> 00:14:11,890
in indifference and flat,

373
00:14:11,890 --> 00:14:13,809
or at least train up a
model that's going to

374
00:14:13,809 --> 00:14:14,889
potentially find these

375
00:14:14,889 --> 00:14:16,960
units in Dickerson flat.

376
00:14:16,960 --> 00:14:19,794
And so we're gonna
go ahead and use

377
00:14:19,794 --> 00:14:21,490
a supervised
classification

378
00:14:21,490 --> 00:14:22,779
and random forests.

379
00:14:22,779 --> 00:14:26,290
So here's a
brief depiction

380
00:14:26,290 --> 00:14:29,049
of what happens with
random forests.

381
00:14:29,049 --> 00:14:31,179
So we divide the space

382
00:14:31,179 --> 00:14:34,449
up with some sort of
a partition and then

383
00:14:34,449 --> 00:14:36,249
continue partitioning
and keep

384
00:14:36,249 --> 00:14:39,279
partitioning all the
way down until we

385
00:14:39,279 --> 00:14:41,800
get to the level of depth

386
00:14:41,800 --> 00:14:44,740
that we have specified
in the model.

387
00:14:44,740 --> 00:14:48,694
And then we can examine

388
00:14:48,694 --> 00:14:54,305
the tree for the random
forest and see how,

389
00:14:54,305 --> 00:14:56,329
we can see how the
actual choices

390
00:14:56,329 --> 00:14:57,649
are being made at

391
00:14:57,649 --> 00:14:59,149
each particular
break and make

392
00:14:59,149 --> 00:15:02,069
sure that they actually
make sense to us.

393
00:15:02,260 --> 00:15:06,185
So I'm dividing
my data up into

394
00:15:06,185 --> 00:15:09,380
70% training data,
30% test data,

395
00:15:09,380 --> 00:15:11,000
which is pretty standard

396
00:15:11,000 --> 00:15:13,519
statistical modeling
methodology

397
00:15:13,519 --> 00:15:14,780
you can tell because
I just pulled

398
00:15:14,780 --> 00:15:16,249
this graphics straight off

399
00:15:16,249 --> 00:15:18,334
of somebody
else's website.

400
00:15:18,334 --> 00:15:20,480
Then in Python, I'm taking

401
00:15:20,480 --> 00:15:22,310
my soldiers gap data
and splitting it

402
00:15:22,310 --> 00:15:24,875
into training and
testing sets and

403
00:15:24,875 --> 00:15:26,375
modeling it using

404
00:15:26,375 --> 00:15:28,520
scikit-learn has
various methods,

405
00:15:28,520 --> 00:15:29,810
mostly focusing on

406
00:15:29,810 --> 00:15:32,344
support vector and
random forest.

407
00:15:32,344 --> 00:15:33,860
And then the support,

408
00:15:33,860 --> 00:15:36,349
the soldiers gap data are

409
00:15:36,349 --> 00:15:39,860
then validated using
model metrics.

410
00:15:39,860 --> 00:15:41,479
And once a model

411
00:15:41,479 --> 00:15:43,655
has reasonably
high accuracy,

412
00:15:43,655 --> 00:15:45,829
I go ahead and fit that to

413
00:15:45,829 --> 00:15:48,200
new data from
Dickerson flat.

414
00:15:48,200 --> 00:15:51,064
Export that to
my GIS software,

415
00:15:51,064 --> 00:15:52,190
join it back up to

416
00:15:52,190 --> 00:15:54,950
the original data
and then display

417
00:15:54,950 --> 00:15:56,569
the map and convert to

418
00:15:56,569 --> 00:16:00,064
raster for
display purposes.

419
00:16:00,064 --> 00:16:06,065
And it turned out
in the modeling.

420
00:16:06,065 --> 00:16:09,515
That bands 11.12
did turn out to be

421
00:16:09,515 --> 00:16:11,840
a super high
importance and

422
00:16:11,840 --> 00:16:14,629
banned for which is this,

423
00:16:14,629 --> 00:16:16,639
these go blue, green, red.

424
00:16:16,639 --> 00:16:18,439
So the red band from

425
00:16:18,439 --> 00:16:21,019
the visible spectrum
also was important.

426
00:16:21,019 --> 00:16:23,149
So that's kind
of interesting.

427
00:16:23,149 --> 00:16:26,029
And then for the results,

428
00:16:26,029 --> 00:16:28,400
here it is, it looks
like a geologic map,

429
00:16:28,400 --> 00:16:30,620
if you know what a
geologic map looks like.

430
00:16:30,620 --> 00:16:33,679
So what we've got
remember let's see,

431
00:16:33,679 --> 00:16:37,220
let's go back to,
yeah, right there.

432
00:16:37,220 --> 00:16:38,495
So there's
difference in flat

433
00:16:38,495 --> 00:16:39,649
in its original all

434
00:16:39,649 --> 00:16:44,135
monolithic, homogeneous
single unit.

435
00:16:44,135 --> 00:16:47,090
Again. Here we've
got a couple

436
00:16:47,090 --> 00:16:50,150
of units poking
in from outside.

437
00:16:50,150 --> 00:16:56,105
And then here's my new
map with the results.

438
00:16:56,105 --> 00:16:57,920
And it's looking,

439
00:16:57,920 --> 00:16:59,329
it's looking
quite coherent.

440
00:16:59,329 --> 00:17:02,750
In fact, in the
sense that units

441
00:17:02,750 --> 00:17:06,349
that are grouped
together makes sense.

442
00:17:06,349 --> 00:17:09,739
And so from a
geological perspective,

443
00:17:09,739 --> 00:17:14,375
this map, actually,
it's pretty good.

444
00:17:14,375 --> 00:17:17,300
This is, so that's where

445
00:17:17,300 --> 00:17:19,490
we ended up
with at the end

446
00:17:19,490 --> 00:17:24,185
of the first round
of my processing.

447
00:17:24,185 --> 00:17:26,660
Then what I was
looking at 0.

448
00:17:26,660 --> 00:17:29,524
And then the idea is
to overlay that on

449
00:17:29,524 --> 00:17:31,459
Google Earth and find

450
00:17:31,459 --> 00:17:34,219
an actual outcrop
that we can go visit.

451
00:17:34,219 --> 00:17:36,680
This is my slide from

452
00:17:36,680 --> 00:17:39,019
last year when I presented

453
00:17:39,019 --> 00:17:41,254
the original work
and this was

454
00:17:41,254 --> 00:17:44,645
the quads to map
this summer.

455
00:17:44,645 --> 00:17:47,014
And that was for
the summer of 2022.

456
00:17:47,014 --> 00:17:48,679
So what happens in

457
00:17:48,679 --> 00:17:51,364
my department
is that we run

458
00:17:51,364 --> 00:17:53,899
a field mapping class in

459
00:17:53,899 --> 00:17:56,735
the summertime
and the meat,

460
00:17:56,735 --> 00:18:01,024
the main researcher
on that, Martin,

461
00:18:01,024 --> 00:18:04,069
martin picks

462
00:18:04,069 --> 00:18:06,379
the the quads that
we're gonna do.

463
00:18:06,379 --> 00:18:07,729
Then we get
together all of

464
00:18:07,729 --> 00:18:09,140
the information
for those quads.

465
00:18:09,140 --> 00:18:11,419
And then students
go out and spend

466
00:18:11,419 --> 00:18:12,679
a couple of weeks
in the field

467
00:18:12,679 --> 00:18:14,150
actually mapping
those quads.

468
00:18:14,150 --> 00:18:16,190
So these are the
quads that were on

469
00:18:16,190 --> 00:18:18,799
the docket to be
mapped last summer.

470
00:18:18,799 --> 00:18:21,754
So I went ahead and
got started on that.

471
00:18:21,754 --> 00:18:23,509
Now the interesting
thing about that,

472
00:18:23,509 --> 00:18:24,409
and this looks like it's

473
00:18:24,409 --> 00:18:26,150
a little bit out of order.

474
00:18:26,150 --> 00:18:31,999
Is that there's
physiographic provinces

475
00:18:31,999 --> 00:18:36,274
in just defined
everywhere.

476
00:18:36,274 --> 00:18:38,959
And physiographic
provinces are packages

477
00:18:38,959 --> 00:18:41,149
of stuff like I

478
00:18:41,149 --> 00:18:43,370
was talking about
earlier terrains.

479
00:18:43,370 --> 00:18:45,215
These are packages of

480
00:18:45,215 --> 00:18:47,060
rocks that are similar and

481
00:18:47,060 --> 00:18:51,500
also typography and
stuff like that.

482
00:18:51,500 --> 00:18:53,060
So geomorphic lily,

483
00:18:53,060 --> 00:18:54,560
these things are
gonna be similar.

484
00:18:54,560 --> 00:18:55,639
So all of this
stuff that I'm

485
00:18:55,639 --> 00:18:57,320
wiggling my mouse
over down here in

486
00:18:57,320 --> 00:19:00,050
the lower right-hand
corner is

487
00:19:00,050 --> 00:19:04,430
one defined
physiographic province.

488
00:19:04,430 --> 00:19:05,839
And then over here on

489
00:19:05,839 --> 00:19:07,879
the left where I'm
wiggling my mouse,

490
00:19:07,879 --> 00:19:09,770
that's another
physiographic province.

491
00:19:09,770 --> 00:19:10,909
And finally it just up to

492
00:19:10,909 --> 00:19:12,740
the top here is
the third one.

493
00:19:12,740 --> 00:19:14,509
So what we can
see here is that

494
00:19:14,509 --> 00:19:16,490
my van and craft quads are

495
00:19:16,490 --> 00:19:18,694
actually split
down the middle

496
00:19:18,694 --> 00:19:21,770
by two different
physiographic provinces.

497
00:19:21,770 --> 00:19:23,629
So that's gonna
be a little bit

498
00:19:23,629 --> 00:19:25,250
of a challenge for
coming up with

499
00:19:25,250 --> 00:19:29,300
training data for
particular set.

500
00:19:29,300 --> 00:19:31,070
And so what I ended up

501
00:19:31,070 --> 00:19:33,724
doing for the
students to go out

502
00:19:33,724 --> 00:19:35,930
immediately in the
summer was I just

503
00:19:35,930 --> 00:19:38,554
did an unsupervised
classification.

504
00:19:38,554 --> 00:19:42,244
So this is what I got
from just running.

505
00:19:42,244 --> 00:19:44,540
I mean, I tweaked it
and stuff like that.

506
00:19:44,540 --> 00:19:46,099
But this was sort of what

507
00:19:46,099 --> 00:19:48,529
the machine-learning
models came

508
00:19:48,529 --> 00:19:49,264
up with for an

509
00:19:49,264 --> 00:19:50,599
unclassified or

510
00:19:50,599 --> 00:19:53,179
an unsupervised
classification of

511
00:19:53,179 --> 00:19:55,400
the van and craft
quads and Van is

512
00:19:55,400 --> 00:19:57,799
the top quad and craft
is the lower flawed.

513
00:19:57,799 --> 00:20:00,920
And these are the two
quads that we started

514
00:20:00,920 --> 00:20:05,149
off to try to use
machine learning on.

515
00:20:05,149 --> 00:20:07,369
Here's a bigger picture of

516
00:20:07,369 --> 00:20:10,099
the physiographic
regions of the area.

517
00:20:10,099 --> 00:20:12,980
And so it
outlined in green

518
00:20:12,980 --> 00:20:14,570
down here is the section

519
00:20:14,570 --> 00:20:17,129
that we're looking at.

520
00:20:18,820 --> 00:20:21,499
And that's

521
00:20:21,499 --> 00:20:23,405
where that slide
was supposed to be.

522
00:20:23,405 --> 00:20:24,349
Okay.

523
00:20:24,349 --> 00:20:26,284
Then so these are

524
00:20:26,284 --> 00:20:28,070
this is the actual
geologic map.

525
00:20:28,070 --> 00:20:28,309
This is

526
00:20:28,309 --> 00:20:30,695
the existing geologic
map for that area.

527
00:20:30,695 --> 00:20:33,574
This is the van wad
and the craft quad.

528
00:20:33,574 --> 00:20:35,689
And we did see in the
craft quad that there

529
00:20:35,689 --> 00:20:37,670
was a little section of
yellow down in here.

530
00:20:37,670 --> 00:20:39,200
So somebody has previously

531
00:20:39,200 --> 00:20:40,534
mapped part of this.

532
00:20:40,534 --> 00:20:49,010
But what we see in
this map is that

533
00:20:49,010 --> 00:20:53,090
there are several
places that

534
00:20:53,090 --> 00:20:57,530
are potentially map
useful as training sites.

535
00:20:57,530 --> 00:20:58,910
Let me go back
to this map.

536
00:20:58,910 --> 00:21:00,035
That's why this
one is here.

537
00:21:00,035 --> 00:21:01,490
So Van van is

538
00:21:01,490 --> 00:21:04,625
red craft has some
yellow in it.

539
00:21:04,625 --> 00:21:07,220
Down here to
the lower left,

540
00:21:07,220 --> 00:21:08,840
I've got Burns
Butte and burns.

541
00:21:08,840 --> 00:21:10,700
These two quads
are mapped in

542
00:21:10,700 --> 00:21:13,819
green there in this
physiographic province.

543
00:21:13,819 --> 00:21:15,380
And then over here on
this side I've got

544
00:21:15,380 --> 00:21:17,434
Warm Springs and
mosquito mountain.

545
00:21:17,434 --> 00:21:18,710
And those are
mapped in, in

546
00:21:18,710 --> 00:21:20,029
high resolution in this

547
00:21:20,029 --> 00:21:21,514
physiographic province.

548
00:21:21,514 --> 00:21:27,110
So those, those two
training sites might be

549
00:21:27,110 --> 00:21:30,229
good candidates
for training up

550
00:21:30,229 --> 00:21:34,039
to try to apply those
to ban and craft.

551
00:21:34,039 --> 00:21:37,115
But it turns out that

552
00:21:37,115 --> 00:21:40,309
even closer to
van and craft,

553
00:21:40,309 --> 00:21:43,070
I've got Buchanan and

554
00:21:43,070 --> 00:21:45,994
stinking water pass
right down here below.

555
00:21:45,994 --> 00:21:49,894
And so I ended up
looking at these.

556
00:21:49,894 --> 00:21:52,070
I can see that
the rocks on

557
00:21:52,070 --> 00:21:52,969
this side over here and

558
00:21:52,969 --> 00:21:54,380
stinking water pass do

559
00:21:54,380 --> 00:21:56,869
seem similar to
the rocks on

560
00:21:56,869 --> 00:21:58,159
the right-hand side of

561
00:21:58,159 --> 00:21:59,989
our maps and
van and craft.

562
00:21:59,989 --> 00:22:01,699
And the rocks on

563
00:22:01,699 --> 00:22:05,134
the left-hand side
of this in Buchanan

564
00:22:05,134 --> 00:22:08,300
are more similar
perhaps to the stuff on

565
00:22:08,300 --> 00:22:11,494
the left-hand side
of this divide.

566
00:22:11,494 --> 00:22:12,920
Now, the thing about this

567
00:22:12,920 --> 00:22:14,390
physiographic provinces

568
00:22:14,390 --> 00:22:16,039
thing is that it's kind

569
00:22:16,039 --> 00:22:17,914
of a really big
picture thing.

570
00:22:17,914 --> 00:22:20,059
Like this line that
I'm looking at here,

571
00:22:20,059 --> 00:22:22,894
could literally be tens,

572
00:22:22,894 --> 00:22:24,559
tens of miles on,

573
00:22:24,559 --> 00:22:26,284
to the left or right.

574
00:22:26,284 --> 00:22:29,060
So this is kind of
a general cut at it

575
00:22:29,060 --> 00:22:31,610
in order to nail it down.

576
00:22:31,610 --> 00:22:33,799
But this but this
line could do could

577
00:22:33,799 --> 00:22:36,410
be over here or it
could be over here.

578
00:22:36,410 --> 00:22:39,029
It could be all the
way over here perhaps.

579
00:22:39,640 --> 00:22:41,779
But looking at this map

580
00:22:41,779 --> 00:22:42,994
and spending some
time looking at it,

581
00:22:42,994 --> 00:22:44,329
I did notice there's

582
00:22:44,329 --> 00:22:46,654
some nice material down
here and burns view

583
00:22:46,654 --> 00:22:48,860
that I could sample
and I may end up using

584
00:22:48,860 --> 00:22:50,360
that because I don't

585
00:22:50,360 --> 00:22:52,339
have to only sample
from one plot.

586
00:22:52,339 --> 00:22:54,664
I can take samples
from anywhere I want.

587
00:22:54,664 --> 00:22:57,515
So it's possible
that will,

588
00:22:57,515 --> 00:23:00,259
that we will make

589
00:23:00,259 --> 00:23:01,429
this model a
little bit more

590
00:23:01,429 --> 00:23:03,560
sophisticated by
sampling from some areas

591
00:23:03,560 --> 00:23:05,719
that maybe are not,
that don't exist.

592
00:23:05,719 --> 00:23:08,854
But for this effort,

593
00:23:08,854 --> 00:23:10,384
I just ended up using

594
00:23:10,384 --> 00:23:14,255
Buchanan and
stinking water pass.

595
00:23:14,255 --> 00:23:16,370
I went ahead and did

596
00:23:16,370 --> 00:23:18,259
a lot of sampling
on this one.

597
00:23:18,259 --> 00:23:22,579
So that's my stratified
sampling strategy

598
00:23:22,579 --> 00:23:27,455
for my Buchanan and
stinking water pass.

599
00:23:27,455 --> 00:23:32,845
And then I want to
introduce some of you.

600
00:23:32,845 --> 00:23:34,329
I wasn't sure
how many people

601
00:23:34,329 --> 00:23:35,635
would show up today,

602
00:23:35,635 --> 00:23:39,219
but I did a demo
of this in the,

603
00:23:39,219 --> 00:23:40,420
in the geology department

604
00:23:40,420 --> 00:23:42,460
for my grad students.

605
00:23:42,460 --> 00:23:44,769
I do a lot of work
in Python and

606
00:23:44,769 --> 00:23:46,885
R. And I'm gonna do

607
00:23:46,885 --> 00:23:49,030
any final work
in Python or

608
00:23:49,030 --> 00:23:49,629
R because of

609
00:23:49,629 --> 00:23:52,119
the reproducibility
of the results.

610
00:23:52,119 --> 00:23:54,339
Because you can actually
step through it

611
00:23:54,339 --> 00:23:56,875
and understand exactly
what's going on.

612
00:23:56,875 --> 00:23:58,824
But I want to encourage
people to use this.

613
00:23:58,824 --> 00:24:00,310
It's called the
orange, and

614
00:24:00,310 --> 00:24:01,405
it's for data mining.

615
00:24:01,405 --> 00:24:02,905
And this is an example

616
00:24:02,905 --> 00:24:05,740
of a workflow where
I've got this,

617
00:24:05,740 --> 00:24:06,850
this is what we want.

618
00:24:06,850 --> 00:24:09,009
I ended up doing
my training data

619
00:24:09,009 --> 00:24:10,420
coming into this.

620
00:24:10,420 --> 00:24:14,355
So this is a visual
modeling program.

621
00:24:14,355 --> 00:24:16,084
I've got my training
data coming in.

622
00:24:16,084 --> 00:24:17,720
I'm selecting my columns

623
00:24:17,720 --> 00:24:19,220
and selecting the map unit

624
00:24:19,220 --> 00:24:22,955
as the target that I'm
trying to predict.

625
00:24:22,955 --> 00:24:25,249
So I've got all
of my columns of

626
00:24:25,249 --> 00:24:27,740
data coming in in

627
00:24:27,740 --> 00:24:30,979
a CSV with one
column that's got

628
00:24:30,979 --> 00:24:33,019
the map unit
that goes along

629
00:24:33,019 --> 00:24:35,375
with those ten column,

630
00:24:35,375 --> 00:24:37,594
ten columns of
satellite data.

631
00:24:37,594 --> 00:24:40,084
So I set the map
unit is the target.

632
00:24:40,084 --> 00:24:41,779
I, I sample it down,

633
00:24:41,779 --> 00:24:44,719
so I'm not grabbing
all 144,000.

634
00:24:44,719 --> 00:24:46,655
I'm using only 10%.

635
00:24:46,655 --> 00:24:48,380
Then I'm sending that off

636
00:24:48,380 --> 00:24:49,910
to a random forest model,

637
00:24:49,910 --> 00:24:51,589
a logistic
regression model,

638
00:24:51,589 --> 00:24:52,955
and a neural
network model.

639
00:24:52,955 --> 00:24:55,280
And then looking at the
testing and scoring

640
00:24:55,280 --> 00:24:57,784
in the confusion matrix
and all of that.

641
00:24:57,784 --> 00:24:59,645
And then down at
the bottom of

642
00:24:59,645 --> 00:25:02,000
the workflow here I've
got another dataset

643
00:25:02,000 --> 00:25:03,905
coming in and this is
the prediction data

644
00:25:03,905 --> 00:25:07,355
that I had to rename
the sum of the columns.

645
00:25:07,355 --> 00:25:09,170
And then I'm sending

646
00:25:09,170 --> 00:25:10,834
that through to
the predictions.

647
00:25:10,834 --> 00:25:12,440
And that's
grabbing the model

648
00:25:12,440 --> 00:25:13,880
from random
forest over here,

649
00:25:13,880 --> 00:25:15,214
sending it to
the predictions,

650
00:25:15,214 --> 00:25:17,569
merging that with the
data that I brought in,

651
00:25:17,569 --> 00:25:19,505
and then creating

652
00:25:19,505 --> 00:25:22,640
a data table and
then exporting

653
00:25:22,640 --> 00:25:24,830
it as a CSV
that I can then

654
00:25:24,830 --> 00:25:25,880
integrate back with my

655
00:25:25,880 --> 00:25:28,354
GIS software
to make a map.

656
00:25:28,354 --> 00:25:30,769
There's the
confusion matrix

657
00:25:30,769 --> 00:25:33,589
from the different units.

658
00:25:33,589 --> 00:25:35,359
And we can see right

659
00:25:35,359 --> 00:25:37,129
away that this
one unit here,

660
00:25:37,129 --> 00:25:40,204
T MBA, it has a lot of,

661
00:25:40,204 --> 00:25:42,830
a lot of problems
and it also ends up

662
00:25:42,830 --> 00:25:47,450
being the most ubiquitous
unit in the thing.

663
00:25:47,450 --> 00:25:48,620
So I'm going to have
to do something

664
00:25:48,620 --> 00:25:49,730
to take care of that.

665
00:25:49,730 --> 00:25:50,660
And I haven't
quite figured

666
00:25:50,660 --> 00:25:52,550
out what to do
about that yet.

667
00:25:52,550 --> 00:25:54,830
But here it is.

668
00:25:54,830 --> 00:25:56,074
This is the actual map

669
00:25:56,074 --> 00:25:59,120
of the predictions from,

670
00:25:59,120 --> 00:26:00,679
from my random
forest model.

671
00:26:00,679 --> 00:26:02,734
And we can see once
again that this is,

672
00:26:02,734 --> 00:26:05,420
it's looking coherent
like a geologic map.

673
00:26:05,420 --> 00:26:08,494
We've got stuff grouping
together over here

674
00:26:08,494 --> 00:26:13,205
that is genetically
related.

675
00:26:13,205 --> 00:26:14,870
We've got stuff
down in here

676
00:26:14,870 --> 00:26:16,429
that looks quite similar.

677
00:26:16,429 --> 00:26:17,120
And then we've got

678
00:26:17,120 --> 00:26:18,559
a streak of stuff
going through

679
00:26:18,559 --> 00:26:21,499
here and some nice
groupings down in here.

680
00:26:21,499 --> 00:26:23,885
So it seems to
be pretty good.

681
00:26:23,885 --> 00:26:27,530
Then if we overlay
that on the,

682
00:26:27,530 --> 00:26:29,524
with the hillshade map,

683
00:26:29,524 --> 00:26:31,729
we can see that
we can actually

684
00:26:31,729 --> 00:26:33,994
find out crops
along in here.

685
00:26:33,994 --> 00:26:34,670
Now I need to do

686
00:26:34,670 --> 00:26:36,425
a little bit more
sophisticated work

687
00:26:36,425 --> 00:26:38,495
on finding the outcrops.

688
00:26:38,495 --> 00:26:41,809
And that is just really
a matter of time.

689
00:26:41,809 --> 00:26:44,315
Because what I can do

690
00:26:44,315 --> 00:26:47,239
is look for exposures
where there's

691
00:26:47,239 --> 00:26:49,769
a lack of vegetation
and hopefully you

692
00:26:49,769 --> 00:26:51,829
see that that would end up

693
00:26:51,829 --> 00:26:54,949
being a good
outcrop there.

694
00:26:54,949 --> 00:26:57,664
Now, this summer we did,

695
00:26:57,664 --> 00:27:00,199
the students
did go out and

696
00:27:00,199 --> 00:27:03,364
they went ahead and
sampled the data.

697
00:27:03,364 --> 00:27:05,645
From the field.
And so this is

698
00:27:05,645 --> 00:27:07,625
the data from
craft point from

699
00:27:07,625 --> 00:27:12,605
Angela stepson who
shared her data with me.

700
00:27:12,605 --> 00:27:15,079
And I'm still integrating

701
00:27:15,079 --> 00:27:17,989
this with the data
from the model,

702
00:27:17,989 --> 00:27:19,580
but it's looking
quite good.

703
00:27:19,580 --> 00:27:22,024
We're starting to see
similar patterns.

704
00:27:22,024 --> 00:27:24,770
So this idea of
integrating the,

705
00:27:24,770 --> 00:27:26,614
the modeling with
the field work

706
00:27:26,614 --> 00:27:29,509
is really where I'd
like to go with this.

707
00:27:29,509 --> 00:27:32,479
And that is, I would
like us to be able

708
00:27:32,479 --> 00:27:35,659
to have these models in

709
00:27:35,659 --> 00:27:37,834
a Cloud-based
system and then put

710
00:27:37,834 --> 00:27:40,220
data into it at
night after we

711
00:27:40,220 --> 00:27:41,359
get back from mapping in

712
00:27:41,359 --> 00:27:42,830
the field and then be able

713
00:27:42,830 --> 00:27:46,559
to update the model
appropriately.

714
00:27:46,559 --> 00:27:49,420
And I guess I, I guess
I got through that

715
00:27:49,420 --> 00:27:50,560
fairly quickly
because this

716
00:27:50,560 --> 00:27:52,525
is my final
thoughts slide.

717
00:27:52,525 --> 00:27:54,550
So the idea is that

718
00:27:54,550 --> 00:27:57,924
this approach is yielding
promising results.

719
00:27:57,924 --> 00:28:00,534
But other data
can be added.

720
00:28:00,534 --> 00:28:03,280
Increase the accuracy
I need to deal with

721
00:28:03,280 --> 00:28:07,315
whatever t MBA
unit because

722
00:28:07,315 --> 00:28:10,060
that one is actually
about half of

723
00:28:10,060 --> 00:28:13,389
the data that I'm

724
00:28:13,389 --> 00:28:15,489
using is of that
particular type.

725
00:28:15,489 --> 00:28:15,790
So there's

726
00:28:15,790 --> 00:28:18,355
something interesting
going on there.

727
00:28:18,355 --> 00:28:20,425
Then I need to tackle.

728
00:28:20,425 --> 00:28:22,270
The most important thing
though, of course,

729
00:28:22,270 --> 00:28:24,610
is I want to
use this to go

730
00:28:24,610 --> 00:28:28,620
after funding for
getting better data.

731
00:28:28,620 --> 00:28:31,054
Because right now
I'm limited to

732
00:28:31,054 --> 00:28:33,139
just the freely
available Sentinel data

733
00:28:33,139 --> 00:28:34,339
at the scale of 20

734
00:28:34,339 --> 00:28:36,350
m. But there's

735
00:28:36,350 --> 00:28:40,324
much higher resolution
available data.

736
00:28:40,324 --> 00:28:42,844
And with some funding
from the state,

737
00:28:42,844 --> 00:28:44,329
we could probably get

738
00:28:44,329 --> 00:28:47,539
some pretty good
results from that.

739
00:28:47,539 --> 00:28:50,540
And the idea is
still though,

740
00:28:50,540 --> 00:28:52,489
is that integrating
this with field work is

741
00:28:52,489 --> 00:28:54,529
going to be the
best we don't want.

742
00:28:54,529 --> 00:28:58,144
I am not proposing
in any way that we

743
00:28:58,144 --> 00:29:02,015
start to replace
geologic mapping.

744
00:29:02,015 --> 00:29:04,160
The actual activity
of geologic mapping,

745
00:29:04,160 --> 00:29:05,509
which involves going
out and hitting

746
00:29:05,509 --> 00:29:07,174
rocks with hammers
and all of that.

747
00:29:07,174 --> 00:29:08,210
We're not going to replace

748
00:29:08,210 --> 00:29:09,215
that with satellite data,

749
00:29:09,215 --> 00:29:11,180
but we're going
to enhance it.

750
00:29:11,180 --> 00:29:12,410
So there we go.

751
00:29:12,410 --> 00:29:13,775
That's my talk.

752
00:29:13,775 --> 00:29:15,559
Probably meet maybe one of

753
00:29:15,559 --> 00:29:17,060
the shortest system
science talks ever,

754
00:29:17,060 --> 00:29:18,359
right?

755
00:29:20,980 --> 00:29:23,999
That's terrific.

756
00:29:25,570 --> 00:29:27,994
So yeah, you all got back

757
00:29:27,994 --> 00:29:30,560
an extra half-hour except

758
00:29:30,560 --> 00:29:31,580
that we're probably
going to want

759
00:29:31,580 --> 00:29:32,675
to ask some questions.

760
00:29:32,675 --> 00:29:35,189
Yes, please ask questions.

761
00:29:35,800 --> 00:29:38,600
You mentioned Python
a little while ago.

762
00:29:38,600 --> 00:29:39,800
How critical is Python?

763
00:29:39,800 --> 00:29:41,119
Do the work that you do?

764
00:29:41,119 --> 00:29:45,364
Oh, good question.
It's not.

765
00:29:45,364 --> 00:29:47,554
I'm I'm agnostic.

766
00:29:47,554 --> 00:29:50,480
I use are, I

767
00:29:50,480 --> 00:29:53,134
actually I prefer
are for some stuff.

768
00:29:53,134 --> 00:29:56,239
Because the way
that Python works,

769
00:29:56,239 --> 00:29:57,890
some of the ways
that some of

770
00:29:57,890 --> 00:30:00,260
the idioms in Python
don't work as

771
00:30:00,260 --> 00:30:03,139
well for modeling
as well as R.

772
00:30:03,139 --> 00:30:05,045
But it's one of
those things where

773
00:30:05,045 --> 00:30:06,544
the orange though,

774
00:30:06,544 --> 00:30:08,089
that orange thing that
I was showing you,

775
00:30:08,089 --> 00:30:09,094
the orange data mining,

776
00:30:09,094 --> 00:30:10,519
that's a little Python.

777
00:30:10,519 --> 00:30:13,820
So what you can do,
that's kinda cool.

778
00:30:13,820 --> 00:30:15,559
And I haven't actually
tried this yet.

779
00:30:15,559 --> 00:30:17,119
As you can get
your model all

780
00:30:17,119 --> 00:30:18,769
working there in orange in

781
00:30:18,769 --> 00:30:20,240
the visual thing and then

782
00:30:20,240 --> 00:30:21,920
export it to Python.

783
00:30:21,920 --> 00:30:23,750
And all of your
stuff is documented

784
00:30:23,750 --> 00:30:25,640
as far as which widgets
using and stuff.

785
00:30:25,640 --> 00:30:27,215
And then you can go
in and tweak stuff.

786
00:30:27,215 --> 00:30:29,824
So that's actually
kinda promising.

787
00:30:29,824 --> 00:30:31,955
I guess, where I was
going with this.

788
00:30:31,955 --> 00:30:33,349
What kind of
advice would you

789
00:30:33,349 --> 00:30:34,970
recommend to our
students who

790
00:30:34,970 --> 00:30:36,229
want to be well-prepared

791
00:30:36,229 --> 00:30:38,300
for the research world,

792
00:30:38,300 --> 00:30:42,065
a computationally
oriented research world.

793
00:30:42,065 --> 00:30:43,220
Most of them are
wondering, should

794
00:30:43,220 --> 00:30:44,479
I learn our Python?

795
00:30:44,479 --> 00:30:46,279
Both are there other
tools I should learn?

796
00:30:46,279 --> 00:30:47,660
Should I wait until
the latest thing

797
00:30:47,660 --> 00:30:49,714
emerges and I'm
ready to learn?

798
00:30:49,714 --> 00:30:53,959
Yeah, I know. You can't.

799
00:30:53,959 --> 00:30:56,344
I think everybody
needs to learn Python.

800
00:30:56,344 --> 00:30:58,130
And the thing
is, what I've

801
00:30:58,130 --> 00:31:00,199
been telling
students lately is

802
00:31:00,199 --> 00:31:02,269
that you need to

803
00:31:02,269 --> 00:31:04,580
learn these things well
enough that you can

804
00:31:04,580 --> 00:31:06,439
understand other
people's code

805
00:31:06,439 --> 00:31:07,609
so that when you run

806
00:31:07,609 --> 00:31:09,244
into somebody else's code

807
00:31:09,244 --> 00:31:10,985
and you want to
use it yourself,

808
00:31:10,985 --> 00:31:12,049
you can figure
out where the

809
00:31:12,049 --> 00:31:13,550
variables are and
how to change

810
00:31:13,550 --> 00:31:14,899
the data is that it's

811
00:31:14,899 --> 00:31:17,639
pointing to and all of
that sort of stuff.

812
00:31:17,710 --> 00:31:21,499
That to me, two
most, most,

813
00:31:21,499 --> 00:31:23,569
most working
scientists these

814
00:31:23,569 --> 00:31:25,759
days need to be
at that level.

815
00:31:25,759 --> 00:31:28,220
At least. I don't expect
people to be able

816
00:31:28,220 --> 00:31:29,075
to sort of like

817
00:31:29,075 --> 00:31:31,519
write-up program
from scratch.

818
00:31:31,519 --> 00:31:33,650
Because that
actually is much,

819
00:31:33,650 --> 00:31:35,119
much harder than

820
00:31:35,119 --> 00:31:36,814
tweaking other
people's code.

821
00:31:36,814 --> 00:31:38,719
And tweaking other
people's code is actually

822
00:31:38,719 --> 00:31:41,479
a more valuable
skill in a lot of

823
00:31:41,479 --> 00:31:43,160
ways than being
able to just

824
00:31:43,160 --> 00:31:45,769
write a program from a
script from scratch.

825
00:31:45,769 --> 00:31:49,444
But I also have, I come
from being a hacker.

826
00:31:49,444 --> 00:31:51,920
I'm a hacker from
the '80s and

827
00:31:51,920 --> 00:31:55,505
I I did IT and stuff.

828
00:31:55,505 --> 00:31:56,855
So I haven't really,

829
00:31:56,855 --> 00:31:59,960
I have a very
different attitude

830
00:31:59,960 --> 00:32:01,220
towards using software.

831
00:32:01,220 --> 00:32:02,254
I just like, Oh,

832
00:32:02,254 --> 00:32:03,500
look, it's another
piece of software.

833
00:32:03,500 --> 00:32:03,740
Cool.

834
00:32:03,740 --> 00:32:05,359
Let's just dive in
and see what it does.

835
00:32:05,359 --> 00:32:07,325
You know, a lot
of people go.

836
00:32:07,325 --> 00:32:08,149
I know.

837
00:32:08,149 --> 00:32:12,649
I do think when you guys
are accessing data,

838
00:32:12,649 --> 00:32:13,670
you obviously have to use

839
00:32:13,670 --> 00:32:15,949
GIS type of
tools primarily

840
00:32:15,949 --> 00:32:19,700
or GIS compatible use.

841
00:32:19,700 --> 00:32:21,544
I'm not sure what
I'm saying here,

842
00:32:21,544 --> 00:32:22,849
but I've had
people ask me,

843
00:32:22,849 --> 00:32:24,139
should I learn
SQL, which is

844
00:32:24,139 --> 00:32:26,210
just ancient language,
are dealing with.

845
00:32:26,210 --> 00:32:28,070
With databases and I

846
00:32:28,070 --> 00:32:29,630
pretty much say, why not?

847
00:32:29,630 --> 00:32:30,769
And I'm just
wondering what

848
00:32:30,769 --> 00:32:32,239
other people think and
you could comment,

849
00:32:32,239 --> 00:32:33,529
but others in the
room could comment

850
00:32:33,529 --> 00:32:35,570
as well, right?

851
00:32:35,570 --> 00:32:36,500
Yeah.

852
00:32:36,500 --> 00:32:38,480
Rick, what are you
what do you tell

853
00:32:38,480 --> 00:32:41,570
your students to
do for you for?

854
00:32:41,570 --> 00:32:46,145
Well, so overall,
we don't have,

855
00:32:46,145 --> 00:32:49,294
I don't see the
students who are

856
00:32:49,294 --> 00:32:53,060
on a trajectory to do
computational work.

857
00:32:53,060 --> 00:32:54,140
So I'm kind of

858
00:32:54,140 --> 00:32:56,165
absolved of that
responsibility.

859
00:32:56,165 --> 00:32:57,589
However, there is, within

860
00:32:57,589 --> 00:32:58,850
the culture of
public health is

861
00:32:58,850 --> 00:33:02,089
the distinction
between what

862
00:33:02,089 --> 00:33:03,379
the working scientists do,

863
00:33:03,379 --> 00:33:05,345
perhaps at the
state level,

864
00:33:05,345 --> 00:33:09,350
versus what Students will

865
00:33:09,350 --> 00:33:11,000
pick up on their
own through

866
00:33:11,000 --> 00:33:13,969
their own maybe
hobbyist approach

867
00:33:13,969 --> 00:33:17,029
to learning coding
in particular.

868
00:33:17,029 --> 00:33:19,940
And so I've seen

869
00:33:19,940 --> 00:33:21,830
among my colleagues
that there really

870
00:33:21,830 --> 00:33:25,369
is a teachable moment

871
00:33:25,369 --> 00:33:27,875
related to career
development

872
00:33:27,875 --> 00:33:28,715
where they say like,

873
00:33:28,715 --> 00:33:29,960
you know, it's how

874
00:33:29,960 --> 00:33:31,999
great that you can
do these things.

875
00:33:31,999 --> 00:33:36,334
But if you really
want to get a job,

876
00:33:36,334 --> 00:33:37,760
it's important
that you learn

877
00:33:37,760 --> 00:33:39,485
these other platforms.

878
00:33:39,485 --> 00:33:41,870
That's kinda the
guidance of the faculty.

879
00:33:41,870 --> 00:33:44,285
Use what other platforms

880
00:33:44,285 --> 00:33:45,620
just just listed for you.

881
00:33:45,620 --> 00:33:48,680
Well, so I know I'm
dating myself here

882
00:33:48,680 --> 00:33:50,149
because I haven't really

883
00:33:50,149 --> 00:33:52,715
like paid attention
too much.

884
00:33:52,715 --> 00:33:55,429
Overall.

885
00:33:55,429 --> 00:33:59,179
The big distinction
when I was paying

886
00:33:59,179 --> 00:34:02,840
attention was whether
students were ready.

887
00:34:02,840 --> 00:34:03,890
This was actually before

888
00:34:03,890 --> 00:34:05,330
the revolution happened

889
00:34:05,330 --> 00:34:07,069
and encompasses
whether students

890
00:34:07,069 --> 00:34:08,989
were ready to use.

891
00:34:08,989 --> 00:34:12,140
Sudan is,

892
00:34:12,140 --> 00:34:13,730
was popular in

893
00:34:13,730 --> 00:34:16,939
public health
epidemiological studies.

894
00:34:16,939 --> 00:34:19,744
And let's see.

895
00:34:19,744 --> 00:34:21,575
I'm kinda blocked in now.

896
00:34:21,575 --> 00:34:23,060
That's okay. I didn't
mean to put you

897
00:34:23,060 --> 00:34:24,574
on the spot really.
I didn't know.

898
00:34:24,574 --> 00:34:25,865
If people have advice,

899
00:34:25,865 --> 00:34:27,109
just one of the
questions that we

900
00:34:27,109 --> 00:34:28,760
have to try to
answer a lot.

901
00:34:28,760 --> 00:34:30,170
How technically
should I go?

902
00:34:30,170 --> 00:34:31,925
And if I move into
technical direction,

903
00:34:31,925 --> 00:34:32,810
What's gonna be the most

904
00:34:32,810 --> 00:34:34,009
useful in this field,

905
00:34:34,009 --> 00:34:35,300
that field, et cetera.

906
00:34:35,300 --> 00:34:37,310
Totally well, I do
have a corollary bit

907
00:34:37,310 --> 00:34:38,434
of information though.

908
00:34:38,434 --> 00:34:40,549
I went to a
conference that

909
00:34:40,549 --> 00:34:43,790
was populated
by a number of

910
00:34:43,790 --> 00:34:47,225
researchers from this
great place about

911
00:34:47,225 --> 00:34:48,679
the U-Dub called

912
00:34:48,679 --> 00:34:49,790
the Institute for Health

913
00:34:49,790 --> 00:34:51,470
Metrics and Evaluation.

914
00:34:51,470 --> 00:34:53,389
But the conference
was labeled as

915
00:34:53,389 --> 00:34:54,680
a global health conference

916
00:34:54,680 --> 00:34:55,700
and I teach global health.

917
00:34:55,700 --> 00:34:57,860
And so I pulled
aside one of

918
00:34:57,860 --> 00:35:00,290
the presenters
and asked him

919
00:35:00,290 --> 00:35:01,399
how I could get

920
00:35:01,399 --> 00:35:05,449
my students employed
after shock.

921
00:35:05,449 --> 00:35:08,389
And his first
question was,

922
00:35:08,389 --> 00:35:11,719
Can they program
databases? Right?

923
00:35:11,719 --> 00:35:14,074
And I said when we're

924
00:35:14,074 --> 00:35:16,460
much more quickly
hire someone

925
00:35:16,460 --> 00:35:18,545
that could program
a database

926
00:35:18,545 --> 00:35:20,479
than someone who has

927
00:35:20,479 --> 00:35:22,039
substantive knowledge
about global health.

928
00:35:22,039 --> 00:35:22,819
And we'll just teach

929
00:35:22,819 --> 00:35:23,900
them all about
global health.

930
00:35:23,900 --> 00:35:27,350
So that was
deflating, realistic.

931
00:35:27,350 --> 00:35:30,620
So good for System
Science, right?

932
00:35:30,620 --> 00:35:32,075
And good for data science.

933
00:35:32,075 --> 00:35:33,199
Well, our students are

934
00:35:33,199 --> 00:35:34,519
learning and I'll
stop talking.

935
00:35:34,519 --> 00:35:35,959
I know there might
be other questions.

936
00:35:35,959 --> 00:35:37,339
Our students are realizing

937
00:35:37,339 --> 00:35:39,859
that if they can't
claim to be,

938
00:35:39,859 --> 00:35:41,600
to some degree, reasonably

939
00:35:41,600 --> 00:35:42,904
well-trained in
data science.

940
00:35:42,904 --> 00:35:45,120
They're going to have
a hard time for 80,

941
00:35:45,120 --> 00:35:46,835
90% of the jobs out there

942
00:35:46,835 --> 00:35:48,995
knowing system
science isn't enough.

943
00:35:48,995 --> 00:35:51,425
It's, it's positive
on their resume.

944
00:35:51,425 --> 00:35:53,180
And so a lot of our
students are spending

945
00:35:53,180 --> 00:35:54,874
some time over and see us

946
00:35:54,874 --> 00:35:58,609
getting at least a
credible list of

947
00:35:58,609 --> 00:36:00,979
data science
credentials one way

948
00:36:00,979 --> 00:36:02,029
or the other to

949
00:36:02,029 --> 00:36:03,259
compliment their
system science.

950
00:36:03,259 --> 00:36:04,820
I think that's smart
that they're doing

951
00:36:04,820 --> 00:36:06,739
that we haven't actually

952
00:36:06,739 --> 00:36:08,629
lead that maybe as

953
00:36:08,629 --> 00:36:10,430
much as we should in
terms of advising.

954
00:36:10,430 --> 00:36:11,900
But that's why I'm
asking these questions.

955
00:36:11,900 --> 00:36:14,099
Have experienced people.

956
00:36:14,140 --> 00:36:16,819
I see people trying to do

957
00:36:16,819 --> 00:36:18,830
stuff in SQL, but that,

958
00:36:18,830 --> 00:36:21,020
that strikes me is
just being cute

959
00:36:21,020 --> 00:36:24,200
because there's other
tools that you can use.

960
00:36:24,200 --> 00:36:25,669
We'll do it much better.

961
00:36:25,669 --> 00:36:27,289
And proving
that you can do

962
00:36:27,289 --> 00:36:29,164
it in SQL is kinda cute.

963
00:36:29,164 --> 00:36:31,399
But it's not the
way to do it.

964
00:36:31,399 --> 00:36:34,699
It's like, yeah, I
could do that with SQL,

965
00:36:34,699 --> 00:36:36,050
but I can do it also in

966
00:36:36,050 --> 00:36:38,555
one line in
Python. You know.

967
00:36:38,555 --> 00:36:40,759
I mean, yeah, there's

968
00:36:40,759 --> 00:36:43,204
also like there's a
whole thing in GIS,

969
00:36:43,204 --> 00:36:44,629
like where people try
to prove that they

970
00:36:44,629 --> 00:36:46,549
can do stuff using
command line.

971
00:36:46,549 --> 00:36:48,695
Like I can do
Command Line GIS.

972
00:36:48,695 --> 00:36:49,580
And it's like, it's

973
00:36:49,580 --> 00:36:50,884
cute that you can do that.

974
00:36:50,884 --> 00:36:52,099
But in the real-world,

975
00:36:52,099 --> 00:36:53,299
you're going to grab
your mouse and go

976
00:36:53,299 --> 00:36:55,920
drag the thing
on the screen.

977
00:36:55,930 --> 00:36:58,879
So I see a lot of people

978
00:36:58,879 --> 00:37:01,174
doing weird
stuff out there,

979
00:37:01,174 --> 00:37:03,470
like with funny
programming languages

980
00:37:03,470 --> 00:37:05,750
or stuff just to kinda
prove that they can.

981
00:37:05,750 --> 00:37:09,709
But Python and R are
going to be here.

982
00:37:09,709 --> 00:37:11,329
I mean, those are,
those are the

983
00:37:11,329 --> 00:37:12,560
hardcore like if

984
00:37:12,560 --> 00:37:15,080
you know Python and R
and system science,

985
00:37:15,080 --> 00:37:17,344
then you can go out
and do amazing things.

986
00:37:17,344 --> 00:37:19,415
They personally, yeah.

987
00:37:19,415 --> 00:37:22,325
So maybe, maybe you
mentioned this.

988
00:37:22,325 --> 00:37:23,869
I got a bit distracted
there in the middle,

989
00:37:23,869 --> 00:37:25,490
but how big, how big

990
00:37:25,490 --> 00:37:27,319
is your datasets for this?

991
00:37:27,319 --> 00:37:28,610
And did you, did
you execute it

992
00:37:28,610 --> 00:37:30,254
all on a local computer?

993
00:37:30,254 --> 00:37:32,180
Did you run some of
this in the cloud?

994
00:37:32,180 --> 00:37:33,800
Yeah. Great question.

995
00:37:33,800 --> 00:37:35,270
I actually did.
I kinda hit

996
00:37:35,270 --> 00:37:37,355
the limits on my
local computer.

997
00:37:37,355 --> 00:37:38,554
I'm actually remote it in.

998
00:37:38,554 --> 00:37:40,520
I've got a really
big workstation

999
00:37:40,520 --> 00:37:43,055
at Portland State
in my office.

1000
00:37:43,055 --> 00:37:45,109
And I remote into
it from home and

1001
00:37:45,109 --> 00:37:47,555
run all of my modeling
and stuff there.

1002
00:37:47,555 --> 00:37:49,789
And some of these datasets

1003
00:37:49,789 --> 00:37:52,414
are enough to even
slow that sucker down.

1004
00:37:52,414 --> 00:37:55,550
There were 300,000 points

1005
00:37:55,550 --> 00:37:58,175
in the final prediction.

1006
00:37:58,175 --> 00:38:07,319
So and then they trained
on 144,000 points.

1007
00:38:07,660 --> 00:38:10,414
And I did some
random sampling

1008
00:38:10,414 --> 00:38:13,444
to get the
computation down.

1009
00:38:13,444 --> 00:38:15,200
But you're absolutely
right, Glenn,

1010
00:38:15,200 --> 00:38:17,540
what I need to do
next is actually move

1011
00:38:17,540 --> 00:38:19,280
this entire process out

1012
00:38:19,280 --> 00:38:20,644
to Google Earth Engine.

1013
00:38:20,644 --> 00:38:22,774
Because Google Earth
Engine has access

1014
00:38:22,774 --> 00:38:25,129
in the Cloud to all
that sudden old data.

1015
00:38:25,129 --> 00:38:26,660
And then I can
just run my code

1016
00:38:26,660 --> 00:38:27,800
in Python right out

1017
00:38:27,800 --> 00:38:29,720
there and not have

1018
00:38:29,720 --> 00:38:31,999
to transfer any data
back and forth.

1019
00:38:31,999 --> 00:38:34,175
Yeah, that sounds
that sounds great.

1020
00:38:34,175 --> 00:38:35,930
Yeah, that's, that's
the way that's

1021
00:38:35,930 --> 00:38:38,640
the way we need
to go with this.

1022
00:38:38,650 --> 00:38:43,340
I see only one small
I have a qualm.

1023
00:38:43,340 --> 00:38:44,734
Let's put it that way.

1024
00:38:44,734 --> 00:38:46,744
Oh, good to hear
from you, Steve.

1025
00:38:46,744 --> 00:38:52,715
Well, maybe lidar data

1026
00:38:52,715 --> 00:38:53,870
should be
available for much

1027
00:38:53,870 --> 00:38:55,159
of the state, right?

1028
00:38:55,159 --> 00:38:58,594
Oh, actually, no. We're
working on that now.

1029
00:38:58,594 --> 00:38:59,840
Yeah, but yeah.

1030
00:38:59,840 --> 00:39:00,274
Go ahead.

1031
00:39:00,274 --> 00:39:02,240
It seems to me that's

1032
00:39:02,240 --> 00:39:04,220
your primary tool for

1033
00:39:04,220 --> 00:39:07,505
location of
spaces, slopes.

1034
00:39:07,505 --> 00:39:08,735
Oh, yeah, definitely.

1035
00:39:08,735 --> 00:39:11,614
And watersheds, etc.

1036
00:39:11,614 --> 00:39:14,060
Yeah. So once you
have that data

1037
00:39:14,060 --> 00:39:17,344
in and you organize

1038
00:39:17,344 --> 00:39:19,984
data in terms of

1039
00:39:19,984 --> 00:39:21,710
the Lidar slopes that

1040
00:39:21,710 --> 00:39:22,189
are because you're going

1041
00:39:22,189 --> 00:39:24,094
to end up with
between two spaces,

1042
00:39:24,094 --> 00:39:26,224
because you've got pixels.

1043
00:39:26,224 --> 00:39:28,369
You now I've got a wave

1044
00:39:28,369 --> 00:39:29,869
spatially
organizing it would

1045
00:39:29,869 --> 00:39:33,394
seem to me you're currently
separated pixels.

1046
00:39:33,394 --> 00:39:35,749
The problem I
see for mapping

1047
00:39:35,749 --> 00:39:38,119
a geological problem
with a set of

1048
00:39:38,119 --> 00:39:40,864
separated
locations is that

1049
00:39:40,864 --> 00:39:43,100
the space in Portland is

1050
00:39:43,100 --> 00:39:45,724
not the space in
Eastern Oregon,

1051
00:39:45,724 --> 00:39:48,049
but you're being
treating basically them

1052
00:39:48,049 --> 00:39:49,579
each is separated pixels

1053
00:39:49,579 --> 00:39:51,560
we characteristics
based on light.

1054
00:39:51,560 --> 00:39:53,555
And now you're doing a
little bit of large,

1055
00:39:53,555 --> 00:39:55,834
of a large area geometry.

1056
00:39:55,834 --> 00:39:58,219
But you really
want to handle

1057
00:39:58,219 --> 00:39:59,930
two pixels that
are side-by-side

1058
00:39:59,930 --> 00:40:01,970
as though they're
side-by-side.

1059
00:40:01,970 --> 00:40:04,174
And solving that problem,

1060
00:40:04,174 --> 00:40:05,749
which I haven't given

1061
00:40:05,749 --> 00:40:09,239
enough thought to know
how confused I am,

1062
00:40:10,150 --> 00:40:12,770
would seem to me to be

1063
00:40:12,770 --> 00:40:15,590
a key aspect of

1064
00:40:15,590 --> 00:40:18,065
coming up with a
geological map.

1065
00:40:18,065 --> 00:40:19,850
That's a really
good point, Steve.

1066
00:40:19,850 --> 00:40:20,929
And actually
that it totally

1067
00:40:20,929 --> 00:40:22,520
ties into the work
that Marty and I are

1068
00:40:22,520 --> 00:40:25,834
doing together because
we're taking pixels

1069
00:40:25,834 --> 00:40:29,555
of land use data from
a satellite map,

1070
00:40:29,555 --> 00:40:31,204
the national
land cover map.

1071
00:40:31,204 --> 00:40:32,660
And we're looking
at that pixel

1072
00:40:32,660 --> 00:40:34,009
plus the pixels
of the north,

1073
00:40:34,009 --> 00:40:35,464
south, east, and west.

1074
00:40:35,464 --> 00:40:36,980
And we're
treating those as

1075
00:40:36,980 --> 00:40:39,785
new variables
in our model.

1076
00:40:39,785 --> 00:40:43,235
Yes, That's the
general idea.

1077
00:40:43,235 --> 00:40:44,929
I'm not sure
because you're

1078
00:40:44,929 --> 00:40:46,775
dealing with slopes
and gravity,

1079
00:40:46,775 --> 00:40:50,330
whether or not you
want to make it that

1080
00:40:50,330 --> 00:40:55,460
unconstrained in a
geoloc mapping problem.

1081
00:40:55,460 --> 00:40:59,240
But it seems to me
that you have to

1082
00:40:59,240 --> 00:41:00,575
do that in order to have

1083
00:41:00,575 --> 00:41:03,184
any possibility
of success.

1084
00:41:03,184 --> 00:41:04,715
And I'd wonder about

1085
00:41:04,715 --> 00:41:08,119
the geological
relationship of

1086
00:41:08,119 --> 00:41:09,890
that set of pixels,

1087
00:41:09,890 --> 00:41:11,959
which gave us his
poor results.

1088
00:41:11,959 --> 00:41:17,074
Slope angle
boundaries, right?

1089
00:41:17,074 --> 00:41:19,160
And that definitely,
So slope is definitely

1090
00:41:19,160 --> 00:41:20,300
one of the other
variables that

1091
00:41:20,300 --> 00:41:22,055
I want to put
into the model.

1092
00:41:22,055 --> 00:41:25,189
Because certain types
of units are gonna be

1093
00:41:25,189 --> 00:41:30,769
more inclined to be

1094
00:41:30,769 --> 00:41:34,535
flatline versus being
actually inclined.

1095
00:41:34,535 --> 00:41:37,730
That's actually
an attribute of

1096
00:41:37,730 --> 00:41:41,809
a geologic unit is how
Angular does it yet.

1097
00:41:41,809 --> 00:41:44,360
So, yeah, that's
a, that's a great,

1098
00:41:44,360 --> 00:41:46,670
That's definitely
part of it.

1099
00:41:46,670 --> 00:41:48,829
The pixels are
close enough

1100
00:41:48,829 --> 00:41:50,300
together that it's
not a problem.

1101
00:41:50,300 --> 00:41:52,295
But what I think you
might be saying is that

1102
00:41:52,295 --> 00:41:54,799
each pixel is being
treated independently,

1103
00:41:54,799 --> 00:41:56,495
whereas it should be

1104
00:41:56,495 --> 00:41:58,414
treated as part of
a neighborhood?

1105
00:41:58,414 --> 00:42:01,489
Yes. And perhaps I'm not

1106
00:42:01,489 --> 00:42:02,540
quite sure how to define

1107
00:42:02,540 --> 00:42:04,564
neighborhood for
geological purposes.

1108
00:42:04,564 --> 00:42:07,564
I'm not a geologist,
but nevertheless,

1109
00:42:07,564 --> 00:42:08,750
it would seem
to me that that

1110
00:42:08,750 --> 00:42:10,849
would be important for

1111
00:42:10,849 --> 00:42:15,395
extracting what a compact,
what would it be?

1112
00:42:15,395 --> 00:42:18,529
A compact set of
symbol of where

1113
00:42:18,529 --> 00:42:20,060
our compact set of pixels

1114
00:42:20,060 --> 00:42:22,310
should be similar, right?

1115
00:42:22,310 --> 00:42:24,110
And that would
seem to me to be

1116
00:42:24,110 --> 00:42:26,509
an important thing to
try to get away from.

1117
00:42:26,509 --> 00:42:28,220
Part of what we
have basically with

1118
00:42:28,220 --> 00:42:30,829
qualia is that we see
things as similar,

1119
00:42:30,829 --> 00:42:33,289
somehow unlinked
them in space.

1120
00:42:33,289 --> 00:42:35,089
We don't have a means
of doing that yet

1121
00:42:35,089 --> 00:42:38,434
in in datasets as
far as I know.

1122
00:42:38,434 --> 00:42:43,940
But that was that was
that stood out to me.

1123
00:42:43,940 --> 00:42:45,410
And it would seem
to me that that

1124
00:42:45,410 --> 00:42:47,389
would be an essential

1125
00:42:47,389 --> 00:42:49,384
to use that and
trying to predict

1126
00:42:49,384 --> 00:42:51,305
your predictive
model without that,

1127
00:42:51,305 --> 00:42:53,240
it seems to me to be.

1128
00:42:53,240 --> 00:42:55,895
It's hard to tell
how good your data

1129
00:42:55,895 --> 00:42:58,789
is without that in terms
of your prediction.

1130
00:42:58,789 --> 00:43:02,300
Yes, I agree.
Steve, I think,

1131
00:43:02,300 --> 00:43:04,400
I think bringing slope in

1132
00:43:04,400 --> 00:43:07,910
to the model is going
to help it quite a bit.

1133
00:43:07,910 --> 00:43:09,800
The two other things

1134
00:43:09,800 --> 00:43:11,539
I'd bring up is you
probably have got

1135
00:43:11,539 --> 00:43:15,934
data from other
industries on

1136
00:43:15,934 --> 00:43:19,399
some land-use questions
which may affect what

1137
00:43:19,399 --> 00:43:20,929
your pixels are
saying as far as

1138
00:43:20,929 --> 00:43:23,164
the frequencies that
are displaying.

1139
00:43:23,164 --> 00:43:26,554
Roads e.g. will, will

1140
00:43:26,554 --> 00:43:28,279
definitely affect
the way that picked

1141
00:43:28,279 --> 00:43:32,045
up a road would

1142
00:43:32,045 --> 00:43:34,580
have a reflection
right next to him.

1143
00:43:34,580 --> 00:43:36,110
Piece of non-road which

1144
00:43:36,110 --> 00:43:37,579
were basically
you'd want to

1145
00:43:37,579 --> 00:43:39,710
take the road out of
your geological map.

1146
00:43:39,710 --> 00:43:40,400
Oh, good.

1147
00:43:40,400 --> 00:43:41,719
I'm so glad that
you mentioned that,

1148
00:43:41,719 --> 00:43:42,620
Steve, that's one of

1149
00:43:42,620 --> 00:43:43,595
the things that I've been,

1150
00:43:43,595 --> 00:43:45,649
that that's another
improvement in the model

1151
00:43:45,649 --> 00:43:46,490
and I forgot to add to

1152
00:43:46,490 --> 00:43:47,975
my final thoughts slide.

1153
00:43:47,975 --> 00:43:49,399
And that is that I can use

1154
00:43:49,399 --> 00:43:51,020
masking to actually

1155
00:43:51,020 --> 00:43:52,430
mask out areas that

1156
00:43:52,430 --> 00:43:54,064
I don't want to
include in the model.

1157
00:43:54,064 --> 00:43:56,584
So I can say, take
out this road,

1158
00:43:56,584 --> 00:43:57,935
take out this forest.

1159
00:43:57,935 --> 00:44:00,139
This forest is just
confusing my models,

1160
00:44:00,139 --> 00:44:01,550
so let's just mask it out.

1161
00:44:01,550 --> 00:44:03,559
So yeah, we can totally
do stuff like that

1162
00:44:03,559 --> 00:44:05,765
to an agricultural areas.

1163
00:44:05,765 --> 00:44:06,770
Yep. Yep.

1164
00:44:06,770 --> 00:44:08,569
That would, I
think dramatically

1165
00:44:08,569 --> 00:44:10,040
improve things
to begin with.

1166
00:44:10,040 --> 00:44:10,805
Yeah.

1167
00:44:10,805 --> 00:44:12,859
That's I don't think I

1168
00:44:12,859 --> 00:44:14,675
have anything
else to say to

1169
00:44:14,675 --> 00:44:16,550
suggest except
that when you

1170
00:44:16,550 --> 00:44:19,535
mentioned wanna give
people a sense of size.

1171
00:44:19,535 --> 00:44:21,770
School bus is 40 ft long.

1172
00:44:21,770 --> 00:44:23,794
Oh, good. Thank You.

1173
00:44:23,794 --> 00:44:25,894
Do school but a
little bit about

1174
00:44:25,894 --> 00:44:28,610
two-thirds to
three-quarters of

1175
00:44:28,610 --> 00:44:30,290
two school buses
and links,

1176
00:44:30,290 --> 00:44:36,319
hey hey purse,
tennis courts.

1177
00:44:36,319 --> 00:44:43,715
The length of a tennis
court is 23.8 m. Okay.

1178
00:44:43,715 --> 00:44:48,485
So that's a good
illustration.

1179
00:44:48,485 --> 00:44:49,520
You were talking
about what's

1180
00:44:49,520 --> 00:44:51,559
20 m? Yeah,
that's perfect.

1181
00:44:51,559 --> 00:44:53,749
Yeah. A little
bit shorter than

1182
00:44:53,749 --> 00:44:56,150
the length of a
tennis court, right?

1183
00:44:56,150 --> 00:44:57,365
That's perfect. Yeah.

1184
00:44:57,365 --> 00:44:58,264
Thank you. Alright.

1185
00:44:58,264 --> 00:44:59,615
Now that I have before,

1186
00:44:59,615 --> 00:45:01,830
let me ask you
another question.

1187
00:45:01,870 --> 00:45:06,359
Could you have used

1188
00:45:06,490 --> 00:45:09,875
among the machine-learning

1189
00:45:09,875 --> 00:45:11,405
methods that you used?

1190
00:45:11,405 --> 00:45:13,894
Or is it the case that

1191
00:45:13,894 --> 00:45:15,080
because our outcome isn't

1192
00:45:15,080 --> 00:45:17,090
the insight kid or isn't

1193
00:45:17,090 --> 00:45:19,759
then the aren't assist
them that it is an

1194
00:45:19,759 --> 00:45:20,840
integrated into

1195
00:45:20,840 --> 00:45:23,705
this general
machine-learning,

1196
00:45:23,705 --> 00:45:25,864
a suite that you,

1197
00:45:25,864 --> 00:45:27,484
that'll be
difficult to use

1198
00:45:27,484 --> 00:45:29,585
Ockham to compare
it to say

1199
00:45:29,585 --> 00:45:32,299
random far as neural nets

1200
00:45:32,299 --> 00:45:34,385
and support
vector machines.

1201
00:45:34,385 --> 00:45:36,349
I think the tricky
thing with using

1202
00:45:36,349 --> 00:45:38,120
Aquaman this would
be coming up

1203
00:45:38,120 --> 00:45:39,530
with a rational way to

1204
00:45:39,530 --> 00:45:44,120
discretize the continuous
satellite signal.

1205
00:45:44,120 --> 00:45:46,249
Because we don't know

1206
00:45:46,249 --> 00:45:48,755
where to put the
class breaks in.

1207
00:45:48,755 --> 00:45:50,975
Okay, So, but
if we didn't,

1208
00:45:50,975 --> 00:45:53,120
if we did some analysis
on it and looked at

1209
00:45:53,120 --> 00:45:56,149
the distributions of
each of those signals.

1210
00:45:56,149 --> 00:45:57,379
We might be able
to come up with

1211
00:45:57,379 --> 00:45:58,715
reasonable class
breaks and

1212
00:45:58,715 --> 00:46:00,080
divide each one into like

1213
00:46:00,080 --> 00:46:05,074
55 things, then
we could do that.

1214
00:46:05,074 --> 00:46:06,499
And I did think

1215
00:46:06,499 --> 00:46:07,249
about that a couple of

1216
00:46:07,249 --> 00:46:08,149
days ago I was
thinking, you know,

1217
00:46:08,149 --> 00:46:11,299
aka might do a pretty
good job of predicting

1218
00:46:11,299 --> 00:46:13,639
these units if I can get

1219
00:46:13,639 --> 00:46:17,405
the data munge around
the right way,

1220
00:46:17,405 --> 00:46:20,869
you know, the
methods that you

1221
00:46:20,869 --> 00:46:25,475
use all treated the data
as continuous data.

1222
00:46:25,475 --> 00:46:27,499
Is that right? Yeah, yeah,

1223
00:46:27,499 --> 00:46:30,365
That market is higher.

1224
00:46:30,365 --> 00:46:32,089
And I just finish

1225
00:46:32,089 --> 00:46:35,749
this study where
we compare, OK,

1226
00:46:35,749 --> 00:46:38,210
on the two support
vector machines and

1227
00:46:38,210 --> 00:46:41,330
neural nets,
outcome did better.

1228
00:46:41,330 --> 00:46:44,240
And we had a
discrete too high as

1229
00:46:44,240 --> 00:46:47,240
data for outcome and
Bayesian networks.

1230
00:46:47,240 --> 00:46:49,129
And the support
vector machines

1231
00:46:49,129 --> 00:46:50,690
and neural nets treated

1232
00:46:50,690 --> 00:46:52,265
the data as continuous

1233
00:46:52,265 --> 00:46:54,529
and still outcome
did better.

1234
00:46:54,529 --> 00:46:57,020
So it's worth
thinking about.

1235
00:46:57,020 --> 00:46:59,135
It is, yeah.

1236
00:46:59,135 --> 00:47:03,290
The question is, you

1237
00:47:03,290 --> 00:47:07,579
took training
data on areas of

1238
00:47:07,579 --> 00:47:09,259
high resolution and then

1239
00:47:09,259 --> 00:47:11,794
you try to use that to

1240
00:47:11,794 --> 00:47:14,120
classify areas that are

1241
00:47:14,120 --> 00:47:16,654
at low resolution,
right? Or two, yes.

1242
00:47:16,654 --> 00:47:21,319
So wouldn't it be
desirable to kind of

1243
00:47:21,319 --> 00:47:25,160
test what you've learned

1244
00:47:25,160 --> 00:47:26,405
on the training data,

1245
00:47:26,405 --> 00:47:27,979
on something else that

1246
00:47:27,979 --> 00:47:29,539
you know the answers to

1247
00:47:29,539 --> 00:47:34,399
just to verify
that in fact,

1248
00:47:34,399 --> 00:47:39,170
your validly
applying it to

1249
00:47:39,170 --> 00:47:44,029
new to different
terrain? Yes.

1250
00:47:44,029 --> 00:47:46,429
But so you mean like

1251
00:47:46,429 --> 00:47:48,770
samples from the
actual well,

1252
00:47:48,770 --> 00:47:50,179
let's say if, if the

1253
00:47:50,179 --> 00:47:51,740
area used for the
training data,

1254
00:47:51,740 --> 00:47:54,305
if you divide
that into two,

1255
00:47:54,305 --> 00:47:57,200
you train on one
part and you test on

1256
00:47:57,200 --> 00:48:00,229
the other part of the
high resolution just

1257
00:48:00,229 --> 00:48:01,909
to be sure that in

1258
00:48:01,909 --> 00:48:03,799
fact it does
generalize even

1259
00:48:03,799 --> 00:48:08,239
within that that
similar terrain.

1260
00:48:08,239 --> 00:48:09,095
Right.

1261
00:48:09,095 --> 00:48:10,670
Maybe I should
do a train test

1262
00:48:10,670 --> 00:48:12,035
validate split.

1263
00:48:12,035 --> 00:48:14,989
Yeah. Yeah. Yeah,
that's a good idea.

1264
00:48:14,989 --> 00:48:17,420
I should do that
because right now yeah,

1265
00:48:17,420 --> 00:48:19,264
I'm just using the
train test splits

1266
00:48:19,264 --> 00:48:23,089
and getting getting in

1267
00:48:23,089 --> 00:48:29,330
the range of 70% accuracy
in the train. Yeah.

1268
00:48:29,330 --> 00:48:30,289
Right.

1269
00:48:30,289 --> 00:48:32,270
But, you know, you're
applying it to

1270
00:48:32,270 --> 00:48:34,624
a totally different
kind of terrain.

1271
00:48:34,624 --> 00:48:36,319
So you don't know,

1272
00:48:36,319 --> 00:48:38,270
you haven't checked that

1273
00:48:38,270 --> 00:48:40,219
you are training as well,

1274
00:48:40,219 --> 00:48:44,089
generalize well, even
to similar terrain?

1275
00:48:44,089 --> 00:48:45,739
Well, no, I think I'm

1276
00:48:45,739 --> 00:48:47,930
verifying it within
the training set.

1277
00:48:47,930 --> 00:48:49,340
I'm validating it.

1278
00:48:49,340 --> 00:48:50,614
Very right.

1279
00:48:50,614 --> 00:48:52,880
And I'm making sure
that the place that I'm

1280
00:48:52,880 --> 00:48:55,580
going to next is
related to it.

1281
00:48:55,580 --> 00:48:56,810
So I should be seeing

1282
00:48:56,810 --> 00:48:59,249
the same kinds of
stuff up there.

1283
00:48:59,920 --> 00:49:02,180
That was that
physiographic

1284
00:49:02,180 --> 00:49:03,649
Gavin terrains where I had

1285
00:49:03,649 --> 00:49:06,499
one tree over here and
one over here. Yeah.

1286
00:49:06,499 --> 00:49:09,484
So I am trying to I'm
trying to do that.

1287
00:49:09,484 --> 00:49:11,614
That's a good
idea. I should,

1288
00:49:11,614 --> 00:49:13,820
I should actually
spend more time

1289
00:49:13,820 --> 00:49:17,370
validating the data
from the training sets.

1290
00:49:21,610 --> 00:49:23,509
By Rick.

1291
00:49:23,509 --> 00:49:26,285
Oh, yeah. I have
a question.

1292
00:49:26,285 --> 00:49:27,244
Yeah. Oh, okay.

1293
00:49:27,244 --> 00:49:28,760
Yeah.

1294
00:49:28,760 --> 00:49:31,684
Did I read correctly
and you're upset that

1295
00:49:31,684 --> 00:49:33,829
part of part of it

1296
00:49:33,829 --> 00:49:36,319
included an ask or
raising awareness that

1297
00:49:36,319 --> 00:49:39,260
you all could use help
with this project.

1298
00:49:39,260 --> 00:49:41,585
Are you trying to
recruit students?

1299
00:49:41,585 --> 00:49:42,800
I would love to have

1300
00:49:42,800 --> 00:49:44,764
some students
involved with it.

1301
00:49:44,764 --> 00:49:47,359
So as I look at the read,

1302
00:49:47,359 --> 00:49:49,520
write territory
here, it seems

1303
00:49:49,520 --> 00:49:51,530
like you've got a lot
of work to be done.

1304
00:49:51,530 --> 00:49:54,679
How long do you think
this will take to flesh

1305
00:49:54,679 --> 00:49:57,799
out the red areas

1306
00:49:57,799 --> 00:50:01,529
with higher res
information?

1307
00:50:04,480 --> 00:50:06,739
What's the rate
of progress,

1308
00:50:06,739 --> 00:50:07,850
I guess is what
I'm saying.

1309
00:50:07,850 --> 00:50:09,559
I know. Yeah,
that's the problem.

1310
00:50:09,559 --> 00:50:10,174
Right.

1311
00:50:10,174 --> 00:50:11,870
Like so part of
my motivation

1312
00:50:11,870 --> 00:50:13,564
for this too is
that I've been

1313
00:50:13,564 --> 00:50:15,890
working with the geologic
map of Oregon for

1314
00:50:15,890 --> 00:50:17,120
like 20 years now since

1315
00:50:17,120 --> 00:50:18,709
I've been in
the department.

1316
00:50:18,709 --> 00:50:21,754
And it hasn't
changed much like

1317
00:50:21,754 --> 00:50:25,264
this level of poorly
mapped stuff.

1318
00:50:25,264 --> 00:50:27,245
And I'm not supposed
to say poorly mapped.

1319
00:50:27,245 --> 00:50:29,629
I'm actually, I'm
supposed to say mapped at

1320
00:50:29,629 --> 00:50:32,119
low resolution
because the mapping

1321
00:50:32,119 --> 00:50:34,295
that was done was
high-quality mapping.

1322
00:50:34,295 --> 00:50:36,199
It's just that it was
at a scale that's

1323
00:50:36,199 --> 00:50:38,510
not as good as what
we needed to be now.

1324
00:50:38,510 --> 00:50:40,805
So it's not that it
was poorly mapped.

1325
00:50:40,805 --> 00:50:42,650
It's just that it's
mapped at low resolution.

1326
00:50:42,650 --> 00:50:44,165
But anyway, yeah,
but the the,

1327
00:50:44,165 --> 00:50:45,499
the low resolution
areas of

1328
00:50:45,499 --> 00:50:47,059
the map have not
changed much.

1329
00:50:47,059 --> 00:50:49,519
And the priorities for
the state, of course,

1330
00:50:49,519 --> 00:50:50,599
are the places
where there's

1331
00:50:50,599 --> 00:50:52,639
high population density.

1332
00:50:52,639 --> 00:50:54,829
So I don't see us

1333
00:50:54,829 --> 00:50:56,509
getting the rest
of this map done

1334
00:50:56,509 --> 00:50:58,235
using our current methods

1335
00:50:58,235 --> 00:51:00,365
for 20 years or more.

1336
00:51:00,365 --> 00:51:02,270
So I'm hoping,
I'm hoping with

1337
00:51:02,270 --> 00:51:03,709
my satellite
approach that maybe

1338
00:51:03,709 --> 00:51:06,329
we can get it done
in like five years.

1339
00:51:06,430 --> 00:51:08,960
Well, my question is

1340
00:51:08,960 --> 00:51:10,129
a little self-serving
and I'll be

1341
00:51:10,129 --> 00:51:11,329
turning 60 years old

1342
00:51:11,329 --> 00:51:12,919
this year and a
couple of months.

1343
00:51:12,919 --> 00:51:14,179
And so I'm
starting to think

1344
00:51:14,179 --> 00:51:15,499
about retirement,

1345
00:51:15,499 --> 00:51:16,804
write down the road.

1346
00:51:16,804 --> 00:51:20,389
And I think I've harbored
this interest in

1347
00:51:20,389 --> 00:51:22,070
my retirement
of being one of

1348
00:51:22,070 --> 00:51:24,019
those volunteer
scientists that

1349
00:51:24,019 --> 00:51:25,819
goes out and
does fieldwork,

1350
00:51:25,819 --> 00:51:29,434
go help with biology
transects or so?

1351
00:51:29,434 --> 00:51:31,490
On Wednesday, I
chaired a big meeting

1352
00:51:31,490 --> 00:51:32,509
and it was nerve-wracking

1353
00:51:32,509 --> 00:51:33,650
and to reward myself,

1354
00:51:33,650 --> 00:51:36,425
I attended the
Wednesday geology talk

1355
00:51:36,425 --> 00:51:39,199
on local faultlines and

1356
00:51:39,199 --> 00:51:40,745
it was very good actually.

1357
00:51:40,745 --> 00:51:41,314
Yeah.

1358
00:51:41,314 --> 00:51:42,530
And she was so great.

1359
00:51:42,530 --> 00:51:44,569
It was just really
a rich talk and it

1360
00:51:44,569 --> 00:51:47,569
made me think
about weather.

1361
00:51:47,569 --> 00:51:49,190
There wouldn't be room for

1362
00:51:49,190 --> 00:51:50,450
someone not trained in

1363
00:51:50,450 --> 00:51:51,620
geology to go help you

1364
00:51:51,620 --> 00:51:53,450
with Huff at
data collection.

1365
00:51:53,450 --> 00:51:55,954
There's lots of room
for that, okay?

1366
00:51:55,954 --> 00:51:58,009
Yeah. In fact, actually

1367
00:51:58,009 --> 00:52:00,169
the US Geological
Survey has had

1368
00:52:00,169 --> 00:52:02,434
a thing for a
while called like

1369
00:52:02,434 --> 00:52:03,829
mapping core or geo

1370
00:52:03,829 --> 00:52:05,165
core or something
like that.

1371
00:52:05,165 --> 00:52:08,059
Where they just like we
would ask volunteers

1372
00:52:08,059 --> 00:52:11,450
to go out and collect
samples or GPS,

1373
00:52:11,450 --> 00:52:14,610
something in or whatever
needed to be done.

1374
00:52:23,860 --> 00:52:26,915
Percy, one of
your interests

1375
00:52:26,915 --> 00:52:30,184
you said was fractals.

1376
00:52:30,184 --> 00:52:35,659
What have you
thought at all?

1377
00:52:35,659 --> 00:52:38,480
To what extent
that these maps

1378
00:52:38,480 --> 00:52:40,474
fractal and can you,

1379
00:52:40,474 --> 00:52:43,070
can you use
fractal kind of

1380
00:52:43,070 --> 00:52:45,709
metrics to talk
about regions?

1381
00:52:45,709 --> 00:52:47,630
And is fractal?

1382
00:52:47,630 --> 00:52:51,889
Is the idea of fractals
relevant as possible?

1383
00:52:51,889 --> 00:52:54,845
I mean, we see
fractals a lot in

1384
00:52:54,845 --> 00:52:57,215
GIS in terms of
measuring coast

1385
00:52:57,215 --> 00:52:58,699
near the classic example

1386
00:52:58,699 --> 00:52:59,930
of the coastline, right?

1387
00:52:59,930 --> 00:53:01,099
Right.

1388
00:53:01,099 --> 00:53:05,824
And so what we
would probably,

1389
00:53:05,824 --> 00:53:07,850
we've seen this
actually in

1390
00:53:07,850 --> 00:53:09,845
our work with the
national land cover data

1391
00:53:09,845 --> 00:53:12,199
when we went from the
90 meter grid cell size

1392
00:53:12,199 --> 00:53:13,835
down to the 30-meter
grid cell size.

1393
00:53:13,835 --> 00:53:14,900
The patterns that we were

1394
00:53:14,900 --> 00:53:16,040
picking up when we had

1395
00:53:16,040 --> 00:53:17,734
90 meter resolution were

1396
00:53:17,734 --> 00:53:19,069
the same patterns
that we saw at

1397
00:53:19,069 --> 00:53:20,450
the 30 meter resolution,

1398
00:53:20,450 --> 00:53:22,190
but the model is a
little bit different.

1399
00:53:22,190 --> 00:53:23,899
But it was still
basically it

1400
00:53:23,899 --> 00:53:25,534
was that sort of thing.

1401
00:53:25,534 --> 00:53:29,105
So I would expect that

1402
00:53:29,105 --> 00:53:33,619
if we change the
resolution of the data,

1403
00:53:33,619 --> 00:53:37,669
that we would get
similar patterns.

1404
00:53:37,669 --> 00:53:38,570
Yeah.

1405
00:53:38,570 --> 00:53:41,130
But I'm not sure.

1406
00:53:41,470 --> 00:53:43,849
I don't know if that
would show up in

1407
00:53:43,849 --> 00:53:47,299
the actual band signals.

1408
00:53:47,299 --> 00:53:48,350
It's hard to see how that

1409
00:53:48,350 --> 00:53:50,164
would actually
express itself.

1410
00:53:50,164 --> 00:53:52,339
I mean, I'm I'm guessing
it would be there,

1411
00:53:52,339 --> 00:53:53,299
but I'm not sure how it

1412
00:53:53,299 --> 00:53:56,945
would it would
show itself.

1413
00:53:56,945 --> 00:53:58,699
Just because something is

1414
00:53:58,699 --> 00:54:01,370
observed is not mean,
it's predictive.

1415
00:54:01,370 --> 00:54:02,974
True.

1416
00:54:02,974 --> 00:54:06,290
And so we often
see fractals,

1417
00:54:06,290 --> 00:54:07,580
but it's not clear
how you'd use

1418
00:54:07,580 --> 00:54:09,960
the information
for prediction.

1419
00:54:13,360 --> 00:54:15,244
Yeah.

1420
00:54:15,244 --> 00:54:17,720
And I'm I'm, I'm curious

1421
00:54:17,720 --> 00:54:20,989
about you mentioned
the forest model.

1422
00:54:20,989 --> 00:54:22,220
Yeah.

1423
00:54:22,220 --> 00:54:27,035
Now, what is the
mathematical transform

1424
00:54:27,035 --> 00:54:29,195
it's talking about
and how is it used?

1425
00:54:29,195 --> 00:54:30,964
I haven't heard of this.

1426
00:54:30,964 --> 00:54:33,454
So random forest is

1427
00:54:33,454 --> 00:54:36,870
a collection of
decision trees.

1428
00:54:37,720 --> 00:54:41,104
So you start with
a decision tree.

1429
00:54:41,104 --> 00:54:43,669
You look at say, band
one and you say,

1430
00:54:43,669 --> 00:54:46,699
at what point in
band one does this

1431
00:54:46,699 --> 00:54:50,479
divide into unit
a versus unit B?

1432
00:54:50,479 --> 00:54:52,580
And then you do
that with Bantu

1433
00:54:52,580 --> 00:54:54,244
and bands or you make

1434
00:54:54,244 --> 00:54:55,939
decision cuts on each of

1435
00:54:55,939 --> 00:54:58,569
the variables and say,

1436
00:54:58,569 --> 00:55:03,950
I could show you
a decision tree,

1437
00:55:03,950 --> 00:55:06,080
right, this second binary

1438
00:55:06,080 --> 00:55:07,399
choice and each one.

1439
00:55:07,399 --> 00:55:09,380
And then you combine
the binary choices.

1440
00:55:09,380 --> 00:55:10,970
Yeah, Basically,
yeah, it's like a,

1441
00:55:10,970 --> 00:55:12,995
it's a, it's a,
it's an ensemble.

1442
00:55:12,995 --> 00:55:14,990
It's an ensemble
technique.

1443
00:55:14,990 --> 00:55:16,760
So each decision tree and

1444
00:55:16,760 --> 00:55:18,109
each decision
tree is actually

1445
00:55:18,109 --> 00:55:20,120
based on a subset
of the data.

1446
00:55:20,120 --> 00:55:22,430
So it doesn't use
the full dataset.

1447
00:55:22,430 --> 00:55:25,340
It uses maybe three
columns out of the ten.

1448
00:55:25,340 --> 00:55:27,620
And instead of using
all hundred and

1449
00:55:27,620 --> 00:55:28,999
44,000 rows and only

1450
00:55:28,999 --> 00:55:30,785
uses the first
20,000 rows.

1451
00:55:30,785 --> 00:55:32,555
And so each little
random forest,

1452
00:55:32,555 --> 00:55:34,670
each decision
tree is based on

1453
00:55:34,670 --> 00:55:37,070
some micro set
of the data.

1454
00:55:37,070 --> 00:55:38,599
And then each of

1455
00:55:38,599 --> 00:55:40,190
the little decision
trees votes

1456
00:55:40,190 --> 00:55:41,480
when it sees a new data.

1457
00:55:41,480 --> 00:55:44,990
And the ensemble
aggregates

1458
00:55:44,990 --> 00:55:46,010
those votes and then

1459
00:55:46,010 --> 00:55:47,689
comes up with a
final decision.

1460
00:55:47,689 --> 00:55:51,485
Does it use, does
it redo the rent,

1461
00:55:51,485 --> 00:55:52,940
the random development of

1462
00:55:52,940 --> 00:55:56,194
the Random Forest
repeatedly?

1463
00:55:56,194 --> 00:55:58,385
Know, you trade it up

1464
00:55:58,385 --> 00:56:00,769
iteratively until it
gets to the point

1465
00:56:00,769 --> 00:56:01,939
where it's a
stable model on

1466
00:56:01,939 --> 00:56:03,230
each iteration it starts

1467
00:56:03,230 --> 00:56:05,030
with a different
random forest.

1468
00:56:05,030 --> 00:56:08,195
Yeah. Okay.

1469
00:56:08,195 --> 00:56:10,759
And so then you basically,

1470
00:56:10,759 --> 00:56:12,139
so you could do that

1471
00:56:12,139 --> 00:56:13,670
on a relatively
small subset of

1472
00:56:13,670 --> 00:56:14,959
data and see
if it would it

1473
00:56:14,959 --> 00:56:16,670
would go to a
larger subset?

1474
00:56:16,670 --> 00:56:18,330
Definitely.

1475
00:56:19,150 --> 00:56:21,349
Okay. And whether or not

1476
00:56:21,349 --> 00:56:22,879
the weather, I'm sorry,

1477
00:56:22,879 --> 00:56:24,559
whether or not the
random forest first

1478
00:56:24,559 --> 00:56:27,110
developed would extend
is what I meant.

1479
00:56:27,110 --> 00:56:28,744
Yes. Okay, thanks.

1480
00:56:28,744 --> 00:56:30,125
Yeah.

1481
00:56:30,125 --> 00:56:32,720
So is, is the retina

1482
00:56:32,720 --> 00:56:36,395
just discretizing
the data?

1483
00:56:36,395 --> 00:56:43,279
Now, notice that it
decides where to cut.

1484
00:56:43,279 --> 00:56:45,590
Where to cut continua.

1485
00:56:45,590 --> 00:56:48,755
Yeah, let's discretize,

1486
00:56:48,755 --> 00:56:51,380
in effect,
discretize the data.

1487
00:56:51,380 --> 00:56:52,879
Yeah, you're right,
it does actually

1488
00:56:52,879 --> 00:56:54,200
kinda, well.

1489
00:56:54,200 --> 00:56:55,550
And then could
you use random,

1490
00:56:55,550 --> 00:56:58,550
Could you use the
random forest cuts

1491
00:56:58,550 --> 00:57:00,424
to discretize it for aka?

1492
00:57:00,424 --> 00:57:04,410
Exactly, Yeah,
that's a great idea.

1493
00:57:05,830 --> 00:57:08,029
Hey guys, I'm
gonna go ahead

1494
00:57:08,029 --> 00:57:09,244
and turn the
recording off.

1495
00:57:09,244 --> 00:57:10,219
You can hang around,

1496
00:57:10,219 --> 00:57:11,360
but this way we've
got a kind of

1497
00:57:11,360 --> 00:57:13,790
a nice stopping point
right here. Exactly.

1498
00:57:13,790 --> 00:57:15,480
Thanks, Wayne.