Post on 12-Jun-2020
Effect of Spatial Resolution on Characterizing Soil Properties from Imaging Spectrometer Data
Debsunder Dutta Praveen Kumar
Jonathan GreenbergUniversity of Illinois at Urbana Champaign
Presented at HyspIRI Science and Applications Workshop - Caltech
October 2015 1
ProblemStatementandResearchQuestions
1. Whatisthefeasibilityofquantificationofthesoilproperties/constituents usingairborneimagingspectroscopydata?
2. WhatistheEffectofSpatialResolution(ScalingUp)ontheCharacterizationofSoilConstituents?
2
Approach– FeasibilityofCharacterizingSoilConstituentsoverLargeAreas
3
500 1000 1500 2000 25000
2000
4000
Sand Silt Clay NDVI<=0.7
Wavelengths (nm)
Reflecta
nce%
(x100)
500 1000 1500 2000 2500
02000
4000
Sand Silt with NDVI<=0.7
Wavelengths (nm)
Reflecta
nce%
(x100)
Soils arecomplexHeterogeneoussystem
Generallyanover-determinedproblemwithp>>n
Developamodelingframeworkapplicable
overlargeareas
Evaluatetheconsistencyofresultsoverthelandscape
TakeadvantageoftheAVIRISspectra(224bands@10nm)coveringthefullVis-NIRandSWIRregionofthespectra
-sparse“narrowband”modelswillbeabletocapturesoilattributes
Gaininsightsintothemodelstructure
Characterizingsoilattributesoverlarge
areas
Veryfewfieldobservations
!" #!" $!!" $#!" %!!" %#!"
!"#!!"
$!!!"$#!!"%!!!"%#!!"&!!!"&#!!"'!!!"
&(#" )(#" $&(#" $)(#" %&(#" %)(#"
!"#$%&'()*+%
,*-*
./"#.*%0
1233
%45%
6"7*8*#9/:%0#(5%
!*&+,!*)%"-." !*)%,$*%)"-." $*%),$*+)"-." $*+),%*%(-." %*%(,%*#-."
100 data points76 data points with NDVI <0.7
training and test set split 80% and 20%
224 predictor set
10
Reduced set of 45 predictors
10 10 10 5
Observed Soil Texture Data
clay
silty clay
sandy clay
clay loam silty clay loam
sandy clay loam
loam
silty loam
sandy loam
siltloamy sandsand
102030405060708090
10
20
30
40
50
60
70
80
90
1020
3040
5060
7080
90
[%] Sand 50−2000 µm
[%] C
lay
0−2 µm
[%] S
ilt 2−50 µ
m
3 6
8
91011
12
13
14
16
1718
19
20
21
23
24
25
26
2728 29
30
31
3233
34
3536
37
38
39
40
42
4344
46
48
49
50
51
52
57
58
61
62
63
64
65
666768
71
72 75
76
80
81
8283
8485
86
87
8889
91
92
9394
96
98
99
100
Predicted Soil Texture Data
clay
silty clay
sandy clay
clay loam silty clay loam
sandy clay loam
loam
silty loam
sandy loam
siltloamy sandsand
102030405060708090
10
20
30
40
50
60
70
80
90
1020
3040
5060
7080
90
[%] Sand 50−2000 µm
[%] C
lay
0−2 µm
[%] S
ilt 2−50 µ
m
3
6
8
9
10
11
12
13
14
16
17
18
1920
21
23
24
2526 27
28
29
30
31
32
33
3435
36
37
38
39
40
42
4344
46
48 49
50
51
52
57
58
61
62
63
64
656667687172
75
768081
8283
84
85
86
87
8889
91
92
9394
96
98
99
100
Lasso 50 times
Bootstrapping
50 times for final model coefficients
Final Model Coefficients
-”lasso”methodsuitedforp>>n
y = β0 + β1X1 + β2X2 + ...+ βnXn
Applicationofmodelsspatiallyoverlargeareas
Capturesvariabilityinattributesspatially?
Duttaet.al.,OntheFeasibilityofCharacterizingSoilPropertiesFromAVIRISData, IEEETGRS,Sept2015:doi:10.1109/TGRS.2015.2417547
4
!( !(
!(
!(!(!( !( !(
!(!(!(
!(!(
!(!(
!(!(!(
!(!(!(!(!( !(!(!(
!( !(!(!(
!(!(!(
!(
!( !(
!(!(!(
!(
!(!(
!(!(!(!(
!(
!(!(!(!(
!(
!(!(!(!(!(
!( !(
!(!(!(
!(
!(
!(!(!(!(
!(
!(!(!(!(!(!(!(
!(
!(!(!( !(
!(
!(!(
!(!(!(!(!(!(
!(
!(!(
!(!(
!( !(
Sources: Esri, DeLorme, HERE, TomTom, Intermap, increment P Corp., GEBCO,USGS, FAO, NPS, NRCAN, GeoBase, IGN, Kadaster NL, Ordnance Survey, EsriJapan, METI, Esri China (Hong Kong), swisstopo, and the GIS User Community
0 10 205 Kms µ
Historic Meander Patternsof the Mississippi River
O' Bryan's Ridge
Region R1:
Region R2: DataandStudyRegion
• 27th July2011between14:04and15:00localtime(19:04– 20:00GMT).
• Altitude9.0– 9.1kmresultinginpixelresolutionof7.6m.
• Grabsamplesat100differentlocations.
• Labanalysisoftextureandchemicalconstituents, Organicmatter,Ca,Mg,K,Al,B,S,Fe,Zn,Cu,PandMn
Results– SpatialMapsofTextureandOrganicMatter
5
Results– SpatialOrganizationandLandscapeFeatures
6
Thespatialmapswhichrevealsconsistentspatialorganizationincludinglegacylandscapefeaturesandimmediatefinescaledisturbancesonthelandscape.
Results– SpatialMapsofChemicalConstituents
7
0 10 205 Kms
Mg [mg/kg]0 - 200
200 - 400
400 - 600
600 - 800
800 - 1,000
1,000 - 1,200
1,200 - 1,400
1,400 - 1,600
1,600 - 1,800
1,800 - 2,000
2,000 - 5,000
Erroneous Values
Water
Vegetation
0 10 205 Kms
Al [mg/kg]0 - 200
200 - 400
400 - 600
600 - 800
800 - 1,000
1,000 - 5,000
Erroneous Pixels
Water
Vegetation
0 10 205 Kms
Ca [mg/kg]0 - 1,000
1,000 - 2,000
2,000 - 3,000
3,000 - 4,000
4,000 - 5,000
5,000 - 6,000
6,000 - 7,000
7,000 - 8,000
8,000 - 15,000
Erroneous Pixels
Water
Vegetation
0 10 205 Kms
K [mg/kg]0 - 200
200 - 400
400 - 600
600 - 800
800 - 1,000
1,000 - 5,000
Erroneous Pixels
Water
Vegetation
Historic meander paths of the Mississippi River
Historic meander paths of the Mississippi River
Historic meander paths of the Mississippi River
Historic meander paths of the Mississippi River
GuidingResearchQuestions
1. Whatisthefeasibilityofquantificationofthesoilproperties/constituents usingairbornehyperspectraldata?
– “Lasso”algorithmbasedframeworkfoundtobefeasibletoquantifysoilconstituentsoverlargeareaswithlimitedsoilsampledataandfieldspectroscopy.
– Methodisapplicableequallywellforsoiltextureandchemicalconstituentsandprovidesspatialmapswhichrevealsconsistentspatialorganizationincludinglegacylandscapefeaturesandimmediatedisturbancesonthelandscape.
2. WhatistheEffectofSpatialResolution(ScalingUp)ontheCharacterizationofSoilConstituents?
– Feasibilityofapplicationofthedata-miningbasedmethodforquantifyingsoilconstituentsfromspacebasedsatelliteplatforms?
– Developingasuitablesetofmetricsforevaluationofperformanceandconsistencyofresultsacrossscales?
8
Distribution of spatial median and lasso
modeled value of soil constituent for a set of
pixels at L1
Within pixel distribution of soil constituent
Distribution of within pixel vari-ance for a set of
pixels at L1
Original
Upscaled Resolution (L1)
lasso predicted
value
pixel spectra
within pixel variance for a single pixel lo-cation for upscaled
resolution L1
2. Measure of deviation of soil 3. Error in prediction of
Approach forEvaluatingEffectofScaleontheCharacterizationofSoilConstituents
1
Convolution andresampling
Evaluatetheconsistencyofpointscaleresultsacross
scales
Evaluatetheconsistencyofspatialdistributionacross
scales
Within PixelVariance
ScaleupImagesfromFinetoCoarseResolutions
-applicationof”lasso”basedbootstrapframeworkatdifferentresolutions
DeviationofPredictionfrommedian
Pdfofconstituents
Refle
ctan
ce
Wavelength
Refle
ctan
ce
Wavelength
Refle
ctan
ce
Wavelength
Observed Value
Pred
icte
d Va
lue
Observed Value
Pred
icte
d Va
lue
Observed Value
Pred
icte
d Va
lue
Scal
ing
up o
f im
ages
from
fine
to c
oars
e sp
atia
l res
olut
ions
Upscaled pixel reflectances and model (lasso) structure at different spatial resolutions
Point scale results of soil-constituents (based on lasso method) and comparison with observatiional data
Eval
uate
the
cons
iste
ncy
of p
oint
scal
e re
sults
in te
rms o
f obs
erve
d vs
pre
dict
ion
plot
s, so
il te
xtur
e tr
iang
le a
nd m
odel
stru
ctur
e (b
and
impo
rtan
ce)
acro
ss d
iffer
ent s
patia
l res
olut
ions
X
X
f(x) [
pdf]
f(x) [
pdf]
pdf of soil constituent 1 and soil constituent 2
pdf of soil constituent 1 and soil constituent 2
Com
paris
on o
f pdf
of s
oil c
onst
ituen
ts a
cros
s ent
ire
land
scap
e (s
tudy
are
a) a
t diff
eren
t res
olut
ions
Difference of spa-tial median and lasso modeled
value of soil con-stituent
Spat
ial m
ap o
f soi
l con
stitu
ent
at d
iffer
ent r
esol
utio
ns
Within pixel distribution of soil constituent
Distribution of within pixel variance for a set of
pixels at L1
Original
Upscaled (L1)
Lasso predicted
Eval
uate
the
cons
iste
ncy
betw
een
lass
o m
odel
pre
dict
ed
and
stat
istic
al sc
alin
g of
soil
cons
titue
nts
at d
iffer
ent r
esol
utio
ns
Landscape of Area A under
Study
Comparison of point scale results and model structure based on scaling up imaging spectroscopy data for different resolutions
Comparison of spatial distribution of constituents obtained by model appli-cation to that obtained by statistical scaling for different resolutions
X
X
f(x) [
pdf]
pdf of soil constituent 1 and soil constituent 2
Observed Value
Pred
icted
Valu
eC
Refle
ctan
ce
Wavelength
SoilTextureTriangle
Observedvsmodelpredictions
ModelStructure
Duttaet.al.,inpreparation,2015
PointScaleEvaluationofResults– ObservedvsModelPrediction
10
0 20 40 60 80 100
02
04
06
08
01
00
Observed Sand [%]
Pre
dic
ted
Sa
nd
[%
]
+
+
+
+
+
+
+
++ +
+
+++
+
++ +
+
+ +
+ ++
+
+
+
+
+
+
+
+
+
+
+ ++
++
++
+
+
+
++
+++
+
+
+
++
+
+
+
+
+
+ +++
+++
+
+
+
+
+ +
+
+
+
++
+
+
+
+
+
+
++
+
+
+
+
++
+ : Sand
R2
0.691
30
50
70
% R
MS
Test
Err
or
0 20 40 60 80
02
04
06
08
0
Observed Silt [%]
Pre
dic
ted
Silt
[%]
+
+
+
+
+
+
+ ++
+
++
+ +
++
+
+
+
++
+
++
+
+ +++ +
+++
+
+
+
+
+
++
+
+
+
+
+
+
+
++
+
+
+
++ +
+
++
+
++++
+++
+ ++
+
++
+
++
++
+
+
+
+ +++
+
++
+
+
++
+ : Silt
R2
0.505
20
40
60
80
% R
MS
Test
Err
or
0 10 20 30 40 50 60
01
02
03
04
05
06
0
Observed Clay [%]
Pre
dic
ted
Cla
y [
%]
+
+
+
+
+
+
+
++
+
+
++
+++
+
++
+
+
+
++
+
+
+
+
+
+
+
+
++
++
+
+ +
++
+
++
+
+
++ +
++
+
+
++
+
+
+
+
++ ++++
+
+
+
+
+
+++
+
+ +
+
+
+
+
+
+
+
++
+
+
++
+ +
+ : Clay
R2
0.798
20
60
100
% R
MS
Test
Err
or
0 20 40 60 80 100 120
02
04
06
08
01
20
Observed Sand [%]
Pre
dic
ted
Sa
nd
[%
]
+
++
+ +
+++ ++
++
+
+
++
+
+
+
+
+ +++++
+ ++
+
+
+
+
+++
+
+
++
+
++
+
+
++++
+
++
+
+
+
+
+
+
+
+ +++
++ ++
+
+
+
+ +
+
+
+
++
+
++
++
+
+ +
+
+
++
++
+ : Sand
R2
0.736
30
50
70
% R
MS
Test
Err
or
0 20 40 60 80
02
04
06
08
0
Observed Silt [%]
Pre
dic
ted
Silt
[%] +
+
+
++
+
++
+
++
+ +++
++++
++
+ +
++
++
+
++
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
++
++
+
+
+
+
++
+
+
+
+
++++
++
+
+
+
+
+
++
++
+
++ +
++
+
+
+
+++
+ +
+ +
+
+ : Silt
R2
0.646
10
20
30
40
50
% R
MS
Test
Err
or
0 20 40 60
02
04
06
0
Observed Clay [%]
Pre
dic
ted
Cla
y [
%]
+
++
+
+
+
+
++
+
+ +
+
++
++
+
+
+
+
++
+++
+
++
+
+ +
+
++
+
+ +
+
+
+
+
+
+
+
++
+ ++
++
+
++
+
+
+
+
++ ++ ++
+
++
+
+
++
+
+
+++
+
++
++
+
++
+
+++
+ +
+ : Clay
R2
0.812
50
150
250
% R
MS
Test
Err
or
0 20 40 60 80 100
02
04
06
08
01
00
Observed Sand [%]
Pre
dic
ted
Sa
nd
[%
]
+
++
+
+
+++
+
++++
++
++
+
++
+
++
+
+ ++
++
+
+ ++
+
+
++ +
+
+
++
+
+++
+
+
+
+
++
+
++
+
+ +++
++
++ +
+ +
++
++
+
+
++
++
+
+
+
++
+ +
+
+
+ : Sand
R2
0.644
50
150
250
% R
MS
Test
Err
or
0 20 40 60 80
02
04
06
08
0
Observed Silt [%]
Pre
dic
ted
Silt
[%]
+
++
+
+
+
++
+
+
+
+
+
+
+
+
++ +
+
+
+
+
++
+
++
+
+
+
+
+
+
+
+
++
+
+
++
+
+
++
+
++
+ +
+
+
++
+
++++
++
+
+
+
++
+ +
+
+
++
++
+
+
+ ++
+
+
+
+
++
+ : Silt
R2
0.686
20
40
60
80
% R
MS
Test
Err
or
0 20 40 60 80
02
04
06
08
0
Observed Clay [%]
Pre
dic
ted
Cla
y [
%]
+
+
+++
+
++
+
+
+
+
++
++
+
+
+
+
+
+++
+
+
+
+ ++
+
+
+
+
++++
+
+
+
+
++
+ +
+++
+
++
+
+ +
+
++ ++++
+
+
+
+++
+
+ ++
+
++
++
+
++
+++
+
+ +
+ : Clay
R2
0.791
0100
200
300
% R
MS
Test
Err
or
(a) 15.2m (b) 30.4m (c) 60.8m
0 2 4 6 8
02
46
8
Observed SOM [%]
Pre
dic
ted
SO
M [
%]
+
+
+
+
++
+
++++
++
+
+
++
+
+ ++
+
+
+
+
+
+
++
+++
+
+
++
+
+
+
+
+
+
+
+
+
+
++ ++
+
+
+
+
+
++
+
+
++ ++
+++
+
+
++
+++
+
+
+
+
+
++
+ +
++
++
+
+
+
+
+
+ : SOM
R2
0.732
10
20
30
40
50
% R
MS
Test
Err
or
0 2 4 6 8
02
46
8
Observed SOM [%]
Pre
dic
ted
SO
M [
%]
+
+
++
++
+++
+
+++
+
+
++
+
++
+ +
+
+
+
+
+
+++
+
+
+
++ +
+++
+
+
+
+
+
++++ +
++
+
++
+
++
+
+
++ ++
++++
+
+
+++
+
+
++
+
+
++
++
++
+
+
+
+
++
+
+ : SOM
R2
0.695
20
40
60
80
% R
MS
Test
Err
or
0 2 4 6 8
02
46
8
Observed SOM [%]
Pre
dic
ted
SO
M [
%]
+
+ +
++
+
+++
+ +
+
++
+
+
+
++
++
++
+
++
+
+
+
+
++
+
+
++
++
+
+
+
++++ ++
+
+
+
+
+
+
+
+
+
++ ++
++
+
+
+++
+
+
+
++
+
++
+
+
+
+
+
+
+
+
+ + ++ : SOM
R2
0.718
50
100
200
% R
MS
Test
Err
or
10m 15.2m 20m 30.4m 45m 60.8m 90mSand 0.689 0.691 0.655 0.736 0.490 0.644 0.657Silt 0.564 0.505 0.540 0.646 0.596 0.686 0.361Clay 0.781 0.798 0.799 0.812 0.781 0.791 0.762SOM 0.638 0.732 0.659 0.695 0.533 0.718 0.557Ca 0.784 0.738 0.745 0.758 0.712 0.739 0.748Mg 0.791 0.798 0.750 0.751 0.744 0.767 0.705K 0.700 0.691 0.696 0.653 0.687 0.698 0.658Al 0.774 0.814 0.737 0.776 0.709 0.760 0.748Fe 0.579 0.535 0.624 0.425 0.453 0.586 0.464
SummaryofR2
valuesforalltheresolutions
PointScaleEvaluationofResults– SoilTextureTriangles
11
clay
silty claysandy clay
clay loam silty clay loam
sandy clay loam
loamsilty loam
sandy loamsiltloamy sand
sand
102030405060708090
10
20
30
40
50
60
70
80
90
1020
3040
5060
7080
90
[%] Sand 50−2000 μm
[%] C
lay
0−2
μm
[%] Silt 2−50 μm
2
3
4
6
7
8
9
10 11
12
13
14
1516
1718
19
20
21
22
23
24
2526
27
28
29
30
31
32
33
34
3536
3738
39
40 42
4344
46
4849
50
51
525354
5556
57
58
5960
61
62
63
656667687172
73
75
76
78
79
808182
83
8485
86
87
88
89
90
91
92
93 94
95
96
9798
99100
clay
silty claysandy clay
clay loam silty clay loam
sandy clay loam
loamsilty loam
sandy loamsiltloamy sand
sand
102030405060708090
10
20
30
40
50
60
70
80
90
1020
3040
5060
7080
90
[%] Sand 50−2000 μm
[%] C
lay
0−2
μm
[%] Silt 2−50 μm
2
3 6
7
8
9
10
11
12
13
1415
16
1718
1920
21
22
23
24
2526
27
2829
30
31
3233
34 35
36
37
38
39
4042
43
44
46
47
48
49
50
5152
5354
55
56
57
58
5960
61
62
63
656667687172
73
7475
76
78
8081
82
83
8485
86
87
8889
9091
92
9394
95
9697
98
99100
clay
silty claysandy clay
clay loam silty clay loam
sandy clay loam
loamsilty loam
sandy loamsiltloamy sand
sand
102030405060708090
10
20
30
40
50
60
70
80
90
1020
3040
5060
7080
90
[%] Sand 50−2000 μm
[%] C
lay
0−2
μm
[%] Silt 2−50 μm
2
3
6
78
9
10
11
12
13
14
15
16
18
19
20
21
22
23
24
25
2628
29
30
31
33
3435
36
37
38
39
40
4142
4344
46
48
49
5152
5354
55
5657
58
5960
61
6263
656667687172
75
78
8081
82
83
848586
87
88
90
91
92
9394
9596
9798
99100
Observed Soil Texture Data Predicted Soil Texture Data
clay
silty claysandy clay
clay loam silty clay loam
sandy clay loam
loamsilty loam
sandy loamsiltloamy sand
sand
102030405060708090
10
20
30
40
50
60
70
80
90
1020
3040
5060
7080
90
[%] Sand ��ï���� μm
[%] C
lay
0ï�
μm
[%] Silt �ï50 μm
3 6
8
91011
12
13
14
16
1718
19
20
21
23
24
25
26
2728 29
30
31
3233
34
3536
37
38
39
40
42
4344
46
48
49
50
51
52
57
58
61
62
63
64
65
666768
71
72 75
76
80
81
8283
8485
86
87
8889
91
92
9394
96
98
99
100
clay
silty claysandy clay
clay loam silty clay loam
sandy clay loam
loamsilty loam
sandy loamsiltloamy sand
sand
102030405060708090
10
20
30
40
50
60
70
80
90
1020
3040
5060
7080
90
[%] Sand ��ï���� μm
[%] C
lay
0ï�
μm
[%] Silt �ï50 μm
3
6
8
9
10
11
12
13
14
16
17
18
1920
21
23
24
2526 27
28
29
30
31
32
33
3435
36
37
38
39
40
42
4344
46
48 49
50
51
52
57
58
61
62
63
64
656667687172
75
768081
8283
84
85
86
87
8889
91
92
9394
96
98
99
100
(a) (b)
(c) (d) (e)
(a)Theobserved(b)modelpredictedpointsat7.6mairborneAVIRISresolution(c)modelpredictedup-scaled15.2m (d)30.4mand(e)60.8Thesamplenumbersareindicatedoneachofthedots.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. XX, NO. X, XXXX 2015 25
TABLE IDESIGN KERNEL SIZES FOR UPSCALING THE AVIRIS DATA
Convolve and Resample Resolution Kernel Size
10 m 3⇥ 315.2 m 5⇥ 520 m 7⇥ 730.4 m 11⇥ 1145 m 15⇥ 1560.8 m 21⇥ 2190 m 29⇥ 29
TABLE IITHE DIFFERENT PARAMETERS USED FOR ATMOSPHERIC CORRECTION OF AVIRIS SCENES
Scene 1 Scene 2 Scene 3 Scene 4 Scene 5 Scene 6
Flight Altitude (km) 9.092 9.108 9.088 9.072 9.093 9.108Flight Heading (deg) 42.451 214.64 44.16 210.33 46.39 210.01Ground Elevation (km) 0.0914 0.09095 0.0924 0.09412 0.0961 0.0994Solar Zenith Angle (deg) 44.24 40.8 37.5 34.3 31.2 28.1Solar Azimuth Angle (deg) 100.29 103.6 107.3 111.3 116.1 121.5Pixel Size (m) 7.6 7.6 7.6 7.6 7.6 7.6
TABLE IIICLASSIFICATION ACCURACY OF USDA MODELED SOIL TEXTURE CLASSES AS COMPARED WITH OBSERVED SOIL TEXTURE CLASSES
Spatial Resolution Exact Classification[%] Close Classification[%] Incorrect Classification[%]
10 m 46.67 28.89 24.4415.2 m 42.86 38.46 18.6820 m 40.22 40.22 19.5730.4 m 43.96 43.96 12.0945 m 43.33 36.67 20.0060.8 m 46.51 40.70 12.7990 m 37.35 33.73 28.92
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. XX, NO. X, XXXX 2015 6
originally used to identify them during data collection andlaboratory analysis, and (b) model predicted textural results at7.6m original airborne AVIRIS resolution. Fig. 8(c) through8(e) show the model predicted textural results using upscaleresolution data at 15.2 m, 30.4 m and 60.8 m respectively.
1) Effect of coarsening the spatial resolution on modelprediction results for soil texture and organic matter?: Fromthe USDA soil texture triangles (Fig. 8) it is found that thelaboratory and modeled soil classification results agree well.The spread of the data in different classes is the same forboth observed and modeled results and across the data at thevarious spatial resolutions. Moreover, inspection of the datalabels and comparison with the observed data reveals thatmost of them are classified correctly, and we categorized theclassifications as ‘Exact Match’, ‘Close Classification’, and‘Incorrect Classification’ defined as follows.
1) If the observed and the model predicted soil propertiesbelong to the same USDA soil texture class, we call ita coincident match or exact classification.
2) If a soil property does not fall in the same categorywe compute the deviation in the observed and modelpredicted values for sand and clay percentages for thesample and the total deviation as:
�
sand%
= |�observed
sand%
��
predicted
sand%
|
�
clay%
= |�observed
clay%
��
predicted
clay%
|
�
total%
= �
sand%
+�
clay%
3) If the total deviation (�total%
) is less than or equal to25% we call it a ‘close’ match, otherwise we call it an‘incorrect’ classification.
The results of the classification analysis for all the differentspatial resolutions are presented in Table. III. The resultsindicate that across all the different spatial resolutions (exceptthe large 90m pixel resolution) about 40 - 47 % of the samplesare classified exactly and about 28 - 41% samples indicateclose classification with the USDA soil texture classes. Theseresults show us that the ensemble lasso method performswell not only for individual models but also for all the threedifferent soil texture models combined together.
B. Model Structure of Soil Constituents at Different SpatialResolutions
The model structure is investigated in terms of the explana-tory predictor variables (wavelengths) for the different soilconstituents at all of the eight different spatial resolutions(namely 7.6m, 10m, 15.2m, 20m, 30.4m, 45m, 60.8m and90m) used for the study. For a particular soil constituent itis interesting to know the forms of inter-relationships betweenpredictor bands of the models as we upscale the data. Figure9 shows the relationships between model structure (predictorbands) at different spatial resolutions for clay, silt, sand andsoil organic matter. It Illustrates which bands participatesin the prediction model (eq. 9) across different resolutions.
When a band participates in at least 3 different resolutionsit is identified in Fig. 9(a). Similarly when it participatesin atleast 4 different resolutions it is identified in Fig. 9(b).Figs. 9(c) and (d) represent participation in 5 and 7 differentresolutions. It is found that there are a number of bands whichemerge as important predictors across most of the spatialresolutions. A number of bands which are important predictorfor not one constituent but for a couple or all of the fourconstituents. The model structure for soil chemical constituentscalcium, magnesium, potassium and aluminum show similarpatterns [see supplementary info. fig. S2 (a) - (d)]. The bandswhich are predictor variables for the combination of otherspatial resolutions for the textural and chemical properties arepresented in the supplementary material [see figs. S3 and S4].
It is found that the number of common predictors acrossmany different spatial resolutions are higher for the chemicalconstituents than the soil texture models. The summary figurefor the importance of wavelengths across different scalesfor textural as well as chemical constituents is presented infig.10. We find that wavelengths in the blue region of thespectrum (360 - 460 nm) is important for all the texturalproperties and chemical constituents. The empirical modelsfor the prediction of the different soil constituents are basedon spectral characteristics of pure elements, compounds andradicals as well as the intercorrelations between differentwavelengths as soil is a complex mixture. Even in an automaticmodel development framework we have found that some ofthe pure spectroscopic characteristics are well pronounced inthe models. For the clay models the 2.2µm feature which is ahydroxyl absorption feature and is a signature of clay mineralskaolinite and montmorillonite is found to be a predictor acrossthe models for all the spatial resolutions. Some of the otherfeatures for these minerals such as the 1.4 µm water absorptionis captured across some of the spatial resolutions. Similarlyfor the chemicals a close examination [fig. 10] reveals that thebands at 1.1, 1.4 and 2.2 µm are selected as predictor bands forcalcium and magnesium across many of the spatial resolutions.These band are associated with and are specific signaturesof Ca-OH and Mg-OH. Thus it may be concluded that theempirical modeling framework selects the important spectralsignatures associated with a chemical constituent or a texturalproperty as well as utilizes an underlying correlation structureamong various bands across upscaled spatial resolutions forquantification of the soil constituents. The results of sectionIV - A and B indicate that the model results are consistentand establishes the potential applicability of the method acrossdifferent spatial resolutions (from airborne to space borne) forquantification of soil texture.
C. Spatial Distribution of Soil Constituents Across the Land-scape at Different Resolutions
1) Change in prediction maps of soil constituents overlarge areas due to coarsening spatial resolutions?: The lassoprediction models developed at different spatial resolutions(for soil texture and organic matter) were applied on a pixelby pixel basis for obtaining quantitative spatial maps of theconstituents over the entire floodplain. These maps help us to
1. IftheobservedandthemodelpredictedsoilpropertiesbelongtothesameUSDAsoiltextureclass,wecallitacoincidentmatchorexactclassification.
2. Otherwisewecomputethetotaldeviation.
1. Ifthetotaldeviation(∆total%)islessthanorequalto25%wecallita‘close’classification,otherwisewecallitan‘incorrect’classification.
PointScaleEvaluationofResults– ModelStructure
12
Sand Silt
Clay OrganicMatter
Wavelength [nm]
Num
ber o
f Spa
tial R
esol
utio
ns
1
2
3
4
5
6
7
Band
1
Band
11
Band
22
Band
33
Band
44
Band
55
Band
66
Band
77
Band
88
Band
99
Band
110
Band
121
Band
132
Band
143
Band
154
Band
165
Band
176
Band
187
Band
198
Band
209
Band
220
365
462
569
655
763
870
976
1082
1187
1273
1382
1492
1602
1711
1821
1907
2017
2127
2237
2347
2456
Wavelength [nm]
Num
ber o
f Spa
tial R
esol
utio
ns
1
2
3
4
5
6
7
Band
1
Band
11
Band
22
Band
33
Band
44
Band
55
Band
66
Band
77
Band
88
Band
99
Band
110
Band
121
Band
132
Band
143
Band
154
Band
165
Band
176
Band
187
Band
198
Band
209
Band
220
365
462
569
655
763
870
976
1082
1187
1273
1382
1492
1602
1711
1821
1907
2017
2127
2237
2347
2456
Wavelength [nm]
Num
ber o
f Spa
tial R
esol
utio
ns
1
2
3
4
5
6
7
Band
1
Band
11
Band
22
Band
33
Band
44
Band
55
Band
66
Band
77
Band
88
Band
99
Band
110
Band
121
Band
132
Band
143
Band
154
Band
165
Band
176
Band
187
Band
198
Band
209
Band
220
365
462
569
655
763
870
976
1082
1187
1273
1382
1492
1602
1711
1821
1907
2017
2127
2237
2347
2456
Wavelength [nm]
Num
ber o
f Spa
tial R
esol
utio
ns
1
2
3
4
5
6
7
Band
1
Band
11
Band
22
Band
33
Band
44
Band
55
Band
66
Band
77
Band
88
Band
99
Band
110
Band
121
Band
132
Band
143
Band
154
Band
165
Band
176
Band
187
Band
198
Band
209
Band
220
365
462
569
655
763
870
976
1082
1187
1273
1382
1492
1602
1711
1821
1907
2017
2127
2237
2347
2456
(a) (b)
(c) (d)
SpatialDistributionofSoilConstituentsAcrosstheLandscape
13
0 1 20.5 Kmsµ 0 1 20.5 Kmsµ
0 1 20.5 Kmsµ 0 1 20.5 Kmsµ
(a) (b)
(c) (d)
!( !(
!(
!(!(!( !( !(
!(!(!(
!(!(
!(!(
!(!(!(
!(!(!(!(!( !(!(!(
!( !(!(!(
!(!(!(
!(
!( !(
!(!(!(
!(
!(!(
!(!(!(!(
!(
!(!(!(!(
!(
!(!(!(!(!(
!( !(
!(!(!(
!(
!(
!(!(!(!(
!(
!(!(!(!(!(!(!(
!(
!(!(!( !(
!(
!(!(
!(!(!(!(!(!(
!(
!(!(
!(!(
!( !(
Sources: Esri, DeLorme, HERE, TomTom, Intermap, increment P Corp., GEBCO,USGS, FAO, NPS, NRCAN, GeoBase, IGN, Kadaster NL, Ordnance Survey, EsriJapan, METI, Esri China (Hong Kong), swisstopo, and the GIS User Community
0 10 205 Kms µ
Historic Meander Patternsof the Mississippi River
O' Bryan's Ridge
Region R1:
Region R2:
7.6m 60.8m
0 8 164 Kms
Clay (60.8m) [%]0 - 20
20 - 40
40 - 60
60 - 80
80 - 100
Erroneous Values
Water
Vegetation
0 8 164 Kms
Clay (7.6m) [%]0 - 20
20 - 40
40 - 60
60 - 80
80 - 100
Erroneous Pixels
Water
Vegetation
0 8 164 Kms
SOM (7.6m) [%]0 - 1
1 - 2
2 - 3
3 - 4
4 - 10
10 - 40
Erroneous Pixels
Water
Vegetation
0 8 164 Kms
SOM (60.8m) [%]0 - 1
1 - 2
2 - 3
3 - 4
4 - 10
10 - 40
Erroneous Values
Water
Vegetation
0 8 164 Kms
Mg (7.6m) [mg/kg]0 - 200
200 - 400
400 - 600
600 - 800
800 - 1,000
1,000 - 1,200
1,200 - 1,400
1,400 - 1,600
1,600 - 1,800
1,800 - 2,000
2,000 - 5,000
Erroneous Values
Water
Vegetation
0 8 164 Kms
Mg (60.8m) [mg/kg]0 - 200
200 - 400
400 - 600
600 - 800
800 - 1000
1000 - 1200
1200 - 1400
1400 - 1600
1600 - 1800
1800 - 2000
2000 - 5000
Erroneous Values
Water
Vegetation
SpatialDistributionofSoilConstituentsAcrosstheLandscape- pdfs
14
� � � � � � � � � � � � ��
��
��
��
��
��
��
��
��
��
�
��
��
��
��
�
��
��
��
��
��
�
� ��
� ��
�10−3.5
10−3
10−2.5
10−2
10−1.5
0 25 50 75 100X (Sand [%])
f(x
) [p
df]
Resolution
� 7.6m
10 m
15.2m
20 m
30.4 m
45 m
60.8 m
90 m
� � ��
��
��
��
��
��
��
��
�� � � � � � � � � �
��
��
��
��
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
�
� �
�
� �
10−5
10−4
10−3
10−2
0 25 50 75 100X (Clay [%])
f(x
) [p
df]
Resolution
� 7.6m
10 m
15.2m
20 m
30.4 m
45 m
60.8 m
90 m
� ��
��
��
��
��
��
��
��
� � � � � � � � � � � �� � �
��
��
��
��
��
��
��
��
��
��
��
�
��
��
�
10−3.5
10−3
10−2.5
10−2
10−1.5
0 25 50 75 100X (Silt [%])
f(x
) [p
df]
Resolution
� 7.6m
10 m
15.2m
20 m
30.4 m
45 m
60.8 m
90 m
� � ��
��
��
��
��
�� � � � � � � �
��
��
��
��
�
��
�
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
�
��
�
�
��
�� �
10−3.5
10−3
10−2.5
10−2
10−1.5
10−1
10−0.5
0.0 2.5 5.0 7.5 10.0X (SOM [%])
f(x
) [p
df]
Resolution
� 7.6m
10 m
15.2m
20 m
30.4 m
45 m
60.8 m
90 m
Region:R1
SpatialDistributionofSoilConstituentsAcrosstheLandscape–DeviationfromStatisticalCentralvalues
15
● ●
●
●
●
●
●
●
●
●
●
0
40
80
120
Cla
y [%
]
PropertyModel ValuesSpatial Median
●
●●
●
●
● ●
3
6
9
10m 15.2m 20m 30.4m 45m 60.8m 90mScale
∆[%
]
● ●
●
● ●
●
●
●
●
0
5
10
15
20
SOM
[%]
PropertyModel ValuesSpatial Median
●
●● ●
●
●
●0
2
4
6
10m 15.2m 20m 30.4m 45m 60.8m 90mScale
∆[%
]
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●● ●●●●
●●●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
2 11 20 29 38 48 56 64 76 86 940
50100
150
Gauss60.8m
Sample Number
Clay [%
]
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●●
●
●●●●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●●●●●
●
●
●●
●
●
●
●●
●●●
●
●
●
●
●●●
●
●
●
● ●
●●
●
●
●
●●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●●
●
●●
●
●●
●
●●●
●
●●
●
●●●
2 11 20 29 38 48 56 64 76 86 94
05
1015
2025
Gauss60.8m
Sample Number
SOM [
%]
●
●●
●●
●
●
●●
●●
●
●●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●●●●●
●
●
●
●
●
●
●
●
●●●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
● ●
●
●
●
●
●
●
●
●●
●
●●
0
1000
2000
3000
4000
Mg
[mg/
kg]
PropertyModel ValuesSpatial Median
● ●
●
● ●●
●
7.510.012.515.017.5
10m 15.2m 20m 30.4m 45m 60.8m 90mScale
∆[%
]
●
●
●●●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
0
500
1000
1500
2000
2500
Fe [m
g/kg
]
PropertyModel ValuesSpatial Median
●
●
●● ●
● ●
68101214
10m 15.2m 20m 30.4m 45m 60.8m 90mScale
∆[%
]●
●
●●
●●●●
●●
●●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
2 11 20 29 38 48 56 64 76 86 94
0500
1000
1500
2000
2500
3000
Gauss60.8m
Sample Number
Mg [m
g/kg]
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●●●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●● ●
●●
●
●●●
●●
●
2 11 20 29 38 48 56 64 76 86 94
0500
1000
1500
2000
2500
3000
Gauss60.8m
Sample NumberFe
[mg/kg
] ●
●●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●●●●●
●●
●
●●●●
●
●
●
●
●●●●●
●
●
●
●
●
●
●
●
●●●●
●●
●●●
●●●
●●
●
●●
●
●
●
●●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
1.89 1.69 1.8 2.11 2.12.71 2.36
1e−02
1e+00
1e+02
1e+04
10m 15.2m 20m 30.4m 45m 60.8m 90mScale
With
in P
ixel
Var
ianc
e SO
M [
sq. %
]
SpatialDistributionofSoilConstituentsAcrosstheLandscape–WithinPixelVariances
16
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
134.15 158.76 146.64193.4 192.13 200.57 213.67
1e+01
1e+03
1e+05
10m 15.2m 20m 30.4m 45m 60.8m 90mScale
With
in P
ixel
Var
ianc
e Cl
ay [
sq. %
]
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●●
●
47287.9738983.63 40914.03
53163.86 54002.9 59270.66 65987.151e+05
1e+07
10m 15.2m 20m 30.4m 45m 60.8m 90mScale
With
in P
ixel
Var
ianc
e M
g [ s
q. m
g/kg
]
●●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●●
47001.68 49531.15 54595.9 52872.23 54170.34 62898.28 57984.01
1e+04
1e+06
10m 15.2m 20m 30.4m 45m 60.8m 90mScale
With
in P
ixel
Var
ianc
e Fe
[ sq
. mg/
kg]
SummaryandConclusions
• Lassoalgorithmbasedmodeling frameworkisapplicableacrossmultiple scalesfromfine tocoarsespatialresolutions.
• Themodel structureacrossmultiple resolutions revealsthatimportantspectralfeaturessuchaswaterabsorption,minerals(clay,OH-,CO3-)arerepresentedacrossmultiple resolutions.
• Thepoint scaleresultsandthewithinpixelvarianceofconstituentsarefound tobeconsistentacrossscales.
• Thepdf oftheconstituentsarealsofound tobesimilaracrossscaleswithslightshiftinthemodesforsomeconstituents.
• Thelassobasedquantificationmethodhasthepotentialtobeapplicablefromspace-basedsensorssuchasHyspIRI.
17
Thankyou!
18