Landsat Spectral Signatures: Studies with Soil Associations

F. C. WESTI;\,
G. D. LUI~IE'"
South Dakota State Uniuersity
Brookings, SD 57006
Landsat Spectral Signatures:
Studies with Soil Associations
and Vegetationt
An accuracy exceeding 90 percent was obtained in
separating four categories of land use for each of six soil
associations.
I:-ITRODUCTION
A:\DSAT I~IAGERY has temporal, synoptic,
and multispectral characteristics possible useful fin large-scale crop inventories.
One such ill\-entory experiment presently
L
is the stratification of regions into relatively
homogeneous partitions within which production tends to he uniform.
The neecl in crop in"entor~- work to partition regions into areas as homogeneous as
ABSTRACT: The effect of soils and uegetation upon Landsat spectral
properties was inuestigated for a 12,950 hectare studu area. Six soil
associations used for corn, small grains, a nd grass were analuzed for
tluO dates during the 1974 growing season. Landsat scenes for April
19 and June 30 lcere studied to try to separate categories of agricultured land use and to assess the inj7uence of soil association on the
spectral signatures of uegetation and bare ground. The April 19 data
were useful to separate cropland from grassland and to locate areas
of open water. Soil differences had a more pronounced influence on
the spectral properties of grassland than on cropland. The JUlie 30
data showed that soil associations could not consistentlu be separated within the data of a single uegetatiue type; howeuer, the results showed that soils did influence all vegetative spectral re.f7ectances to some degree. Because soils did influence vegetative spectrpl reflectance, a generalized training set containing data points
from each of the six soil associations was used to se/wrate four
categories of agricultural land use in the 12,950 hectare test area. An
accuracu of about 94 percelll was obtained.
being conclucted is the LACIE Project, described b~' ~lacDonald (1976). LACIE is designed to estimate wheat production at a regional or countrv level. A pmt of the procedure
'" l'\ow with the Uni"ersitv of Nebraska, Lincoln, Nebraska.
t Authorized for publication as Journal Series
No. 1501, South Dakota Agricultural Experiment
Station.
PHOTOGRAM~IETRIC E;-';GDIEERING
possible was demonstrated by vVigton and
Von Steen (1973) who obtained inconclusive
results when attempting to iIl\'entory crops
on a regional basis. The~' found that training
site selection was difficult because of the di,-ersity of data from a particular crop. Hoffer
et al. (1966) obsen'ed that variation in soils
can cause marked differences in the signatures ofa single crop. And ~dyers et al. (1974)
found that the ratio of Band 5 divided b,'
Band 7 was different for corn and oats that
A:-ID RD10TE SE:-ISING,
Vol. 44, No.3, March 1978, pp. 315-325.
315
316
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING, 1978
were grown on soils of different textures and
landscape positions.
Boundaries on Landsat color composites
separating different cropping patterns often
coincide with soil association boundaries.
Westin and Frazee (1976) used Landsat
scenes to delineate soil association boundaries by using photo interpretation
techniques. Soil associations are groups of
defined and named taxonomic soil units occurring together in an individual and characteristic pattern over a geographic area comparable in many ways to plant associations.
Thus, soil associations partition regions into
relatively homogeneous areas at scales compatible with Landsat.
Kauth and Thomas (1976) indicate that soil
differences appear to be a matter of soil color
and that, after a crop canopy develops, soil
differences tend to be cancelled out. While it
is true that the soils themselves may be
masked, the canopy would be uniform only
if the crops on different soils were planted
on the same date and grew at the same rate.
But soils differ in drainage and other profile
and landscape characteristics with the result
that crops are planted on different dates on
different soil associations and once planted
grow at different rates. Thus, although the
soil itself eventually is masked by crop
canopies, the canopy varies in phenological
characteristics with different soils. The first
objective of this study was to try to identify
soil associations f)'om canopy reflectances.
The second objective was to use training
data selected by soil association to inventory
crops in a 12,950 hectare study area.
MATERIALS AKD METHODS
A study area of 40.2 km by 3.2 km (12,950
hal traversing east and west across northern
Brookings County (about 100 kill nOlth of
Sioux Falls) was selected because it crossed
major soil associations of eastern South
Dakota. The climate of the area is subhumid,
continental with warm summers and cold
winters. Mean annual precipitation is about
55 cm and the average annual temperature
6.6°C. The approximate latitude is 45°N and
the longitude 97°W. The native vegetation
was tall grass. The well drained upland soils
are classified as Chernozems using the 1949
USDA Classification System (Soil Survey
Staff, 1949) and as Mollisols using the Comprehensive System (Soil Survey Stan~ 1975).
The soil associations studied are listed and
classified in Table 1. Corn and oats are the
major crops presently grown, although
wheat, barley, Hax, alhllfa, and grass also are
important.
A complete crop inventory of the 12,950ha study area was made by field survey for
April 19, 1974, and updated on June 30,
1974, the dates of the two Landsat scenes
studied. Field boundaries, acreage, and crop
type were determined. This inventory was
made at a scale of 1:60,000, inked on plastic,
and overlayed on a soil association map
taken hom the Brookings County Soil Survey (Westin et al., 1958). The land-use data
are tabulated in Table 10 under the ground
truth column where they are used to compare with the Landsat estimations.
Growth charts (crop calendars) for the
major crops and grass of this area were prepared. The charts for corn, oats, alf~llfa, Kentucky bluegrass, and smooth bromegrass are
shown in Figure 1.
Landsat photo products of MSS-5 and
MSS-7 bands for April 19 (1635-16381) and
June 30 (1707-16361) were enlarged and
printed at scales of 1:500,000 and 1:60,000.
Color composites for both dates also were
prepared at the University Remote Sensing
Laboratory. The 1:60,000 imagery was
mosaicked and mounted on board equipped
with registration pins. The plastic overlays
of the field patterns and soil associations
were superimposed on these images.
Landsat computer compatible tapes (CCT)
for the April 19 and June 30 dates were obtained, and the digital data for the study area
were assigned a one-digit alphanumeric
code by digital count increments for print
out differentiation. The study area file line
printer output was mosaicked to provide a
visual format of the digital data.
From the overlays of the field boundaries,
vegetation categories, soil associations, and
Landsat data, test fields were selected that
represented vegetative categories on each
soil association. For April 19 three sites each
were selected on hlllow plowed, mixed pasture, altered stubble, dormant fJasture, and
water, and two sites were selected on growing vegetation (bluegrass). For June 30, 8
corn fields (493 data points), 12 oat fields
(503 data points), 7 grass fields (503 data
points), and 500 data points from two lakes
were selected. Information on the vegetation
test sites included crop height and conclition, weed cover, and soil surface condition.
Potential differentiation among categories
was estimated from the preliminary training
sites. Digital val ues of all four bands for each
pixel within the training areas were listed.
Group means and within-group variation
were determined for evaluation of preliminary classification categories.
Computer classification programs used
317
LA DSAT SPECTRAL SIG JATURES
TABLE 1.
Dominant Soil Series
of Association
SOIL ASSOCIATIONS OF THE STUDY AREA
1. Vienna
Loam glacial till
2. Kranzburg
Loess, about 90 cm
thick over loam
glacial till
Alluvium
3. Lamoure
4. Fordville
5. Poinsett
6. Buse
Taxonomic Classification of
Dominant Soil Series
Soil Parent Material
Udic haploboroll, fine
loamy, mixed
Vdic Haploboroll, fine
silty, mixed
Cumulic Haplaquoll, fine
silty, mixed (calcareous)
frigid
Alluvium about 75 cm Pachic Udic Haploboroll,
thick over sand and
fine loamy over sandy or
gravel
sandy skeletal, mixed
Silty glacial drift
Udic Haploboroll, fine
silty, mixed
Loam glacial till
Udorthentic Haploborolls,
fine-loamy, mixed
were K-Class Classifier (Serreyn and eIson, 1973) and a stepwise discriminate
analysis program hom a package ofbiomedical (BMD) computer programs (Dickson,
1974). K-Class is a minimum-distance-tothe-mean classifier that makes decisions
based upon the means and variances within
the test field data. It is considered more versatile than the BMD program because band
combinations can be used. The BMD program computes a set of linear classification
functions by choosing the independent variables in a step-wise manner. Data then are
classified according to distance from a
categorical mean and the probability of their
occurring in each category.
The mode seeking program is an unsupervised, iterative, clustering algorithm developed by Kaveriappa and Nelson (1973).
Results hom the analysis of Landsat data
from April 19 and June 30 are divided into
two subjects: first, to separate categories of
agricultural land use, and second, to assess
the influence of soil associations on the
spectral signatures of vegetation.
Landscape Position
Gently sloping upland
Gently sloping upland
Flat bottomland
Flat terrace
Irregular undulating
upland
Irregular hilly upland
the Lamoure (bottomland, Table 1) and Buse
(hilly upland) associations which generally
are in vigorously growing grass on this date
had darker tones on both MSS bands. The
color composite indicated there were at least
three distinct categories in the study area:
red hues where grass or small grains were
growing, green hues where soils were bare
or covered with residue, and black \-vhere
water was standi ng.
However, ground observations indicated
there were six possible categories of land
use that might be distinguished: clean plowing, mixed pasture, growing vegetation, altered stubble, dormant pasture, and water.
Three fields of each of these categories except growing vegetation (which had two)
were located on the CCT, and the separability of the six categories was attempted using
K-Class for 16 band combinations. The results from the per pixel training data are
shown in Table 2. The best I-band classification was by 7 (59 percent correct); the best
2-band combination was 4 and 6 (61 percent); and the best 3-band was 4, 5, and 6 (59
percent).
RESULTS AND DISCUSSION
LANDSAT DATA: APRIL 19.
J97~
(1636·16381)-SEPARATING CATEGORIES OF LAND
USE
The April data were analyzed to determine the kind ofland-use and vegetation information that could be gained at latitudes of
about 45° from Landsat data early in the
spring. Figure 1 indicates that grasses were
growing, oats had recently been planted, and
the land was being prepared for corn.
Photo interpretation of MSS-5 and MSS-7
prints at a scale of 1:500,000 indicated that
KENTUCKY
BLUEGRASS
SMOOTH
BROMEGRASS
Vertical scale represents period of maximum
growth.
FIG. 1. Growth characteristics for vegetation of
Brookings County.
318
PHOTOGRAMMETRIC ENGINEERING & RElvlOTE SENSING, 1978
TABLE
2.
ACCURACY OF SEPARATING SIX CATEGORIES* OF LAND USE FROM TRAIj\;ING DATA
WITH K-CLASS CLASSIFER USI:-iG 15 BAND COMBI:-iATlONS
APRIL 19, 1974 Landsat
Band
Combination
Overall
Accuracy
Banel
Combi nation
Overall
Accuracy
Banel
Combination
Overall
Accuracy
4
5
6
7
4,5
53
54
56
59
55
4,6
4,7
5,6
5,7
6,7
61
57
55
55
55
4,5,6
4,5,7
4,6,7
5,6,7
4,5,6,7
59
55
58
57
57
* C)('an plll\\ ing:. g:rowing:
\"cg:el.ttiOll. dormant \'egetatioll. mixed pasture. altered stuhble, waler
The training data then were analyzed by
the B~l D and K-Class program (Table 3, left
position). Neither program gave promising
results, and a graphic display ohhe data (Figure 2a) ind icated that the category called
mixed pasture O\'erlapped all categories except water. Altered stuhble also was a variahle category, so hoth of these were deleted
hom the training data and classification of
the !(Hlr remaining categories was attempted. ~lixed pasture, altered stubble, and
dormant pasture all are land-use classes containing a mixture ofland-use classes. Thus, it
is understandable that they would be confused with classes that consist of individual
cO\'er materials alone. Tahle 3, center portion, shows the classification of the four remaining categories. The accuracies increased to 70-80 percent. The highest accuracy was with .f bands using the Br--ID program. A graphic displaY' of these results (Figure 2b) shows that the category dormant
pasture was the main source of confusion.
Some of these fields had weed growth, and
other areas were bare.
A third attempt was made using the three
categories observed on the visual image: water, plowed soils, and growing vegetation.
TABLE
7
5,7
5,6,7
4,5,6,7
1
2
3
LANDSAT DATA-APRIL 19. 197~ (](;Jri-I63IlI)
Ij\;FLLT:"CE OF SOILS OJ\; SPECTRAL PROPERTIES
OF \'EGETATIO:"
The data for a single cover type were isolated, and an attempt was made to identify
the soil association hom these data. The data
!i'om 500 pixels occurring in grass were separated for three soil associations. The BM D
program was used to develop training profiles for each category and to c:lassi fy the
data. The results are shown in Table 4. Band
5 (indicating red light absorption) and band
7 (related to amount of \'egetative biomass)
in combination produced the highest level of
aecurac~'. The Buse soil association (thin
soils on steep uplands, Table 1) was most
easily separated, which seems reasonable
considering the steep soil landscape and
3.
CLASSIFICATION OF TRAINING DATA FOR SIX, FOUR, AND
THREE CATEGORIES OF LAND USE. LANDSAT DATA
APRIL 19, 1974
Lanel Use-Six Categories I
Banel
Combinations
The results shown in Table 3, right pOltion,
indicate an accuracy 01'91 percent. A graphic
display is shown in Figure 2c.
To summarize, the early spring Landsat
data were useful to separate cropland hom
grassland and to locate areas of open water.
In areas of variable acreages of cropland this
information could be of value for preliminary crop estimates.
Lanel Use-Four Categories 2
K-Class
Accuracy
BMD
Accuracy
%
%
Banel
Combinations
59
55
57
58
46
49
52
55
7
7,5
7,5,6
7,5,6,4
Land U se-Three
Categori es 3
K-Class
Accuracy
B ID
Accuracy
%
%
Banel
Combinations
74
74
74
76
73
80
81
82
7
5,7
5,6,7
4,5,6,7
Plowed soil, growing vegetation, mixed pasture, dorm.\llt vegetation, altered stubble, water.
Plowed soil, growing vegetation, dormant vegetation, wilter
Plowed soil, growing vegetation, water
B 10
Accuracy
%
88
89
91
91
319
LANDSAT SPECTHAL SIGNATUHES
Four Categories
'"..>
::;
'"<
:l'"
<
"'<
3.7
"':><
'"<
"'<
:>
4.1
:>
u
z
g
"
r·"
u
I
:;;
x
0.8
~.
;:
/\
;/ w
<
u
wl "z
\
..
.
<
<p
<
u
<
--
/
"u
'"
.j
gv
U>
.3··L
. 5.4
3.4
~~---~C7-
·0.9
FIRST CANONICAL VARIABLE
VEGETATION CLASSIFICATION
VEGETATION CLASSIFICATIO~
LANDSAT
DATA
19 APRIL
19 APRIL
1974
FIG. 2.
"
u
:;;
."'
1.3
,- W
I
Lu _
J
·4. \ L..,-3.-=.----;;0'"'.6,.-------c3; -O
.•
SECOND CANONICAL VARIABLE
VCGETATIOtl CLASSIFICATiON
LANDS!,T DATIl
DliTI\
Ie)
1974
cp = clea~ plOWing
gv = growln~ vegctution
dp - dormant pasture
IN
= water
cp : clean plowing
mpa = mixed pasture
gv = grow~ng vegetation
as = altered stubble
dp = dormant pasture
'oJ = water
z
z
<
3.6
SECOND CANONICAL VARIABLE
LANDSAT
4.3
..>
..>
..>
z
z
<
2c
Three Ca te<jor ies
2b
2a
six Categories
AI'IUL
1974
cp
elC.ln plowinq
gv = growing vcgetatloll
..... = W:1 ter
Graphic displa~· or six-, I<)tlr-, and thrce-categorv land-use separations hy B~ID program.
thin soil profile which would h'1\·e less \·ig- study are separated in the field using manmade criteria. These data indicate that specorous vegetation.
trally
separable soil associations which reA second experiment was to select training
data to compare bare soils for fi\'e soil as- flect plant response could be s\·nthesized
sociations. the data are given in Table 5. The and perhaps be useful for agriculture.
results indicate only f~lir success. The rather LA:-;DSAT DATA--:JU,E 311. 1D7~ (1/117·1(;:)61)
low recognition mav be because in early SEPARATI:\C CATEGORIES OF LA:-;D L·SE A:-;D
spring the soils were all moist and had not PREDICTI:\G ACREAGES OF CROPS
had time to reach their natural drainage
According to \·egetation growth patterns
states. An overall accuracv of 43 percent
was achieved, The Poinsett association and shown in Figure 1, June 30 should provide
the Fordville associations were identified maximum differences in the three major
more easilv than the other soil associations. classes of vegetation: corn, small grain, and
Both ofthe"se associations are better drained, forages (grasses and alfalb). A comparison of
the former because of a sloping position, the MSS-5 and ~,ISS-7 and of the color composlatter because of a gravel substratum, Soil ites with soil association boundaries inclicates there are some tonal differences within
associations including those used in this
TABLE 4.
IDENTIFYING SOIL ASSOCIATIONS FROM GRASS DATA L.A:-JDSAT,
APRIL 19, 1974, FOR THREE SOIL ASSOCIATIONS
Soil Association (Described in Table 1)
Band
Combinations
Fordville
54*
Poinsett
147*
Buse
299*
Overall
500*
Percent Corred Classification
MSS-7
MSS-7-5
MSS-7-5-4
MSS-7-5-4-6
•
~tllllber
of data points
30
52
56
55
45
71
66
70
66
85
83
85
55
75
73
74
320
1978
PHOTOGRAMMETRIC ENGINEERI G & REMOTE SENSING,
TABLE 5. IDENTIFYING SOIL ASSOCIATIONS FROM BARE SOIL DATA LANDSAT,
APRIL 19, 1974, FOR FIVE SOIL ASSOCIATIONS
Soil Association (Described in Table 1)
Band
Combinations
Fordville
273*
Poinsett
322*
Lamoure
371*
Kranzburg
405*
Vienna
368*
Overall
1739*
15
33
38
25
39
43
Percent Correct Classification
MSS-4
MSS-4-7
MSS-4-7-5
75
63
55
59
30
28
1
1
26
35
46
50
• Number of data points.
the same land use due to soil association influence. This was analyzed in the next part
of this study.
The spectral distribution of the proposed
training sites was analyzed to determine the
extent of spectral overlap in the classification categories. Data from all four MSS
bands are shOvvn in Table 6. Note that band 7
is on a 64 scale basis while the other three
bands are on a 128 scale basis.
On MSS-4 there was considerable overlap
among the corn, small grain, and grass data,
but water did not overlap with the other
categories. The wide range in MSS-5 values
for grass data indicates that levels of red light
absorption during photosynthesis varies a
great deal within the grass data, or there is
variability in the amount of bare soil showing since with red light vegetation and soil
are about equally reflective. The fields that
are in poor condition due to grazing differences and possibly drought stress would be
responsible for the higher reflectance values
TABLE 6. REFLECTED BRIGHT ESS OF FOUR
CATEGORIES OF LAND USE ON LANDSAT
CCT FOR JUNE 30, 1974
Land Use
Water
(500)*
Corn
(503)*
Small Grain
(503)*
Grass
(493)*
Bands
Reflectance
MSS-4
-5
-6
-7
MSS-4
-5
-6
-7
MSS-4
-5
-6
-7
MSS-4
-5
-6
-7
17-24
10-21
6-21
0-8
25-31
16-27
22-41
9-20
23-32
13-26
46-72
23-43
23-41
14-44
38-62
19-38
'" Number of training pixels.
due to exposed soil, while those actively
growing would have low reflectance due to
red light absorption by green vegetation. Of
the vegetative categories small grain had the
lowest red reflectance, so would appear in
dark tones on positive prints. The corn data
clearly had lower reflectances than the small
grain and grass data on MSS-6 and MSS-7.
However, there was some overlap between
the grass and small grain data. The small
grain data generally had higher reflectances
than the grass data but, due to the withingroup diversity, many exceptions occulTed.
Relative levels of reflectance among the
wavebands for corn are shown in Figure 3.
On this date corn is actively carrying on
photosynthesis (so has lower reflectance),
and this is indicated by the fact that MSS-5
output is less than MSS-4. However, because the vegetative biomass is low the net
reflectance of the fields is still influenced by
the spectral properties of the soil. The low
reflectance in band 7 is the result of band 7
being on a 64 scale basis while the other
three bands are on a 128 scale basis.
The spectral profile of small grain shown
in Figure 4 is characteristic of growing vegetation having a large biomass. This profile
493
35
samples
."
""
Band
~
;;
:;;;20
."
.
Band
1\
"0
I
I
I
z
'"
~
"'"
\
7
I
0
0
\
\
Band
'- -\
6
i/\ \ /\
\
-(
\
\(/-
27
LANDSAT SPECTRAL
43
REFLECTANCE
FIG. 3. Reflectance characteristics of corn,
Brookings County, 30 June 1974.
321
LANDSAT SPECTRAL SIGNATURES
503
35
Band
4
~
1\,
z
~
Band
6
\1
:\J \
"
~
o
Band
I
~
z
w
\,r/
~
:;;
: .:\
\
samples
6
SPECTRAL
30
.i
I
.I
I·
. I
I
!
Ii
__'\_
;'i
I
66
:;
\
36
LANDSAT
Band
:
~
\
---""-)_~'-=-~~-~\='_'_._/
o Lo
~
7
//,
u
500
7
II
II
< 20
Band
Ii
5
Band
\1
.rl
55
samples
~
~
REFLECTANCE
I
FIG. 4. Reflectance characterisitcs of small
grains, Brookings County, 30 June 1974.
Band
4
r',
I
I
\
I
\
I
\
\
I
'. /
I
\,
I
/
I
\
\
\
'
\
o LO---------:-,------:-.:..-----''-c-:30
consists of low retledance in MSS-5 and
high reflectance in inhared bands.
The spectral prohle of grass is shown in
Figure 5. The data with the lower reflectance values represent most active growth.
As retlectance values for MSS-5 increase, the
levels of photosynthesis decrease. The converse is true for infrared values during this
period. The higher the inhared reflectance
values, the greater is the vegetative biomass.
The water training data (figure 6) showed
decreased reflectance with increased
wavelength.
The spectral distribution of the training
data indicated a limited categorical overlap,
so the BMD program was used to determine
the separability of the data. Use of MSS-7
alone produced on accuracy of 90 percent
(Table 7). The major area of confusion occurred where 10.7 percent of the small grain
was misclassified as grass and where 15.5
percent of the grass was misclassified as
small grain. Although small grain is a grass, it
usually is fertilized while grass in this area is
not fertilized. When MSS-5 was added to the
program (Table 8) the small grain misclassified dropped hom 10.7 percent to 2.9 per-
Ba nd
15
cent. The addition of MSS-5 also dropped
the grass misclassifled from 15.5 percent to
10.1 percent.
Figure 7 is a graphic display lIsing the
BM D program of the 4 category classi fication. A few water pixels were misclassified.
These most likely came hom the shorelines
of morainic depressions.
TABLE 7. CLASSIFICATION OF TRAINING DATA
USING MSS-7 BMD PROGRAM, LANDSAT,
JUNE 30, 1974
Corn
Corn
Small
Grain
Grass
Water
go. I
pcrc~nt
~
;"'i
o
J
\
~--_-----=
58
LANDSAT
Water
0.8
5.2
0
4.2
0.4
1.6
85.1
15.5
0
10.7
84.1
0.8
0
0
97.6
overall accuracy.
Corn
OL-_---=---
Grass
6
o
~
Small
Grain
93.9
.'------"
"
Q.
REFLECTANCE
TABLE 8. CLASSIFICATION OF TRAINING DATA BY
THE BMD PROGRAM, LANDSAT, MSS-5 AND 7,
JUNE 30, 1974
(1999 Pixels)
503 samples
~
LANDSAT SPECTRAL
FIG. 6. Reflectance characteristics of water,
Brookings County, 30 June 1974.
SPECTRAL
REFLECTANCE
FIG. 5. Reflectance characteristics of grass,
Brookings County, 30 June 1974.
Corn
Small
Grain
Grass
Water
Small
Grain
Grass
Water
95.3
2.0
2.6
0
4.8
0.6
1.6
92.3
10.1
0
2.9
89.3
0.4
0
0
98.0
1:-.l3.7 pen:ent o\'endl accuracy
PHOTOGRM.l~1ETRIC ENGINEERING & REMOTE SENSING,
322
S.D
.. "'-....
TABLE 9. CLASSIFICATION OF TRAINING DATA BY
K-CLASS CLASSIFIER, LANDSAT, MSS-5 AND 7,
.I UNE 30, 1974
(1999 data points)
...........
.............
\
\.A..~._~~\:,
001
I
/'
\
\
i
\
I;'
/\
II
~
.'w
/
-ll
Corn
Small
Grain
Grass
Water
95.7
2.0
2.0
0.2
4.9
0.8
4.4
88.5
8.3
0.2
6.6
90.0
0.4
0
0
95.5
CO,"
Corn
Small
Grain
Grass
Water
/
I
1978
/
92.5 pen:cnl O\'erall a<.:<.:llr.lcy
t"IR5T CANONICAL VARIABLE
FIG. 7.
Count~·,
Vegetation classification, Brookings
30 .Iune 1974.
K-Class, then, was used to classify the
training data (Table 9). The results differed
\'er\, little from the BMD results.
si ng these trai n i ng data, the K-Class
program classified all the data points in the
128.6-s(juare km (12,950 hal study area into
one of the felltr categories: corn, small grain,
forage (grass), or water. The results are compared with the ground truth measurements
in Tahle 10. The results for corn, grass, and
water are good, and small grain estimated
were somewhat less accurate. K-Class estimated slightly less grass and corn and
slightly more small grain than was measured
by ground truth.
TABLE 10. FUGHTUNE CLASSIFICATION BY
K-CLASS CLASSIFIER, LANDSAT DATA,
.IUNE 30, 1974, COMPARED WITH
MEASURED GROUND TRUTH
Vegetation
o
LA:\DSAT DATA-JU:\E :30 INFLUENCE OF SOIL
ASSOCIATION ON SPECTRAL PROPERTIES OF
VEGETATION
The same approach taken with the April
19 Landsat data was taken with the June 30
data. That is, an attempt was made to identify soil associations within the data of a
single cover type. This is a test of the homogeniety of soil associations.
An attempt at identifying soil associations
from Small Grain training data using the
BMD program is shown in Table 11. Using
TABLE 11.
Category
Corn
Small grain
Grass
Water
Hectares
K-Class
Ground Truth % Error
4,050
4,274
4,496
130
4,107
3,666
4,660
134
1.38
16.58
3.51
2.98
93.9 percent overall a<.:<.:lIracr
Band 4 alone, two soil associationsLamoure and Vienna-could be identified
about thee-quarters of the time but the other
three soil associations were not identified
consistently. Addition of other bands did not
alter these results markedly. The Lamoure
soils probably stood out because they were
wetter in the spring and planted later.
The mode-seeking program (Kaveriappa
and Nelson, 1973) was used to determine if
different soils within the corn data had
specific modes (Table 12). The data show
that mode 1 contained 95 percent of the
Poinsett soil data, 62 percent of the
Kranzburg soil data, and 53 percent of the
Lamoure soil data. The same problem occur-
IDENTIFYING SOIL ASSOCIATION FROM SMALL GRAIN TRAINING DATA, LANDSAT,
.I UNE 30, 1974, FOR FIVE SOIL ASSOCIATIONS
Percent Correct Classification
Soil Association (Described in Table 1)
MSS-4
MSS-4-6
MSS-4-6-5
MSS-4-6-5-7
Fordville
Poinsett
Lamoure
Kranzburg
Vienna
Overall
0
22
20
20
22
4
22
26
88
82
80
82
0
18
22
18
70
72
70
68
28
30
35
32
323
LANDSAT SPECTRAL SIG! ATURES
TABLE 12. MODE SEEKING OF CORN TRAINI:-IG
DATA USING BANDS 4, 5, 6, AND 7, LANDSAT,
]U:-JE 30,1974 (NUMBERS ARE PERCENT OF
SOIL ASSOCIATION 11' EACH MODE)
Soil Association (Described in Table 1)
Mode Poinsett Lamoure Kranzburg Vienna
Numher
60*
45*
246*
142*
1
2
3
4
5
6
95
3
1
53
44
2
62
18
17
2
1
7
37
44
8
1
1
* -'ulIlber of data points
red with modes 2 and 3. This indicates that,
even though the data from a single soil group
mm' occur dominantlv in one mode, confusio;1 among the differ~nt soil associations occurred because the data from many soil
groups may occur in a single mode.
An attempt to recognize the soil association within the corn data was made by using
the BI\1D program. The results are shown in
Table 13. M 55-6 was the first band entered
in the classification process. Kranzburg soils
were identitled conectly 78 percent of the
time by using this band alone. The addition
of M55-5 into the classification process increased the accuracv of the Lamotne and
Vienna associations, but decreased the Poinsett and Kranzburg accuracies. The best
overall results were obtained when all four
bands were used, but e\'en then correct classification results varied boom 57 to 73 percent
(Table 14).
K-Class, which can compare band combinations, then was used to further analyze the
effect of different MSS combinations for the
recognition of soil associations within the
corn data. The results are shown in Table 15
for four soil associations. The highest level
TABLE 13.
of accuracy in each group is underlined. The
results ar~ \·,triable. For example, use of
1\1S5-7 resulted in identifying the Vienna
soil associations. Overall the best results are
achieved using all 4 bands.
Accurac~' of separating soils fi'()Jll different
types of vegetation with the Bi\ID program
are shown in Table 16. The excellent separation of soils within the grass data is attributed to the wide difference in soil color and
moisture existing between Poinsett
(lIndlilating-silt~· uplands, Table 1) and
Fordville (flat well drained terraces). On
June 30, small grain was growing vigorously,
thus m,tsking the soil and thus presenting a
htirh' uniform appearance. Corn on June 30
was 'about knee high, so much soil still was
exposed.
A comparison of h-Class and the BI\I D
programs for separating soils using corn data
is shown in Table 17. Large differences are
apparent in the two programs when 1\155-6
was used alone. Poinsett soils were recognized at 6,5 percent accuracy by the BI\ID
program while K-Class could not recognize
them at all. The cOIl\'erse is true for Vienna
soils where K-Class identified this soil association at 86 percent accuracy ane! the
BMD program using the same data recognized this soil association with only 11 percent accuracy. The differences between
these programs diminished as more
wavebands were included in the classification program, thus indicating the importance
of using all four bands.
SCM MARY AI\D CONCLUSIOI\S
Analysis of Landsat data from April 19
showed that data collected early in the
spring in cool temperate latitudes can be
used to separate cropland, grassland, and
open water. Soil differences on this date had
a more pronounced influence on the spectral
properties of grassland than on cropland.
IDENTIFYING SOIL ASSOCIATION FROM CORN TRAINING DATA, LANDSAT,
JUNE 30, 1974, FOR FOUR SOIL ASSOCIATIONS
Percent Correct Classification
Soil Association (Described in Table 1)
MSS-6
MSS-6-5
MSS-6-5-7
MSS-6-5-7-4
• Number of data points
Poinsett
60*
Lamoure
45*
Kranzburg
246*
Vienna
142*
Overall
493*
65
52
60
55
38
71
78
73
80
78
75
12
56
61
62
50
65
65
65
72
324
1978
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING,
TABLE 14. IDENTIFYING SOIL ASSOCIATION FROM
CORN TRAINING DATA BY BMD PROGRAM USING
BANDS 6-5-7-4, LANDSAT, JUNE 30, 1974,493
DATA POINTS
(PERCENT OF CASES CLASSIFIED)
Poinsett LamOtHe Kranzburg Vienna
Poinsett
LamotHe
Kranzburg
Vienna
3
73
62
4
3
18
15
18
TABLE 15.
32
16
3
7
72
9
7
57
Analysis of Landsat data from June 30 indicated that several bands enabled the separation of vegetative categories not separable by a single band. A generalized training
set which contained data points from each of
six soil associations was used to separate four
categories of agricultural land use in a 12,950
hectare test area at an accuracy exceeding 90
percent.
Training data were separated by vegetative category, and soil differences within a
IDENTIFYING SOIL ASSOCIATION FROM CORN TRAINING DATA BY K-CLASS
USING 11 BAND COMBINATIONS, LANDSAT, JUNE 30, 1984,
PERCENT CORRECT CLASSIFICATION
Soil Association (Described in Table 1)
Poinsett
60*
LamOtHe
45*
Kranzburg
246*
Vienna
142*
Overall
493*
0
0
0
48
42
47
35
0
12
40
60
0
0
0
27
44
40
0
0
0
40
60
87
56
81
79
77
86
82
86
53
68
69
68
70
70
64
68
68
70
MSS-5
MSS-6
MSS-7
1SS-4 & 6
MSS-5 & 6
MSS-5 & 7
MSS-5-7
MSS 4/6
MSS 5/7
MSS-5,6,7
MSS-4,5,6,7
94
68
75
75
58
77
78
76
66
91
87
79
79
53
71
'" Number of data points
TABLE 16.
SEPARATION OF SOIL ASSOCIATION BY VEGETATION CATEGORY
USING BMD PROGRAM, LANDSAT, JUNE 30, 1974
Percent Correct Classification
MSS
Corn
Small grains
Grass
Soil Association (Described in Table 1)
Bands
Poinsett
LamOtHe
Kranzburg
Vienna
5,6,7
4,5,6
4
55
21
78
79
81
71
55
66
19
Fordville
Overall
*
65
21
60
70
33
'" No Test Data
TABLE 17.
COMPARISON OF FOUR BAND COMBINATIONS AND K-CLASS AND
BMD PROGRAMS TO SEPARATE SOIL ASSOCIATIONS
Percent Correct Classification
Soil Association (Described in Table 1)
Poinsett
60
Lamoure
45
Kranzburg
246
Vienna
142
Overall
493
Bands
K-Class
BMD
K-Class
BMD
K-Class
BMD
K-Class
BMD
K-Class
BMD
MSS-6
MSS-6,5
MSS-5,6,7
MSS-4,5,6,7
0
42
40
60
65
53
55
62
0
44
40
53
36
69
78
73
87
79
79
79
73
73
86
75
76
66
11
55
55
57
68
70
70
71
50
65
65
67
71
72
LANDSAT SPECTRAL SIGNATURES
single cover type were investigated. The
most pronounced soil differences were
found among corn data where some soil associations could be identified over 70 percent of the time.
Results of this study indicate that soil associations could not consistently be recognized with high accuracy within the data ofa
single vegetative type. Probably some of this
is due to tuming practices-different crop
varieties, dates of planting, rates of fertilization, and the like. However, the data
showed that soils did influence all vegetative spectral reflectances to some degree.
Thus, it would seem that training sets would
be more representative of a study area if they
contained data points from all soil associations present.
ACKNOWLEDGMENT
The research reported on herein was suppOlted in part by ASA Grant NGL 42-003007 to Remote Sensing Institute, South
Dakota State University, Brookings, South
Dakota 57007.
REFERE 'CES
Dickson, W. ]. 1974. BMD, Biomedical Computer
Programs. University of Califomia Press, Berkeley, California.
Hoffer, F. lvI., R. A. Holmes, and R. ]. Shay. 1966.
Vegetative, soil, and photographic factors af~
fecting tone in agricultural remote multispectral sensing. Presented at Fourth Symp. on
Remote Sensing of the Environment. April
12-14, 1966.
Kauth, R. ]., and G. S. Thomas. 1976. The Tasselled Cap-A Graphic Description of the
Spectral-Temporal Development of Agricul-
325
ture Crops as Seen by Landsat. Proc. Symp.
on Machine Processing of Remotely Sensed
Data. Purdue University. lEE Cat. 76 CH
1103-1 MPRSD.
Kaveriappa, G. K., and G. D. Nelson. 1973. Unsupervised Iterative Clustering in Pattern
Recognition. Remote Sensing Institute 73-11,
SDSU, Brookings, S. D.
MacDonald, R. B. 1976. The Large Area Crop Inventory Experiment. Second Annual WM. T.
Pecora Memorial Symposium Proceedings,
Sioux Falls, S. D. Oct. 25-29.
Myers, V. I., F. C. Westin, M. L. HOlton, and]. K.
Lewis. 1974. Soil influences in crop identification.In Effective use ofERTS multispectral
data in Northern Great Plains. Remote Sensing Institute, South Dakota State University.
Serreyn, D. V., and G. D. Nelson. 1973. The
K-Class classifier. Remote Sensing Institute,
South Dakota State University. RSI. 73-08.
Simonson, R. W. 1971. soil association maps and
proposed nomenclature. Soil Sci. Soc. Amer.
Proc. 35:959-963.
Soil Survey Staff. 1949. Soil Science, Volume 67,
No.2.
_ _ _ _ . 1975. Soil Taxonomy. Agri. Handbook
No. 436. U. S. Gov't P.O. Washington, D.C.
Westin, F. C., G. ]. Buntley, F. E. Schubeck, L. F.
Puhr, and N. E. Bergstresser. 1958. Soil Survey of Brookings County, South Dakota Bulletin 468. SDSU Agricultural Experiment Station, Brookings, S.D.
Westin, Frederick C., and C. ]. Frazee. 1976.
Landsat Data, Its Use in a Soil Survey Program. Soil Science Society of American Journal 40: 1, 81-89.
Wigton, W. H., and D. H. Von Steen. 1973. Crop
identification and acreage measurement
utilizing ERTS imagery. Presented at Third
Earth Resources Technology Satellite-l
Symp., December 10-14, 1973.
Forthcoming Articles
Alan Austin and Robert Adams, Aerial Color and Color Infrared Survey of Marine Plant
Resources.
Henry W. Brandli, The Night Eye in the Sky.
Hong-Yee Chiu and William Collins, A Spectroradiometer for Airborne Remote Sensing.
H. W. Gausman, D. E. Escobar, j. H. Everitt, A. j. Richardson, and R. R. Rodriguez, Distinguishing Succulent Plants from Crop and Woody Plants.
H. W. Gausman, D. E. Escobar, R. R. Rodriguez, C. E. Thomas, and R. L. Bowen, Ozone
Damage Detection in Cantaloupe Plants.
Dr. Robert W. johnson, Mapping of Chlorophyll a Distributions in Coastal Zones.
Karl Kraus, Rectification of Multispectral Scanner Imagery.
W. Marckwardt, The Accuracy of Orthophotos and Simultaneously Collected Terrain
Height Data.
G. Otepka, Practical Experience in the Rectification of MSS Images.
Dr. Charles K. Paul, Internationalization of Remote Sensing Technology.
I. L. Thomas, A. j. Lewis, and N. P. Ching, Snowfield Assessment from Landsat.