United States
               Environmental Protection
               Agency
 Environmental Monitoring
 Systems Laboratory
 Las Vegas, NV 89193-3478
               Research and Development
 EPA/600/S4-91/027 Jan 1992
EPA       Project Summary
               Watershed  Characterization
               Using  Landsat Thematic  Mapper
               (TM)  Satellite  Imagery:  Blackfoot
               River,  Montana
               K. H. Lee
                This report describes a portion of a
               large regional project undertaken  by
               the Environmental Protection Agency
               (EPA) and water-quality authorities in
               the  States of Montana, Idaho, and
               Washington to identify and analyze fac-
               tors which are affecting water quality
               in an Interconnecting hydrologic sys-
               tem. To achieve this directive in Mon-
               tana, a spatial database is being con-
               structed which will contain satellite de-
               rived land cover, photo-interpreted mac-
               rophyte locations, climate data, topog-
               raphy,  hydrography, and soils.  The
               database will be used by EPA Environ-
               mental Monitoring Systems Laboratory
               (EMSL) to demonstrate the utility of a
               watershed scale  information manage-
               ment system.  This information man-
               agement system is  geared toward
               nonpoint pollution modeling and will
               evolve  into a decision support mecha-
               nism capable of assessing the suitabil-
               ity and feasibility of various alternate
               management scenarios. The data lay-
               ers focus on  elements required  for
               nonpoint source pollution modeling in
               which derivation of factors for soil erod-
               ibility, rainfall, topographic slope-length,
               and vegetation management are gener-
               ated for the watershed. A Geographic
               Information System (GIS) in which  to
               model  alternative land management
               scenarios such as road building, log-
               ging, and fire/burn management. The
               vegetation management factor will be
               partially based on land cover derived
               from Landsat Thematic Mapper  satel-
               lite  imagery.  Vegetative management
               factors combine vegetative cover and
 soil surface conditions Into one nu-
 merical factor.   This report will ad-
 dress only the generation of land cover
 maps for the Blackfoot River Water-
 shed through quantitative remote sens-
 ing techniques.
  This Project Summary was developed
 by EPA's Environmental  Monitoring
 Systems Laboratory, Las Vegas, NV, to
 announce key findings of the research
 project that Is fully documented in a
 separate report of the same title.  (See
 Project Report ordering Information at
 back).

 Introduction
  This report covers  the western third of
 the Blackfoot River Watershed, which is
 part of the greater Lake Pend Oreille Wa-
 tershed. This study is a joint cooperative
 effort between the U.S. EPA and the States
 of Montana,  Idaho,  and Washington to
 understand the water quality issues of the
 Lake Pend Oreille watershed. This part of
 the project deals with several subwater-
 sheds in the western third of the Blackfoot
 River Watershed, and is based in  large
 part upon the interpretation of satellite im-
 agery.
  The entire  Blackfoot River Watershed
 contains approximately 6,022 square kilo-
 meters (2325 sq. mi.). Because of budget
 restrictions, only the western third of the
 area,  2,185 square kilometers (844 sq.
 mi.), was classified for land cover.
  Figure 1 shows the area covered by
this part of the project.
  The Blackfoot  River flows 132 miles
from its  source near  the Continental Di-
vide westward to its junction with the Clark
                                                              Printed on Recycled Paper

-------
                  Pend Oreille River
                                                                                 Study Location

                                                                                    Canada
                                                                                              Montana
                                                                      Butte, MT
Figure 1. Remote Sensing/GIS Support for Section 525 of the Water Quality Act of 1987.
Fork at Bonner, Montana. Factors which
affect the water quality  are  the  logging
and  mining activities, road building, fires,
agriculture, unstable river banks,  munici-
pal sewage effluent, and pulp mill waste.
In Idaho, Lake Pend Oreille  receives all
the Clark Fork discharge; eutrophication
has been associated with the degraded
water quality of the  Clark Fork.

Procedures

Digital Image Processing
  A Landsat-5 Thematic Mapper (TM) digi-
tal image acquired  on 20 July, 1988 was
analyzed to create  a general land-cover
map for  the Blackfoot River Watershed.
The Landsat-5 TM satellite sensors record
the amount of electromagnetic energy in
seven spectral bands: blue,  green, and
red (visible spectrum); three  reflected in-
frared (IR)  bands;  and  one thermal IR
band.  Each pixel  (or cell) covers about
30 meters by 30 meters, except for the
thermal (IR) band which is 120 meters.
  The steps taken  in  the analysis of
Blackfoot River satellite data for land-cover
characterization were: image rectification;
optimum band selection;  unsupervised
classification; and accuracy assessment
of the final classified image.
  ...The Optimum Index Factor  (OIF) indi-
cated that a four-band combination of TM
bands 3 (red), 4, 5,  and 7 reflected (IR)
be used to minimize redundancy in the 7-
band data set.
  Ground Control points (GCPs) taken
from topographic maps were used to re-
align the satellite data, correcting  for dis-
tortions in the image. The Universal Trans-
verse Mercator (UTM) coordinate  system
was used for map control.
  The TM  data  set was rectified using
249 ground control points  which  yielded
an overall average locational error of 23.4
meters.

Classification System
  Classification units  can be selected  in
two distinct ways: (1) by image interpreta-
tion based on previous general knowledge
and experience and specific knowledge of
the area  by the  interpreter, or by  (2)
unsupervised "clustering" of similar pixels
by the computer. Spectral characteristics
of surface materials, such as soil or veg-
etation, are likely to vary from  place to
place. Therefore these  clusters  must'be
checked to consolidate some of them hav-
ing only minor differences in spectral char-
acteristics, but which are part of the same
soil or vegetation unit. This was done,
and  resulted in a total  of  seven major
categories, with the three subcategories,
shown below.
   Digital classification of multispectral im-
agery generally produces a sait-and-pep-
per appearing map. This can arise from a
nonuniform response of the  sensor,  pro-
cessing errors, the pixel size or  mapping
unit,  or  the classification  algorithm. In
smoothing,  every pixel is  spatially ana-
lyzed in conjunction with the pixels  that
surround it. Smoothing is used when  the
final land-cover classification is to have a

-------
Tablo 1. Error Matrix: Blackfoot River Watershed Subset
                                     Reference
              Forest-d    Forest-s    Ag     Range    Barren    Wetland    Water
C Forest-s
L Forest-s
A Ag
S Range
S Barren
E Wetland
D Water
15102
71
0
21
4
0
0
1020
2001
26
2337
41
0
85
157
43
9141
703
4
0
1
263
646
0
1638
191
0
575
38
128
62
159
26
0
0
313
69
63
329
1
121
11
0
0
0
0
0
0
2358
                            Overall Accuracy = 80.5%
           Commission
         User's Accuracy
         Forest-d
         Forest-s
         Ag
         Range
         Barren
         Wetland
         Water
=  89%
=  68%
=  98%
=  32%
=  10%
"  100%
"  78%
minimum mapping unit such as an acre or
hectare.
  The problems of converting raster data
(having a stair-case appearance) and vec-
tor data (shown as a smooth line, arcuate
or straight)  are presently solved  by an
intermediate file structure called SVF. The
problems of this conversion are expected
to be solved  in the near future  by the
utilization of raster data in a grid format in
the vector GIS.
  The following classification of land-cover
features resulted from unsupervised clas-
sification combined with  adequate  refer-
ence information from the area.

  1.  Agricultural lands: cropland and pas-
     ture, found in areas of low relief.
  2.  Barren land: bare, exposed rock.
  3.  Barren land (roads): including rail-
     ways.
  4.  Forest land: having a  tree  crown
     areal density of at least 10 percent.
   A) Dense  coniferous forest  cover
      (greater than 40%).
   B) Dense deciduous forest cover (more
      than 40%).
   C) Old  timber/thinned sites/sparse
      cover (cover less than 40%; more
      than ,10%).
  5.  Rangeland: grasses and shrubs.
  6.  Wetland: was difficult to distinguish
     between  forest and agriculture. Only
     several small areas were identified.
  7.  Water: includes the Blackfoot River
     and tributaries, lakes, and reservoirs.
                                    Omission
                                Producer's Accuracy
Forest-d
Forest-s
Ag
Range
Barren
Wetland
Water
= 99%
= 36%
= 91%
= 50%
= 6%
= 13%
= 100%
                        Accuracy Assessment
                           Evaluation of the accuracy of the land-
                        cover  classification requires comparison
                        between the classification results and ref-
                        erence data for the area.  These  refer-
                        ence data are taken from various sources.
                        Site-specific comparisons  are  made by
                        calculating  the frequency of coincident
                        classes, point by point, and reporting these
                        values in an error matrix (confusion matrix
                        or contingency table).
                           Detailed statements of accuracy are de-
                        rived from the error matrix in the form of
                        overall and individual land-cover category
                        accuracies.  For  each class the  percent
                        commission and percent omission errors
                        are calculated from the  error matrix (con-
                        fusion  matrix or contingency table).
                           The reference  data used to evaluate
                        the final, classified image of the Blackfoot
                        River Watershed subset were taken from
                        photo-interpreted  color  airphotos (1988)
                        flown for a macrophyte survey, NHAP CIR
                        photography  (1984), and soil  survey
                        orthophoto maps (1982).
                           Research indicates that reference data
                        from at least 1 percent of any area should
                        be enough for a valid  evaluation of the
                        accuracy of the interpretation. In the case
                        of the Blackfoot River Watershed, the ref-
                        erence data used covered  1 percent of
                        the area.
                           Table 1  shows the error  matrix of the
                        study- area having an overall map  accu-
                        racy of 80.5%. However, this  does  not
                        consider the accuracies of individual cat-
egories. Producer's accuracy (omission er-
ror) and user's accuracy (commission er-
ror) are important in assessing the reliabil-
ity of a classified image.
  Commission errors are the percentages
indicating the  ratio of those pixels in any
category  which are correctly classified.
Omission  errors are given a percentage to
indicate the probability that  a  reference
sample will be correctly classified.
  Unsupervised classifications  were ob-
tained  for the  eastern two-thirds  of the
entire  Blackfoot  River Watershed.  How-
ever,  not enough  reference data were
available  (i.e.,  air-photos,  maps,  and
ground verification) for an accuracy as-
sessment to be performed for this area.

Results and  Discussion
  The  selection of a four-band data set
containing 61  megabytes instead of the
six-band  sub-scene of 91 megabytes for
the western third of the watershed  re-
sulted  in the elimination of 30 megabytes
f,rom computer computation for the recti-
fied area of 100 km x 100 km.
  It is  projected that the  overall accuracy
of 80.5 percent would not be  a limiting
factor in the calculation of vegetation man-
agement factors for the Blackfoot  River
Watershed. The resolution  of the TM im-
agery  (30 meters)  will probably  have
greater impact on nonpoint pollution-mod-
eling as  the  level of detail  required  for
land cover can become quite specific, even
to the  point of whether the ground cover
is grass or weeds.  The forest class was
divided into dense and thinned cover arbi-
trarily;  this estimated amount of cover
should be useful in the  development of
vegetation management factors.
  The  Blackfoot GIS  contains other the-
matic layers which are commonly  avail-
able (i.e., hydrography,  elevation,  soils,
etc.); these  combined with  recent land
cover estimates will provide the EPA Re-
gions and State agencies with the build-
ing blocks on which further analyses surely
will develop. The GIS database, as con-
structed for this modeling effort,  is an at-
tempt to work on a third-order watershed
where the level of detail required may not
need to be so specific. Updates are much
easier in the GIS environment as this tech-
nology facilitates the incorporation of higher
accuracy  or resolution  information after
baseline development. A tentative list of
management issues in the Blackfoot River
Watershed includes investigations  of wa-
ter quality for fisheries management, com-
parisons  of  silvicultural prescriptions for
sediment yield reduction, and environmen-
tal assessment of potential impacts of pro-
posed  mining activities.
                                                                         &U.S. GOVERNMENT PRINTING OFFICE: 1992 - 648-080/40126

-------
 K. H. Lee is with Lockheed Engineering and Sciences Company, Las Vegas,
     NV89119
 Ross S. Lunetta is the EPA Project Officer (see below).
 The complete report, entitled "Watershed Characterization Using Landsat
     Thematic Mapper (TM) Satellite Imagery: Blackfoot River, Montana,"
     (Order No. PB92-115237/AS; Cost: $17.00, subject to change) will be
     available only from:
         National Technical Information Service
         5285 Port Royal Road
         Springfield,  VA 22161
         Telephone: 703-487-4650
 The EPA Project Officer can be contacted at:
         Environmental Monitoring Systems Laboratory
         U.S. Environmental Protection Agency
         Las Vegas,  NV 89193-3478

 United States
 Environmental Protection
 Agency
Center for Environmental
Research Information
Cincinnati, OH 45268
      BULK RATE
POSTAGE & FEES PAID
         EPA
   PERMIT No. G-35

Official Business
Penalty for Private Use $300
EPA/600/S4-91/027

-------