Mapping Land Cover Types in Amazon Basin Using l km
JERS- 1 Mosaic
Sassan S. Saatchi t, Bruce Nelson z, Erika Podest I, John Holt _
1. Jet Propulsion Laboratory, 4800 Oak Grove Drive, Pasadena, California 91109
2. Instituto Nacional de Pesquisas de Amazonia, CP 478, Ecologia, 69011-970, Manaus,
Amazonia, Brazil
Abstract
In this paper, the 100 meter JERS-I Amazon
mosaic image was used in a new classifier to
generate a 1 km resolution land cover map. The
inputs to the classifier were lkm resolution mean
backscatter and seven first order texture measures
derived from the 100 m data by using a 10 x 10
independent sampling window. The classification
approach included two interdependent stages: 1) a
supervised maximum a posteriori Baysian
approach to classify the mean backscatter image
into 5 general land cover catagories of forest,
savanna, inundated, white sand, and
anthropogenic vegetation classes, and 2) a texture
measure decision rule approach to further
discriminate subcatagory classes based on
taxonimc information and biomass levels.
Fourteen classes were successfully separated at
lkm scale. The results were verified by examining
the accuracy of the approach by comparison with
the IBGE and the AVHRR 1 km resolution land
cover maps.
Introduction
Understanding the human or climate induced
changes of a tropical landscape requires
knowledge of the current status of the ecosystem,
the extent of the land cover types susceptible to
change, and the causes and impacts of such
changes. Recent advances in remote sensing
technologies have partially contributed in
documenting and monitoring these land use
changes. There are, however, several unresolved
problems associated with mapping land cover
types and monitoring the tropics on regional and
continental scales. These derive from limitations
of current remote sensing techniques and the
methodologies used in both defining the land
cover types and identifying the parameters to be
monitored.
Optical remote sensing has been successfully used
for classification of land cover types and the study
of their changes on local to regional scales. High
resolution (30 m) Landsat Thematic Mapper (TM)
data has been the primary source for estimating
the rate of deforestation by INPE (Instituto
Nacional de Pesquisas Espaciais) and the Landsat
Pathfinder Program (Skole and Tucker, 1993;
Justice and Townshend, 1994). These studies
have used visual interpretation and classification
as their primary approach for extracting thematic
information. Most large scale maps derived from-
these show few land cover types. This is due in
part to difficulties in interpreting the spectral
information of Landsat data acquired at different
years or seasons.
During the past decade, several radar sensors have
been deployed in space such as the shuttle
imaging radar (SIR-A, SIR-B, and SIR-C/X-
SAR), ERS-1,2, JERS-1, and Radarsat. Except
the SIR-C/X-SAR system, all radar sensors have
only one channel. Though none were designed
specifically for land cover mapping, several
investigations have demonstrated that the data
provide unique information about the
characteristics of tropical landscapes (Sader,
1987; Foody and Curran, 1994). First, the radar
data can be acquired as frequently as possible due
to insensitivity to atmospheric condition and sun
angle. This allows continental scale high
resolution data for systematic assessment of
deforestation and regrowth processes. Second,
depending on the wavelength, the radar
backscatter signal carries information about forest
structure and moisture condition by penetrating
into the forest canopy. Few studies have
addressed these characteristics by using the radar
data for mapping tropical land cover and
estimating the biomass of regenerating forests
(Foody and Curran 1994; Luckman et al., 1997;
Saatchi et al., 1997; Rignot et al., 1997).
During the Global Rain Forest Mapping (GRFM)
project, JERS-I SAR (Synthetic Aperture Radar)
satellite was used to map the humid tropical
forestsof theworld. The rationalefor the project
was to demonstrate the application of the
spaceborneL-bandradarin tropicalforeststudies.
In particular,the use of data for mapping land
cover types, estimatingthe areaof floodplains,
and monitoring deforestation and forest
regenerationwereof primaryimportance. In this
paper,we examinetheinformationcontentof the
JERS-1SARdatafor mappinglandcovertypesin
theAmazonbasin.
JERS-1 Amazon Mosaic
JERS-I SAR is an L-band spacebome SAR
system launched by the National Space
Development Agency of Japan fNASDA) in
February, 1992. The system operates at 1.275
GHz with horizontal polarization for both
transmission and reception. The spatial
resolution of the system is 18 m in both azimuth
and range. The swath width is 75 km and the
incidence angle of radar at the center of swath is
38.5 °. The single-look images have 4.2 m pixel
spacing in azimuth and 12.5 m in range and the
standard three look image has 12.5 m pixel
spacing in both azimuth and range.
In late 1995, the JERS-1 satellite entered into its
Global Rain Forest Mapping (GRFM) phase to
collect high resolution SAR data over the entire
tropical rainforest. In approximately 60 days, the
satellite obtained wall-to-wall data over the
Amazon basin. Because of cloud cover, similar
coverage with high resolution optical data such as
Landsat could only be provided on annual or
decadal time frames. The JERS-1 coverage of the
Amazon basin is shown in Figure 1.
We have used I00 m resolution JERS-1 data (8
by 8 averaging of high resolution 12.5 m three-
look data) to generate a map of the entire basin
from 1500 images. The spatial mosaicking
technique is based on a mathematical wallpapering
approach which minimizes the propagation of
errors. The inter-scene overlaps both in the
along-track and cross-track directions were used
for individual scene geoiocation. The scenes were
placed on a global coordinate system with the
flexibility of having scenes float freely with
respect to one another. The locations of all scenes
were calculated simultaneously, avoiding any
directional errors. The result is an optimum
seamless mosaic (Figure 2)
Figure 1.
basin.
¢)..... • ,
..[ '_" [zr,,._:i _ i
,l' '¢' /'
, / j'
Ifl_
/ i _-;'-J
/ _ ,_. j
JERS-I coverage map of the Amazon
Figure 2. JERS-1 Amazon mosaic image during
the dry season.
Methodology
In classifying the JERS-I SAR data, we
developed texture measures from the 100 m
JERS-I mosaic over a 10 x l0 window, resulting
in 1 km resolution texture images with
independent pixel intormation, as the windows
were shifted in a blockwise fashion in a 10-pixel
increments. These first order histogram statistics
characterize the frequency of occurrence of grey
levels within the window in the single channel
radar data and are sensitive to window size which
_as determined a priori. The JERS-I mosaic
image in figure 2 shows orbital stripes which are
due to slight radiometric discrepancy between the
calibration of orbital data takes. These calibration
discrepancies which are often less than 0.5 dB in
intensitydo not affect the texture measures and the
classification of the image at a resampled
resolution of 1 km.
Eight texture measures were calculated for
classification from the first order histograms
(mean, variance, energy, entropy, contrast,
kurtosis, skewness, coeffiecient of variation).
Training areas were chosen from the mosaic
image by consulting the RADAMBRASIL
vegetation map of Amazonia. The class
separability using all texture measures were
calculated by using the B-distance (Bhattacharyya
distance). The separability test helped us to
identify the significance of texture measures in
classifying each land cover types.
After choosing a list of texture measures,
we employ a two stage approach to perform a
supervised classification of the JERS-1 mosaic.
In the first stage, we use a maximum a posteriori
Baysian (MAP) classifier on the JERS-I mean
backscatter image at 1 km scale to select five
general categories of land cover types (forest,
nonforest, savanna, floodplain, open water)
In the second stage, the texture measures and the
MAP classified image were used in a hierarchical
decision rule based algorithm to further
discriminate more specific vegetation types within
each of five general land cover categories. The
decision rules were derived by using predictor
variables obtained from the multi-dimensional
separability analysis of the backscatter and texture
measures for each vegetation type. The rules for
texture measures are calculated irr a hierarchical
fashion using the normal distribution for the
statistics of training areas. This assumption was
verified when using the B-distance separability
test. Once the training data set was obtained and
the MAP classified image was produced, the
decision rules were determined in an automated
procedure in order to guarantee the repeatability of
the classifier. Note that the input texture images
to the classifier and the training data sets are
derived from the 100 m JERS-I using a 10 x 10
window. The use of 100 m backscatter image
rather than the 1 km images for collecting the
training data, helps avoiding mixed pixel
information in training data set.
Discussion
A total of twenty sites were chosen for each
vegetation type, of which ten were used for
training the classifier and ten for verifying the
accuracy of classification. To make the selection
of training and validation sites as accurate as
possible, we registered the RADAM map with the
JERS-I mosaic and extracted the backscatter and
texture measures for each land cover category as
indicated on the map. During the selection of
these, we avoided those cases where the RADAM
vegetation types did not match with the general
characteristics of the radar image suggesting a
possible error in the RADAM data.
A total of 14 classes were separated JERS-I
mosaic. The accuracy of classification were
calculated by developing a confusion matrix of
classes, comparison with the RADAM map, and
pixel-by-pixel comparison with the AVHRR land
cover map. The results and detailed discussion
are reported in Saatchi et a1.(1999).
Acknowledgment
This work is performed at the Jet Propulsion
Laboratory, California Institute of Technology,
under contract from National Aeronautic and
Space Administration.
References
Foody, G.M., and Curran, P.J., 1994, J. of
Biogeograhy. 21:223-244.
Luckrnan, A., Baker, J., Kuplich, T.M.,
Yanasse, C., and Frery, A.C., 1997, Remote
Sens. Environ., 60:1-13.
Rignot, E., Salas, W., and Skole, D.L., 1997,
Remote Sens. Environ., 59:167-179.
Saatchi, S.S., Soares, J.V., and Alves, D.S.,
1997, Remote Sens. Environ. , vol. 59: 191-202.
Saatchi, S., Nelson, B., Podest, E., Holt, J.,
1999, Int. J. Remote Sens., In press.
Sader, S.A., 1987, Phtogramm. Eng. Remote
Sens. 53(2): 193-202.
Skole, D.L. and Tucker, C.J., 1993, Science,
260:1905-1910.
Veloso, H.P., Rangel Filho, A.L.R.& Lima,
J.C.A., 1991, IBGE. Rio de Janeiro. 123 pp.