DIGITAL IMAGE PROCESSING AND GIS APPLICATIONS FOR NATURAL RESOURCES
MONITORING AND EVALUATION IN NORTHEASTERN BRAZIL
H.S. Teotia 1,5 , W.C. Kennard 2 , D.L. Civco 3 and K.A. Ulbricht 4,5
Laboratory for Remote Sensing, Department of Naturai Resources
Management and Engineering, The University of Connecticut,
Storrs, CT 06269-4067 USA
ABSTRACT
A portion of a SPOT HRV image data acquired on 5 May 1987 centered
around Santa Luzia and its environs in Paraiba, Brazil, was processed
to extract earth resources information. Digital analysis
supplemented with selective ground truth data gathered by field
surveys in 1988 was applied to the SPOT multispectral data to derive
land cover and other natural resources information. The data were
processed with ERDAS (Earth Resources Data Analysis System) software
operating on a high performance microcomputer. Training areas were
selected for 15 categories of land use and land cover. A supervised
classification was conducted and the digital land use and land cover
map was corrected geometrically and georeferenced to Universal Trans
verse Mercator coordinates. Accuracy was assessed by digitizing
ground truth test polygons and performing a pixel-by-pixel
comparison. Percentage agreement with ground truth data as well as
errors of omission and commission were recorded and analyzed.
By recoding the land use and land cover classes, six geometrically
corrected maps as follows were prepared: 1) land use and land cover,
2) soil associations, 3) land capability, 4) slope classes, 5) land
irrigability, and 6) agro-technical limitations. The first four
resource information layers (maps) were weighted for their relative
importance to land development and irrigation potential and combined
in a logical model to produce a land suitability classification map.
It was found that water, low density urban and barren rocky land,
mixed cultivated and pasture, alluvial cultivated and eroded land,
and forest classes produced the best results in terms of high
percentage agreement with ground truth data (80.0 to 100.0%) and
relatively low percentage commissions (zero to 13.5%). These classes
were easily discriminated from other classes with a high overall
classification accuracy. Other classes such as cotton sparse
caatinga forest and rocky land, cotton cultivated and fallow alluvial
land with dense shrubs/trees could also be distinguished but with
1,. Professor, Federal University of Paraiba, Campina Grande,
Paraiba, Brasil '
2 and 3. Director and Vice Director, respectively, Laboratory for
Remote Sensing
4, Staff Scientist, German Aerospace Research Establishment, DLR -
Optoelektronik , 8031 Oberpfaffenhofen, FRG
5. Visiting Professor at the University of Connecticut when this
paper was prepared.
347
