350
lands and includes natural and artificial reservoirs and lakes.
Elevations of the arpa range from 270 to 900 m above sea level . The
area lies between 36° 50'W to 37° 05'W longitude and 6° 47 1 S to
6° 57'S latitude. The area is drained by a network of the Sabugi
and Farinha rivers which form lakes at Lagoa do Meio, Barra and Acude
Publico de Santa Luzia. Limited parts of the area are irrigated with
reservoir waters, but most crop production is dependent upon natural
rainfal1.
Image Interpretation
Level IB multispectral data from the SPOT-1 HRV scene dated 5 May
1987 were processed with ERDAS (Earth Resources Data Analysis System)
software operating on a PC's Limited 30386-based microcomputer. This
system consists of hardware and software for the extraction and
reformatting, radiometric and geometric correction, classification,
and display of digital images. The system is capable of displaying
512 x 512 pixel images in as many as 16 million colors. The hard
copy color display can reproduce scaled images and maps in as many as
4098 colors and patterns. The ERDAS configuration consists also of a
9-track tape drive and high volume disk storage devices.
The interpretation was divided into two stages: 1) unsupervised
classification of the SPOT image data to determine the grouping of
pixel values within the scene (a total of 20 spectral classes were
identified) and 2) selection of training areas to be used in
supervised maximum likelihood classification. Following a modified
version of the U.S. Geological Survey procedures (Anderson et al.
1976) using the spectral cluster characteristies of the scene as a
guide, a set of 15 combined Level II categories were chosen as
classes for supervised training (Table 1).
Training areas for the 15 classes interactively selected from a SPOT
false-color composite image were displayed on the monitor using 80
training areas for interpretation. Finally, 156 categories were
chosen as classes for final supervised training. At least three
different training polygons dispersed throughout the entire scene
were selected for each of the 15 categories. The training areas were
correlated with ground truth observations gathered in the field
during March-April and October-November 1988.
The relevant statistics (mean, standard deviation, variance, and
co-variance matrices) were generated for all the training areas and
a maximum likelihood classification was applied to the one-million
pixel image. After inspection of the initial classification, certain
categories were aggregated and others deleted to reduce potential
misclassification while still retaining maximum information. This
resulted in the final classification of 15 categories at Level II.