less accuracy. These latter classes also yielded a high agreement 
with ground truth data (89.0%, 79.9% and 83.7%). The corresponding 
percent commissions, however, were higher at 36.4% and 19.5%. Some 
other classes, such as dense urban area and mixed and cultivated 
fallow land, were more spectrally heterogeneous and had low overall 
classification accuracies (58.7% and 53.2%) and high commission 
percentages (27.2% and 58.0%). 
These results indicated that classification became less accurate as 
the level of detail was increased. The relatively low accuracies of 
some categories at Level II could be attributed to cross-confusion 
among related categories. To obtain more reliable information 
regarding the land use and land cover classes, SPOT images from 
several seasons would be necessary. 
BACKGROUND DISCUSSION 
Remote sensing and geographical information systems (GIS) are 
important and advanced tools for the inventory and analysis of 
natural resources for regional and local planning. Various types of 
remote sensing data (MSS, TM, AVHRR, etc.) have been used for natural 
resources management. Hill and Megier (1986) performed a digital 
classification of the Ardeche region of southern France using Landsat 
5 data as part of a region-wide resources inventory. Pettinger 
(1982) performed a comprehensive digital classification of vegetation 
and land cover in Idaho producing maps at different levels of detail 
for natural resources management purposes. LaBash et al . (1989) 
conducted a digital image analysis of Landsat TM data Tn eastern 
Connecticut for regional land use and land cover classification. 
Civco (1989) concluded that knowledge-based image analysis for 
classifying Landsat Thematic Mapper region-based spectral data, 
coupled with ancillary digital spatial information, not only is 
feasible but also preferable to the per-pixel, spectral data only, 
statistical methods more traditionally employed in deriving land use 
and land cover information for natural resources management. This 
approach produces results both more accurate and more visually 
comprehensible than the traditional method of maximum likelihood 
classification. Furthermore, SPOT HRV data are becoming widely 
available and provide opportunities for increased spatial resolution 
and temporal coverage for mapping. For example, the authors have 
initiated a comprehensive land use planning study in the semi-arid 
regions of northeastern Brazil using SPOT HRV data. Guebert et al. 
(1989) concluded that in the geologic and climatic setting, it is 
possible to characterize the infiltration capacities of disturbed 
watersheds of central Pennsylvania by remotely sensed data. 
Because of the relationships between infiltration capacity and 
surface features observed by SPOT (vegetation and rock type), 
infiltration can be generalized into low, moderate and high in 
filtration capacity categories. These relationships hold for 
surfaces of similar rock type and vegetation, although as climate and 
rock type change, the magnitude of the relationship may vary. 
348