Identifying deforestation in Brazil using multiresolution satellite data

Abstract

MSS, LAC, GAC and GOES data were used to delineate the extent of deforestation in Rondonia, Brazil, in order to identify those satellite data sources appropriate for monitoring deforestation on a continental/subcontinental scale. These data were processed to differentiate forest from non-forest (cleared, colonized areas) using two different classification procedures. The first procedure utilizes all available spectral bands of data in conjunction with a maximum likelihood classifier to discriminate cleared areas from primary forest. The technique is called probability thresholding. The second employs the red and nearinfrared spectral data to calculate a vegetation index which is subsequently thresholded from forest/non-forest delineation. Ground reference data were not available; the 80m (spatial resolution) MSS digital data products served as the reference data source. The 1·1 km LAC, 4 km GAC and 0·9 km GOES (visible band) images were compared with the MSS imagery. Areal comparisons indicated that (i) the LAC data are capable of adequately delineating colonization clearings in the Amazon; (ii) the spatial resolution of'uhe GAC data is too large to delineate linear clearings of varying length (tens to hundreds of kilometres) up to 2 km wide reliably, (iii) the visible GOES data were of little utility due to excessive data noise and (iv) probability thresholding procedures discriminated forest from non-forest more accurately than vegetation-index thresholding procedures. The results indicate that LAC data used in conjunction with probability thresholding offer the best data-source/classification-procedure combination. MSS data may be used when and where available as a ground reference data source in order to define the AVHRR threshold which most accurately discriminates cleared areas from primary forest.