Methodology for Bare Soil Detection and Discrimination by Landsat TM Image

ABSTRACT

The objective of this work was to develop and test a remote sensing technique to determine bare soils with pixel information from satellite images. The methodology was tested and improved on a 2,805 km2 area located in the state of São Paulo, Brazil. The pixel data from a Landsat-5/TM image was transformed into reflectances. 294 pixels were evaluated by five factors simultaneously and included the following: color composition image; vegetation index; soil brightness information (soil line concept), and a comparison between spectral curve of the pixel with spectral patterns of soils. A validation procedure was based on the discriminate analysis for the real soil related with each pixel. For this, a soil map was overlaid onto the image, and the pixels were related to its respective soil class. Soil brightness variations were readily observed in the spectral curves and in red-NIR features and corresponded to differences in texture and particle size as well in iron and organic matter content. Although qualitative, the observation of color composition was useful for pixel identification. The soil line concept was very useful as it presented a high R2 coefficient (0.90). Comparison between ground level soil spectral curves with satellite information could assist on the evaluation of the real format of the curves. Discriminate analysis indicated a 99.3% correct classification of the soils. Field work validation indicated 90% significance. The present method could help researchers acquire valuable information (i.e., soil attributes quantification), when soil data must be acquired from satellite images.

Keywords: