INTEGRATION OF QUICKBIRD DATA AND TERRAIN ATTRIBUTES FOR DIGITAL SOIL MAPPING BY ARTIFICIAL NEURAL NETWORKS

This study evaluated different environmental variables in the digital soil mapping of an area in the northern region of Minas Gerais State, using artificial neural networks. The environmental variables terrain attributes (slope and compound topographic index), the quickbird bands 1, 2 and 3, and lithology were evaluated. The importance of each of the variables in the classification was tested. The "Java Neural Network Simulator" was used with the backpropagation learning algorithm. For each dataset a neural network was created to predict the soil mapping units, and the map produced by the nets was compared with the conventional, to show the general accuracy of each one. The best classification was achieved when all variables were used, with an accuracy of 67.4% compared to the the conventional soil map. Of the variables, slope was most significant, because when excluded from the dataset, the classification was worst (accuracy 33.7%). This result showed that the approach can contribute to overcome some problems of soil mapping in Brazil, especially at scales larger than 1:25,000, with faster and cheaper execution, mainly if remote sensing data with high spatial resolution and an affordable price are available, and good digital elevation models to generate the terrain attributes in the geographical information systems.