Digital soil mapping based on map extrapolation between physiographically similar areas

Abstract

The objective of this work was to test methodologies for digital soil mapping (DSM) and to evaluate the possibility of map extrapolation between physiographically similar areas. The reference area for model training was located at the municipality of Sentinela do Sul, in the state of Rio Grand do Sul (RS), Brazil, and the extrapolation was done for the municipality of Cerro Grande do Sul, RS. Models were developed by DSM using environmental variables as predictors, and soil classes – obtained from a conventional soil survey at 1:50,000 scale – as dependent variables. The combined use of two decision trees (DT), trained in two landscapes with different drainage classes, was tested. For Sentinela do Sul, the agreement between the predicted maps with the ones produced by conventional survey was evaluated using error matrices. Since the importance of mapping errors is variable, a weighted error matrix was created to assign different importances to specific mapping errors between different mapping units. Map accuracy of Cerro Grande do Sul was evaluated by ground truth. Map extrapolation yields satisfactory results, with accuracy higher than 75%. The use of models with two DTs divided by homogeneous landscapes generates extrapolated maps with a greater accuracy, evaluated by ground truth.