Soil prediction using artificial neural networks and topographic attributes

Because relief maps show a strict relationship with soils at different spatial levels, distributions of soil units can be inferred from digital topography analyses. Geoprocessing techniques can be used to create parametric relief representations from digital terrain models (DTMs), and these models can be used to calculate primary and secondary topographical attributes, such as the elevation, profile and plan curvature, slope, stream power index, topographic wetness index, and sediment transport index. The classic method of pedological cartography is onerous and time-consuming; as an alternative, pedometric techniques favor the recognition of preliminary mapping units. In this study, a multilayered perceptron artificial neural network (ANN) with an error backpropagation algorithm was used, where topographical and geological attributes were used as input parameters. The classified map was validated by comparison with two preexisting conventional ground maps of the study area. The kappa (K) index, global exactness (GE), and exactness from the point of view of the producer and user were considered in the comparison. The quality of the soil units classified by the ANN was satisfactory, based on the K and GE values from the comparison.

► Topographical and geological attributes were used as input parameters for ANN.
► The resulted map was validated through comparison with two conventional soil maps.
► The use of ANN proved to be feasible for classifying preliminary soil map units.