Strategies for digital soil mapping by artificial neural networks based on soil-landscape

Abstract

The scarcity of land information to enable its proper use, whether for agricultural, environmental and urban design, can be minimized by solutions from the development of new technologies. Accordingly, this study aimed to apply two strategies to obtain digital maps of soil in areas where no preliminary surveys werecarried out conventional pedological. The strategies were implemented based onenvironmental variables that establish relations between the occurrence of soils andtheir positions in the landscape. The study area comprised the municipality of BarraBonita, SP, totaling 11,072 ha. For use in the prediction of soil by the technique ofArtificial Neural Networks (ANN) were used variables: slope, elevation, profilecurvature, plan curvature and convergence index derived from a Digital Elevation Model (DEM), in addition to information geology and geomorphic surfaces identified in the region. In the first strategy, through a cluster analysis (Fuzzy k-means) of variables, we selected five key areas distributed in the study area, soil survey beingconducted semi-detailed level at these sites for recognition of the map units. Instrategy 2, a map was drawn up detailed level of soil from pre-existing data of onlyone key area, located in the center of the region. Identifying the map units were generated files for training and testing of neural networks. Was used the simulator JavaNNS and learning algorithm "backpropagation". Sets environmental variableswere tested by assessing the importance of each variable to predict soil. The networkshowed better performance for the Kappa index was used to generalize their information, obtaining the digital soil maps. By applying cross tabulation analyzed the spatial correspondence between the digital maps and a conventional map of the region. Reference points were collected to validate the performance of digital maps. According to the position in the landscape and the underlying source material, was noticed a tendency of occurrence of soil classes in key areas mapped. The same arrangement was observed in the soil classifications digital. The attributes ofthe terrain elevation and slope exhibited a greater influence on the distinctionbetween the soil by the neural networks in both strategies. The comparison with reference points showed that the digital map produced based on mapping units from the conventional approach detailed outperformed (81.8% agreement) to the mapbased on pedological survey of semi-detailed level (72.7 %). This study showed that to obtain digital maps of soils, use of environmental variables that express the soillandscape relationship, may contribute to the generation of information preeliminares soil in areas not mapped from map units obtained from adjacent areas.