Applying Artificial Neural Networks Utilizing Geomorphons to Predict Soil Classes in a Brazilian Watershed

Abstract

The use of landscape terrain attributes associated with the field information in geographic information systems (GISs) helps to improve the methods applied in soil survey. Geomorphons is a novel technique to map surface elements from digital elevation model and visibility distance (search radius) of a central point in the landscape, which can adopt flexible scales. The main goal of this study was to evaluate the potential for incorporating Geomorphons, which is used to recognize landscape patterns and to improve the soil class predictions by artificial neural networks (ANNs). The procedures involved the acquisition of a cartographic database, creating digital models that represent landscape attributes relevant to paedogenesis on the research site (including Geomorphons of different search radius), sample collection and description of one hundred soil profiles in predefined locations, and finally the supervised classification by neural networks. The covariates used were as follows: elevation, slope, curvature, combined topographic index (CTI), euclidean distance, clay minerals, iron oxide, normalized difference vegetation index (NDVI), geology, and Geomorphons. All models for the terrain attributes have 30-m pixel resolution, and these variables correspond to neurons in the input layer of the neural networks. The output layer of the supervised classification corresponded to the nine dominant soil classes in the study area. To define the appropriate scale of Geomorphons map, sixteen sets of neural networks contain each one of the terrain attributes plus a Geomorphons map calculated from different search radius. For comparative purposes, one of the sets included no Geomorphons. Selection of the appropriate Geomorphons search radius was based on the statistical indexes obtained from a confusion matrix. The results showed that the best classification used the Geomorphons map obtained by forty-five pixels of search radius, in combination with other variables. This classifier presented values to kappa index and global accuracy corresponding to 0.74 and 77.0, respectively.