Mapeamento digital de solos da região Sul de MG usando novas covariáveis ambientais preditoras

Abstract

Proper land use planning is essential to meet the global challenges posed by food demand and climate change mitigation. To overcome these challenges, soil maps are fundamental tools for landscape interpretation and adaptation of productive activities. Digital soil mapping (DSM) has been used to accelerate the production of maps of large areas, at more detailed scales with lower costs. The improvement of machine learning techniques and the experimentation of new digital products are urgent demands for the development of the DSM, in which this work sought to contribute. Information built up over decades of pedological research was gathered to guide the production of a soil map from a quantitative approach. The legacy information was applied to predictive modeling of soil classes using the Random Forest algorithm. Gamma-ray aerospectrometry data and paleoclimatic models, whose application is unprecedented in the MDS in Brazil, were applied for the mapping of 52,982 km², together with environmental variables derived from the Digital Elevation Model (DEM) and current climate models. The extrapolation level of the model was calculated from the Multivariate Environmental Similarity Surface, and the prediction Entropy was calculated according to the Shannon Entropy formula. The overall accuracy of the prediction was 89% for the mapping of 19 soil classes at suborder level according to the Brazilian Soil Classification System. The most important environmental covariables in the spatial prediction were the gamma-ray aerospectrometry data, the paleoclimatic model of the total annual precipitation estimated 20,000 years ago, and the vertical distance from the drainage network derived from the DEM.