Modelos preditivos da compactabilidade e dos limites de consistência de solos tropicais com sensores proximais

Abstract

This master's thesis, developed by Michel Coutinho de Souza at the Department of Soil Science at UFLA, proposes improving of the physical-mechanical characterization of Brazilian tropical soils through advanced predictive modeling and the use of proximal sensors such as pXRF and Nix™ Pro. The work, based on 66 soil horizons from 10 distinct classes collected in LavrasMG with great textural and mineralogical diversity, is structured in two main articles. The first article focuses on predicting the maximum bulk density (BDmax), identifying texture and moisture as the most efficient individual predictors and validating that the use of algorithms such as Partial Least Squares (PLS) and linear regressions (EXP), associated with proximal sensors, significantly reduces costs and laboratory logistics without loss of precision. The second article focuses on modeling Atterberg consistency limits (Liquidity, Plasticity, and Contraction) and optimum moisture content, concluding that the Polidori chart is more suitable for classifying tropical soils than the traditional Casagrande chart. The results of this second study demonstrated that the Random Forest algorithm is superior for estimating consistency limits by capturing complex nonlinear physicochemical interactions, whereas linear and PLS models are highly effective for predicting the optimum moisture content. In short, the research confirms that combining proximal sensors and artificial intelligence generates robust pedotransfer functions, offering a fast and economical alternative for agricultural planning and geotechnical engineering.