Predicting Soil Physical-Hydric Attributes Based on Pedotransfer Functions and Algorithms for Quantitative Pedology

Understanding soils physical-hydric attributes is an important factor for studies focused on investigating how the water content stored and available in hydrographic basins affects soils environmental functions, biodiversity and sustainability. Thus, the current study aims to estimate basic infiltration rate (bir) and saturated hydraulic conductivity (Ksat) associated with soil granulometric and physicochemical properties, based on pedotransfer functions implementation, geospatial analysis and machine learning algorithms. Soil profiles quantitative analysis was carried out by using algorithms for quantitative pedology (AQP) and, afterwards, three pedotransfer functions were implemented using multiple linear regression, regression trees and Random Forest. Results showed AQP as a potential tool for physical-hydric soil attributes in-depth studies, as well as highlighted soils parameters variability and their implication on bir and Ksat behavior. Based on statistical metrics, validation process showed Random Forest as the best performance model at Ksat and bir estimation (R \({ }^2\) = 0.9409, MAE = 0.0015, and RMSE = 0.0033 for Ksat; R \({ }^2\) = 0.9466, MAE = 0.0902, and RMSE = 0.1922 for bir). Tree-based models were capable of modeling soils physical-hydric attributes, whereas regressive models presented low estimate accuracy due to their inability to handle the observed high variability in bir and Ksat attributes and their dependency on stricter statistical assumptions.