Otimização de redes neurais artificiais para a predição aprimorada de funções de pedotransferência em modelagem computacional do solo

Abstract

Pedotransfer functions (PTFs) play a key role in estimating soil hydraulic properties, such as the soil water retention curve, from simpler and less expensive physical attributes. Hydraulic characterization of soils is a recurring challenge, especially in tropical regions, due to spatial heterogeneity and the difficulty of direct measurements in laboratory or field conditions. In this context, artificial neural networks (ANNs) emerge as a promising alternative, capable of capturing complex nonlinear relationships and overcoming the limitations of traditional statistical methods. This study aimed to optimize ANN architectures for the prediction of PTFs using three distinct datasets under different edaphoclimatic conditions: B1 (Brazil – Northeast), B2 (Brazil – Acre), and D1 (Denmark). Hyperparameter tuning and Bayesian optimization techniques were applied, considering different activation functions, number of layers, neurons, and regularization rates. The results demonstrated that the final model achieved expressive performance values, with an average coefficient of determination higher than 0.90, root mean square error (RMSE) around 0.03, and mean absolute error (MAE) close to 0.02, indicating high accuracy in predicting soil volumetric water content. It can be concluded that ANNs show strong potential in PTF modeling, constituting a robust alternative for applications in hydrology, agriculture, and sustainable soil management, although future studies should consider a greater diversity of data and comparisons with other machine learning algorithms.