Assessing changes in soil electrical conductivity under runoff influences using gradient boosting and metaheuristic algorithms

Soil salinization is a significant challenge in irrigated agriculture, yet there is a lack of comprehensive data-driven models that analyze the interactions between salinity and other soil properties. This study addresses the existing research gap by utilizing advanced machine learning models, specifically CatBoost and HGBoost algorithms, to predict electrical conductivity (EC) in the soil of humid tropical catchments in central Panama. These algorithms are then applied one by one and optimized with some of the modern optimizers such as Gray Wolf Optimizer (GWO), Phasor Particle Swarm Optimization (PPSO), and Hybrid Gravitational Search (HGS) to predict with further improved accuracy. After meticulous data acquisition and preprocessing, the dataset, encompassing parameters such as Runoff rate, Day, Month, Hour, and Minute, was subjected to in-depth analyses like correlation matrix and sensitivity assessment. Much of the study was dedicated to model evaluation, encompassing training–testing data bifurcation (80% training and 20% testing) and performance metrics analysis. The results underscored the superior predictive capabilities of the hybrid HGBoost models, especially the HGBoost-HGS model, with an R² value of 0.97046. The study culminates by offering insights into model convergence, runtime, and error metrics, contributing to the broader discourse on salinity prediction in irrigated farming using machine learning.