Prediction of Water Infiltration of Three Types of Soil with Machine Learning in the Sahuayo River Basin.

Infiltration is the process by which surface water enters soil, is part of the regulatory ecosystem services, and is of the greatest importance for environmental and agricultural management. Its spatial and temporal variability is particular for each soil type and intended usage. The objective was to determine which of the land types and land uses provide the most ecosystem services based on their water infiltration and its spatial prediction. This study evaluated the Sahuayo River basin in the state of Michoacán, Mexico, on agricultural and agricultural land uses (Luvisols, Vertisols, and Leptosols). Sixty soil samples were collected at stratified random points, and their organic matter content, texture, and bulk density were assessed via laboratory methods. An analysis of variance with multiple means comparisons was performed on the infiltration variable results. The SCORPAN model was used to predict the infiltration spatial distribution using three machine learning algorithms (with default and hyperparameter tuning) and 11 variables. These variables correspond to topographic attributes, climate, soil properties, and remote sensing data. The model's predictive power was evaluated using its mean square root error (RMSE) and mean absolute error (MAE) values. The nonagricultural Leptosols had the highest infiltration capacity (fp = 2210.6 mm · h−1), whereas the agricultural Vertisols had the lowest infiltration capacity (fp = 275.8 mm · h−1). Both combinations significantly differed from each other. The highest precision model was recorded with the artificial neural network algorithm, which reported the lowest MAE and RMSE values for both variable selection methods in their default configurations and with hyperparameter adjustment. [ABSTRACT FROM AUTHOR]