Harnessing machine learning and geospatial technologies for precise soil erodibility mapping and prediction

Abstract

Soil erosion threatens fertility and sustainability, with soil erodibility influencing erosion rates based on physical and chemical properties. This study aimed to estimate soil erodibility for various land uses using the K-factor from the Wischmeier equation, assess indicators such as the structural stability index, clay ratio, and dispersion ratio, and develop a predictive model for erosion risk using artificial neural networks (ANN) and geospatial technologies. High-resolution spatial maps of erosion risk were created to inform land management and conservation efforts. An ANN model in MATLAB R2024a predicted soil erodibility as well as indicators such as the dispersion ratio, crust formation, and clay ratio. Statistical analyses, including principal component analysis (PCA) and correlation assessment, were performed with OriginPro 2021b to explore relationships between soil properties. Spatial maps of observed and predicted erodibility were created using ArcGIS 10.7.1. Results showed that erodibility values ranged from 0.023 to 0.152 t·ha·hr·MJ^-1·mm^-1 for the observed data and 0.026 to 0.148 t·ha·hr·MJ^-1·mm^-1 for the predicted values. For different land uses, it included 0.09513t·ha·hr·MJ^-1·mm ¹ for cultivated land, 0.060796 t·ha· hr·MJ ¹ · mm ¹ for forest land, and 0.092685 t·ha·hr·MJ^-1·mm^-1 for pasture land. The ANN model demonstrated high accuracy, with R-values of 0.999 for soil erodibility, 0.996 for the structural stability index (SSI), 0.995 for the clay ratio (CR), and 0.904 for the dispersion ratio (DR). This study effectively combines machine learning and geospatial technologies to predict and map soil erodibility, providing insights for erosion control and sustainable land management.