Soil Erosion Prediction in Western Kazakhstan Through Deep Learning with a Neural Network Approach to LS-Factor Analysis

Abstract

With the rapid shifts in environmental conditions, accurately predicting soil erosion has become crucial for the sustainable management of land resources. This study introduces a deep learning-based approach to forecast soil erosion risks in Western Kazakhstan up to 2030, focusing on the LS factor defined by the Universal Soil Loss Equation (USLE). High-resolution digital elevation models (DEMs) from ASTER GDEM and historical data on climate and land use were utilized to train a convolutional neural network (CNN), enabling projections of future LS-factor changes and the corresponding erosion risks. To further improve the accuracy of LS-factor calculations, the System for Automated Geoscientific Analyses (SAGA) was applied using a multiple-flow algorithm. The results forecast a significant rise in erosion risk by 2030, with areas having LS values between 8 and 24 expected to increase by 10%, and those with LS values above 24 by 0.05%, potentially affecting an additional 24,000 km2. The model achieved a 92% accuracy rate, underscoring the effectiveness of deep learning in environmental risk analysis. The integration of SAGA results provides a more detailed understanding of the erosion processes, enhancing the precision of the predictions.