Interpretation of Bayesian-optimized deep learning models for enhancing soil erosion susceptibility prediction and management: a case study of Eastern India

Abstract

Soil erosion poses a significant threat to sustainable land management and agricultural productivity. Addressing this issue requires advanced predictive models that can accurately identify areas at risk and inform soil conservation strategies. This study focuses on the development and interpretation of well-optimized deep learning (DL) models to predict soil erosion probability, aiming to enhance decision-making in land management. Utilizing the Revised Universal Soil Loss Equation (RUSLE) in conjunction with ground-truthing, we identified critical erosion-prone areas. To predict soil erosion probability, we employed Bayesian optimization to fine-tune Deep Neural Network (DNN), Convolutional Neural Network (CNN), Fully Connected Neural Network (FCNN), and DNN-CNN hybrid models. These DL models were verified using a set of metrics. SHAP value analysis as a means of explainable artificial intelligence (XAI) was used to interpret these DL models for better decision-making. RUSLE estimations and ground truthing highlight that Soil erosion rates in the northeastern and northwestern regions are nearing the highest observed at 25 tonnes per hectare annually, largely due to steep slopes and limited vegetation. In contrast, the southern and southeastern areas have lower erosion rates, due to denser vegetation and gentler slopes. Deep learning models, optimized using Bayesian methods, demonstrate high performance in spatially modeling soil erosion probability. The DNN model achieved an accuracy of 0.93, a precision of 0.92, and an F1-score of 0.94, identifying 222.73 sq. km as highly susceptible to erosion, which indicates its strong ability to detect true erosion events. The CNN model identified 49.68% of the study area (503.30 sq. km) as high-risk, with an accuracy of 0.90 and a precision of 0.91. The FCNN model showed a balanced risk distribution, indicating 37.04% of the land (375.25 sq. km) as very low risk and 36.43% (369.11 sq. km) as very high risk, with an accuracy of 0.91. The DNN-CNN hybrid model highlighted 41.58% of the area (421.20 sq. km) as high risk, demonstrating its effectiveness in capturing spatial patterns of erosion susceptibility. SHAP value analysis indicates that land use and soil type (LULC and K-factor) are crucial in erosion predictions, with LULC having a significant predictive influence in the DNN model. These insights facilitate the prioritization of soil conservation measures, enabling decision-makers to focus on the most impactful factors for mitigating soil erosion.

Keywords

• Soil erosion prediction
• deep learning
• Bayesian Optimization
• SHAP values
• sustainable land management