A new application of deep neural network (LSTM) and RUSLE models in soil erosion prediction

Area of Science:

• Environmental Science
• Geoscience
• Data Science

Background:

• Rainfall variability is a major driver of global disasters, increasing soil erosion and related hazards.
• Accurate rainfall prediction is crucial for early detection of soil erosion vulnerability and mitigating storm, drought, and flood impacts.

Purpose of the Study:

• To predict soil erosion probability in Sri Lanka's Central Highlands using deep learning and established models.
• To develop a soil erosion susceptibility map for 2024 and assess future risks.

Main Methods:

• Utilized daily rainfall data (1990-2021) from five agro-meteorological stations for LSTM model simulation.
• Employed the Revised Universal Soil Loss Equation (RUSLE) model integrated with geo-informatics for map generation.
• Validated the 2024 soil erosion susceptibility map against historical data (2000, 2010, 2019) using AUC-ROC.

Main Results:

• LSTM model forecasted monthly rainfall for the next 36 months.
• RUSLE model predicted an average annual soil erosion rate of 11.92 t/ha/yr for the Highlands.
• Approximately 30% of the land area is classified as having moderate to very-high soil erosion susceptibility.
• The soil erosion susceptibility map achieved a high accuracy of 0.93 (AUC-ROC).

Conclusions:

• The study demonstrates the effectiveness of integrating LSTM and RUSLE models for predicting soil erosion susceptibility.
• Findings provide valuable insights for policy-making and land management in erosion-prone regions.
• Further research can explore diverse deep learning models with RUSLE to enhance predictive accuracy.