Machine Learning-Oriented Forecasting Of Soil Degradation Due To Agricultural Land Use Patterns

Abstract

Soil degradation poses a significant threat to agricultural sustainability, reducing soil fertility and crop productivity worldwide. This study presents a machine learning-oriented approach to forecast soil degradation based on agricultural land use patterns, integrating multi-year data on soil properties, climate variables, and land management practices. Four algorithms — Random Forest (RF), XGBoost, Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) — were implemented and evaluated using metrics such as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared (R²). Experimental results revealed that soil degradation was highest in monoculture fields (predicted LSTM value: 74.1) and lowest in fallow lands (27.9). Among the algorithms, LSTM achieved the highest predictive accuracy with RMSE = 5.5, MAE = 3.7, and R² = 0.91, followed by XGBoost (RMSE = 5.8, MAE = 3.9, R² = 0.89), Random Forest (RMSE = 6.5, MAE = 4.2, R² = 0.87), and SVM (RMSE = 7.2, MAE = 4.8, R² = 0.84). Feature importance analysis indicated that soil organic matter, nitrogen content, and crop type were the most influential predictors of degradation. The findings demonstrate that machine learning models, particularly LSTM, can accurately forecast soil degradation, providing a robust tool for sustainable land management, early intervention, and informed decision-making in agricultural planning.