Abstract

Background and Objective

Land subsidence is one of the major challenges in water and soil resource management in Iran. In recent years, integrating remote sensing data with machine learning algorithms has provided a promising approach for the spatial estimation of subsidence. Therefore, this study aimed to evaluate the simultaneous performance of nine commonly used machine learning algorithms in identifying subsidence across the Iranian affected plains. Another objective was to employ the SHAP interpretability method to determine the actual contribution of each environmental variable to the model output and to identify the optimal model for subsidence estimation, with an emphasis on groundwater-level decline as the most controlling factor.

Materials and Methods

In this research, the land subsidence map of Iran for the year 2020 was used as the reference dataset. A total of 17,000 stratified random sample points were extracted from this map. For each point, 20 information layers were prepared, including climatic variables (precipitation and temperature from 90 synoptic stations), topographic indices (Topographic Wetness Index (TWI), slope, aspect, LS factor, valley depth, valley flatness index (MrVBF)), hydrological factors (groundwater-level decline), anthropogenic factors (distance from roads, faults, rivers, residential areas, agricultural lands, and orchards), and spectral indices derived from Sentinel-2 imagery (NDVI, SAVI, EVI, BSI). All layers were resampled to a spatial resolution of 150 meters and standardized to the WGS84 coordinate system. Subsequently, nine machine learning models were implemented, including Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), Decision Tree (DT), Light Gradient Boosting Machine (LightGBM), k-Nearest Neighbors (kNN), Artificial Neural Network (ANN), and Extra Trees (ET). The data were split into training (70%), validation (15%), and testing (15%) subsets. Model performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). Furthermore, because the regression problem was converted into a binary classification using a critical subsidence threshold of 5 cm/year, ROC curves and AUC values were also calculated. To eliminate the "black box" nature of the models and determine the relative contribution of each variable, the SHAP method, based on Shapley game theory, was applied.

Results and Discussion

The results indicated that the ET model, with an R² of 0.992, the lowest RMSE of 1.001 mm, and an MAE of 0.397, demonstrated the highest accuracy among the nine models. This model showed the greatest alignment with InSAR-derived values, with the majority of data points concentrated along the 1:1 line in the scatter plot and falling within the 10% error band. In contrast, the SVM model exhibited very poor performance, with a negative R² (-0.386), which is attributed to its high sensitivity to data scaling and the need for precise hyperparameter tuning. Models based on Random Forest and gradient boosting (CatBoost and XGBoost) ranked next, with R² values above 0.987. ROC analysis confirmed that the Extra Trees model, with an AUC exceeding 0.90, had the best capability for distinguishing critical areas from stable ones. In terms of interpretability, SHAP plots revealed that across all models, and especially in ET and CatBoost, the variables "groundwater-level decline," "population density," and "climate" had the greatest contribution to subsidence estimation.

Conclusion

The Extra Trees model, combined with the SHAP explainer, is introduced as the most accurate and reliable approach for estimating subsidence across Iranian plains. The application of this framework can reduce field monitoring costs and facilitate the prioritization of critical zones for water resource managers. The main limitation of this research is the lack of consistent long-term temporal data on groundwater-level decline across all plains, which is recommended to be addressed in future studies using dynamic hydrological models.

Keywords: Iran, Subsidence monitoring, Soil and water resource degradation, Artificial Intelligence, Optimal management