Land subsidence hazard mapping using a federated learning architecture with satellite and local environmental data

Abstract

Land subsidence has become one of the most critical environmental and geomorphological hazards in recent years, threatening groundwater resources, urban infrastructure, and ecosystem stability. This study presents a novel Federated Learning (FL)-based framework for land subsidence hazard mapping in Poldokhtar County, Iran, without requiring direct access to centralized raw data. The proposed framework integrates Interferometric Synthetic Aperture Radar (InSAR)-derived subsidence rate maps from Sentinel-1 time-series data (2015–2025) with 13 environmental conditioning factors across multiple spatial clients. Five machine learning and deep learning algorithms Random Forest (RF), Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), Deep Neural Network (DNN), and Logistic Regression (LR) were employed as local base models within the FL framework, using FedAvg and FedProx aggregation strategies. The results revealed generally weak to moderate correlations between subsidence and individual environmental factors, highlighting the complex and multi-factorial nature of the subsidence process. Client-based data distribution analysis confirmed a non-IID (Non-Independent and Identically Distributed) condition, with Client 1 representing severe subsidence areas (wider range of negative values) and Client 3 representing areas with weak or negligible subsidence (narrower range near zero). Among the global federated models, RF achieved the highest accuracy (R 2 = 0.963), followed closely by LSTM (R 2 = 0.956) , demonstrating their strong capability in capturing complex spatiotemporal patterns. In contrast, LR exhibited the poorest performance (R 2 = 0.375) with the highest error rates, confirming its limitation to linear relationships. Feature importance analysis identified Digital Elevation Model (DEM), groundwater level, and soil type as the most influential predictor variables associated with subsidence risk in the study area. The findings demonstrate that the proposed FL-based framework not only enhances prediction accuracy through collaborative learning across distributed clients but also preserves data privacy by avoiding raw data sharing. This approach offers a scalable and effective decision-support tool for sustainable groundwater management and land-use planning in subsidence-prone regions.