Integrated Analysis Of Multi-Source Satellite Data Via Deep Learning For Predicting Lake Water Level Dynamics

Abstract

Accurately monitoring lake water levels is critical for identifying potential hazards and developing sustainable water management strategies. While previous studies have primarily relied on single-source satellite data or traditional indices, this study enhances existing approaches by integrating multi-source satellite imagery (Sentinel-2) with atmospheric parameters (Sentinel-5P) and meteorological data (temperature and pressure), providing a comprehensive analysis of hydro- logical dynamics. To support sustainable water management, the water levels of Burdur and Iznik lakes selected as reference in Türkiye were estimated using the Long Short-Term Memory (LSTM) model for the 2018–2025 time period. Additionally, the Simple Water Body Mapping (SWBM) technique, combining Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (MNDWI), and Sentinel-2 Water Index (SWI) indices through region-specific threshold calibration and advanced cloud-filtering methods, was employed, significantly improving the accuracy and robustness of lake-area mapping compared to traditional methods. In the study, real measurements were compared with satellite-derived measurements, and the accuracy of the remote sensing-based results was proven with an error rate of 0.07%. For Lake Burdur, a Root Mean Squared Error (RMSE) of 16.93 was obtained in the training dataset and 20.38 in the test dataset, achieving an explained variance of 91.6%. For Lake Iznik, RMSE values of 10.84 in training and 19.42 in testing were obtained, with an explained variance reaching 92.5%. These results clearly demonstrate the superior capability of integrating multi-source data and advanced deep learning techniques in hydrological modeling, thereby providing a robust and scalable predictive framework that significantly improves decision-making processes for sustainable water resource monitoring and management.