Deep learning-based decoupling of hydro-meteorological signals from Cheonji caldera lake water levels for volcanic precursor detection

Although lake level fluctuations in caldera lakes serve as critical precursors reflecting the hydro-thermal instability of magma systems, their utility as precise monitoring indicators has been limited because volcanic signals are often masked by meteorological variability, such as precipitation and evaporation. This study proposes a novel deep learning-based framework that integrates long-term Landsat and Sentinel-2 optical satellite imagery (1987–2021) to extract candidate volcanic water-level anomalies by decoupling meteorological factors. The proposed framework consists of three key steps: 1) precise segmentation of water bodies from heterogeneous satellite imagery using a multi-encoder SiU-Net model, followed by DEM-referenced water level retrieval via integration with the Copernicus DEM and independent cross-validation against ICESat-2 ATL13 altimetry; 2) quantitative modeling and decoupling of nonlinear meteorological effects through a five-variable (T, H, PP, DiY, SAT) LSTM-based Multivariate Influence Correction Model (MICM) driven by ERA5-Land reanalysis data; and 3) identification of volcanic activity periods using the finally corrected time-series water-level anomalies. Experimental results demonstrate the robustness of the constructed time series, with SiU-Net achieving a high F1-score of over 0.99 for both Landsat and Sentinel-2 imagery. The ERA5-Land-driven MICM achieved a water-level RMSE of 0.4284 m, and the resulting anomaly distribution (μ = +0.063 m, σ = 0.424 m) yielded a 99% confidence interval of ±1.092 m for outlier detection. Three statistically significant water-level anomalies — at DOY 248 of 2000, DOY 216 of 2006, and DOY 267 of 2010 — temporally aligned with independently documented Mt. Baekdu unrest periods, including the 2002–2005 magma injection episode and its 2006–2010 continued-injection phase. These findings suggest that the proposed satellite-based hydrological monitoring approach using deep learning can serve as a complementary screening tool for volcanic surveillance alongside traditional geophysical observations.