Monitoring reservoir storage using remote sensing and large language models

Despite advances in proximal and remote sensing (RS) of surface-water reservoirs, many regions still lack reliable and timely storage records due to sparse or non-public gauging hindering sensor validation. Here we show that validation of synthetic-aperture radar (SAR) water-extent maps against quantitative reservoir storage values extracted from online media information mined with large language models (LLMs) can overcome this limitation. Our innovative framework processes Sentinel-1 SAR imagery in Google Earth Engine to delineate water by sweeping a backscatter threshold T [dB] and a median-filter radius r [m], after which mapped area is translated to storage V_sat [hm³] using a design area-storage relationship. An independent storage reference timeseries V_ref [hm³] is compiled from publicly available online media using a Custom Search Engine (CSE) for document discovery and an LLM (GPT-4o-mini) to extract quantitative values and associated dates, followed by human curation to ensure consistency. Best-fit SAR parameters (T, r) were identified separately for ascending and descending orbits by maximizing Kling-Gupta Efficiency (KGE) while minimizing RMSE. Performance and uncertainty are assessed using paired comparisons and block-bootstrap confidence intervals that account for temporal autocorrelation. The proposed approach is applied to reconstruct a 10-year storage time series of the Aposelemis Dam, the most important domestic water supply project in Crete, currently affected by droughts. The curated CSE/LLM pipeline yielded 82 distinct date-storage pairs. Across orbit-specific configurations, KGE values of 0.96-0.98 and RMSE of 1.5-2.0 hm³ were achieved, with residual analysis indicating limited bias but sensitivity to parameter choice and reference data quality. While uncertainties remain due to SAR threshold selection and heterogeneous media reporting and limitations of the validation references, the results show that combining SAR with LLM-assisted, human-guided validation data provides a scalable pathway for reservoir monitoring in data-scarce regions.