Daily Lake-Surface NDVI Reconstruction Using Multi-Source Machine Learning Under Incomplete Optical Observations

Accurate and continuous monitoring of cyanobacterial blooms is essential for lake ecosystem management, but optical remote-sensing observations are frequently limited by cloud contamination and illumination conditions. To address this limitation, this study proposes a multi-source machine learning framework for daily lake-surface normalized difference vegetation index (NDVI) reconstruction under missing optical observations by integrating Cyclone Global Navigation Satellite System (CYGNSS) observations, ERA5-Land meteorological variables, geographic coordinates, and the CatBoost (version 1.2.10) regression algorithm. Lake Taihu and Lake Chaohu, two representative eutrophic lakes in eastern China, were selected as study areas. An ablation analysis was conducted to evaluate the contributions of different predictor groups. Under the random within-domain evaluation, the full-feature model combining GNSS-R observables, meteorological variables, and geographic coordinates achieved an average test R2 of 0.63 and a root mean square error (RMSE) of 0.15 in GNSS-R-covered regions. Compared with the model using meteorological variables and geographic coordinates alone, the inclusion of GNSS-R observables provided additional predictive information, indicating that GNSS-R served as a supplementary rather than dominant information source. In regions without GNSS-R coverage, meteorological variables and geographic coordinates were used to maintain spatially continuous NDVI reconstruction over the lake surface. The annual results show that the framework can generate daily reconstructed NDVI estimates for both lakes within the study domain. Cross-product comparison with Fengyun-3F (FY-3F) NDVI indicated broad consistency in the major spatial patterns of lake-surface high-NDVI signals, particularly in open-water areas. However, differences in spatial resolution and temporal compositing limit direct pixel-level assessment. Interannual analysis suggested that the frequency and magnitude of positive-NDVI surface signals generally decreased after 2022; however, these signals should be interpreted as integrated lake-surface ecological responses rather than direct quantitative indicators of cyanobacterial bloom intensity. A Shapley Additive Explanations (SHAP) analysis showed that geographic coordinates, GNSS-R surface reflectivity, temperature, and wind speed contributed to the model predictions. Overall, the proposed framework provides a lake-specific empirical approach for maintaining spatially continuous daily NDVI information when optical observations are incomplete.