Machine learning-based multisource remote sensing data fusion for chlorophyll-a downscaling inversion

Chlorophyll-a (Chl-a) concentration serves as a critical indicator for monitoring aquatic ecosystems. However, existing satellite observations face inherent limitations where high-temporal-resolution sensors lack sufficient spatial detail, while high-spatial-resolution satellites are constrained by extended revisit cycles, limited swath coverage, and cloud contamination, making them inadequate for precise coastal water quality monitoring. This study develops an advanced Chl-a retrieval methodology through the integration of multi-source remote sensing data fusion and spatial downscaling techniques. The approach employs least squares regression to harmonize Landsat-8 and Sentinel-2 reflectance data, subsequently identifying 11 optimal spectral band combinations demonstrating strong correlation (R>0.85) as inversion predictors. By implementing an XGBoost machine learning framework, the system successfully downscales Himawari-8's 5km Chl-a products to 30m resolution. Validation results demonstrate the model's robust performance in the Pearl River Estuary during autumn and winter seasons, achieving consistent R² values of 0.84. The derived 30m Chl-a distributions exhibit significant spatiotemporal patterns while maintaining strong correlation with reference data. This methodological advancement provides enhanced data products for marine environmental monitoring and harmful algal bloom early warning systems, while establishing a transferable framework for high-resolution retrieval of other aquatic parameters through innovative multi-sensor synergy and machine learning applications.