Long-term (2000-2024) Daily Water Temperature Reconstruction in the Downstream Jinsha River Based on Multi-source Remote Sensing Fusion and In-situ Calibration

The cascade hydropower development in the downstream Jinsha River has significantly altered the river's thermal regime, exerting profound impacts on the aquatic ecosystem. However, due to severe mixed pixel effects caused by deep-incised canyon terrain, frequent cloud cover, and sparse in-situ monitoring stations, there is a lack of long-term, high-spatiotemporal-resolution water temperature datasets capable of distinguishing the superimposed effects of climate change and reservoir regulation. Addressing the applicability bias of standard remote sensing products in narrow channels and the physical discrepancy between satellite-derived "skin temperature" and ecological "bulk temperature," this study proposes a daily water temperature reconstruction framework based on the Google Earth Engine (GEE) platform, coupling multi-source remote sensing fusion with in-situ calibration.

First, a dynamic water extraction strategy combining the adaptive Otsu method with morphological erosion was employed. This effectively eliminated thermal noise from riverbanks under drastic water level fluctuations, ensuring the acquisition of pure water pixels. Second, a physics-informed machine learning spatiotemporal fusion model was constructed, incorporating Water Fraction and ERA5 meteorological drivers. This model overcomes the limitations of traditional fusion algorithms (e.g., InENVI) in unmixing sub-pixel heterogeneity within canyon waters, successfully generating a daily, 30-meter resolution land surface temperature series from 2000 to 2024. Finally, based on long-term in-situ data from multiple hydrological stations, a "Skin-to-Bulk" bias calibration model accounting for air temperature and wind speed was established to eliminate physical biases in satellite infrared retrieval.

Validation results indicate that the calibrated dataset achieves high consistency with in-situ observations, with a significantly reduced Root Mean Square Error (RMSE), effectively bridging the spatiotemporal gaps of Landsat and MODIS data. Long-term analysis reveals the thermal evolution patterns of the Wudongde-Xiangjiaba reach following cascade impoundment, specifically quantifying the "thermal lag effect" characterized by significantly elevated water temperatures in autumn and winter. This study not only overcomes the technical bottlenecks of water temperature retrieval in complex canyon regions but also provides critical scientific data support for evaluating the ecological effects of large-scale hydraulic engineering and optimizing ecological scheduling.