A cloud-native alternative correction for Landsat-8/9 Collection 2 surface reflectance over inland waters

The Landsat Collection 2 (C2) archive is vital for inland water monitoring, yet the Land Surface Reflectance Code (LaSRC) atmospheric correction for Landsat-8/9 introduces Dark Plume Over Water (DPOW) artifacts. These aerosol extrapolation errors cause severe negative biases in shortwave bands, disrupting long-term consistency. To address this, we developed a cloud-native Alternative Correction (AC) method on Google Earth Engine. This data-driven approach employs random forest regression, trained on spatiotemporally aggregated high-quality water pixels, to reconstruct reliable surface reflectance (SR) from Top-of-Atmosphere observations. Validation against a global in-situ hyperspectral dataset and benchmarking against the physics-based ACOLITE processor demonstrate the robustness of the proposed method. While ACOLITE effectively resolves the negative bias issue, the AC method achieves superior radiometric accuracy, reducing the ultra-blue Root Mean Square Error to 0.019 (compared to 0.029 for ACOLITE and 0.031 for C2 SR). Notably, under high-aerosol conditions, the AC method minimizes the residual spectral distortions often observed in physical inversions, effectively restoring the natural spectral shape. Spatially, the method eliminates DPOW artifacts; furthermore, it removes systematic biases between Landsat-8/9 and legacy sensors (Landsat-4/5/7). By restoring radiometric integrity, this automated solution secures the foundation for reliable long-term global limnology.