An integrated atmospheric-topographic correction framework for land surface reflectance estimation using a spatial-spectral Attention U-Net model

Surface reflectance, a fundamental parameter for deriving high-level satellite products, is typically obtained through atmospheric correction. However, in rugged terrains, topographic effects substantially distort surface reflectance estimates. Many post-topographic correction methods for surface reflectance were proposed and applied to mitigate topographic effects. However, the coupled atmospheric and topographic influences in radiative transfer processes cause significant errors when these corrections are performed separately. For example, we observed over 50 % of pixels in the official Landsat 8 surface reflectance data across a complex terrain exhibited physically implausible negative values, predominantly in shadowed areas. Moreover, traditional pixel-by-pixel methods failed to leverage valuable spatial information for estimation. To address these limitations, we developed an integrated atmospheric and topographic correction framework, Unet-TopoFlat, leveraging a spatial-spectral Attention U-Net algorithm and a novel pseudo-topographic synthetic strategy. The pseudo-topographic synthetic strategy generated sufficient and robust topographically incorporated TOA radiance samples across different terrains, surfaces, and atmospheric conditions, using surface reflectance over flat terrains based on radiative transfer models, atmospheric parameters, random forest regression, and a mountainous radiative transfer parameterization scheme. Using Landsat 8 data as a proxy for evaluation, the Unet-TopoFlat was trained on 47,398 samples (256 × 256 pixels), leveraging multiple datasets. The Unet-TopoFlat model effectively captured the spatial and spectral relationships between TOA radiance and surface reflectance, achieving a relative root mean square error (rRMSE) of 4.5 %–6.2 % across 20,314 samples spanning different terrain, temporal, and spectral bands. Compared to the baseline Unet-FLAT model, which lacked topographic consideration and exhibited substantial uncertainties, Unet-TopoFlat effectively reduced topographic effects, lowering negative reflectance ratios from 55.5 % to 2.8 % while accurately recovering surface information and preserving spectral information. Moreover, the leaf area index (LAI) and snow cover mapping using our estimated surface reflectance were superior to those using official products, and deviations reached up to 2.4 for LAI and 8 % for snow cover mapping at the regional scale. Our proposed framework is not sensor-specific and can be potentially applied to multiple optical remotely sensed data.