Dr. Yichuan Ma’s recent publication on Remote Sensing of Environment

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

Congratulations to Dr. Yichuan Ma, Professor Shunlin Liang, and the research team for their recent publication in Remote Sensing of Environment (Vol. 334, 115188). The paper, titled "An integrated atmospheric-topographic correction framework for land surface reflectance estimation using a spatial-spectral Attention U-Net model," is available via https://doi.org/10.1016/j.rse.2025.115188. The code for the framework is open-source and accessible at https://zenodo.org/records/17846183.

High-level satellite products are essential for monitoring global carbon and water cycles, yet their accuracy heavily depends on the quality of surface reflectance data. In mountainous regions, complex terrain introduces severe distortions, such as deep shadows and intense multiple scattering. Traditional methods typically perform atmospheric and topographic corrections separately, ignoring the coupled interactions between the atmosphere and terrain. Consequently, standard products (e.g., Landsat 8) often exhibit physically implausible negative reflectance values—affecting over 50% of pixels in winter mountain scenes—rendering them unusable for many applications.

To address this critical gap, the team developed Unet-TopoFlat, an integrated correction framework powered by deep learning. The core innovation lies in a "pseudo-topographic synthetic strategy." Since ground-truth reflectance in mountains is scarce, the team synthesized a massive training dataset by combining high-quality reflectance from flat terrains with Digital Elevation Model (DEM) data to simulate topographically distorted Top-of-Atmosphere (TOA) radiance. A spatial-spectral Attention U-Net model was then trained to directly retrieve accurate surface reflectance from TOA observations, effectively learning to disentangle atmospheric and topographic effects simultaneously.

The framework was rigorously validated using Landsat 8 data across the mountainous western United States. The Unet-TopoFlat model demonstrated exceptional performance, achieving a relative root mean square error (rRMSE) of 4.5%–6.2% across spectral bands. Most notably, it reduced the ratio of negative reflectance values in rugged terrain from a staggering 55.5% (in original products) to just 2.8%, effectively recovering valuable surface information in shadowed areas while mitigating saturation in sunlit snow-covered regions.

The practical impact of this improvement is significant. In case studies involving downstream applications, the new surface reflectance data reduced deviations in Leaf Area Index (LAI) estimation by up to 2.4 and improved snow cover mapping overall accuracy by up to 10% compared to standard products. The framework is physically based and sensor-independent, offering a robust solution for generating high-quality surface reflectance data over complex terrains globally.

Figure 1: The workflow of the Unet-TopoFlat framework, illustrating the pseudo-topographic synthetic strategy and the deep learning-based estimation process. 

Figure 2: Comparison of estimated surface reflectance over complex terrains. The proposed method (bottom rows) successfully recovers information in shadowed areas where the original Landsat 8 product (middle rows) fails and produces negative values. 

References:

Ma, Y., Liang, S., Ma, H., He, T., Shi, X., Li, W., ... & Zhang, Y. (2026). An integrated atmospheric-topographic correction framework for land surface reflectance estimation using a spatial-spectral Attention U-Net model. Remote Sensing of Environment, 334, 115188. https://doi.org/10.1016/j.rse.2025.115188

Keywords:

Surface reflectance, Topographic correction, Deep learning, Attention U-Net, Mountainous remote sensing