Retrieval of Surface Solar Radiation through Implicit Albedo Recovery from Temporal Context

Yael Frischholz, Devis Tuia, Michael LehningarXiv 2025 · DOI · arXiv

Abstract

Accurate retrieval of surface solar radiation (SSR) from satellite imagery is essential for solar energy management but is often hindered by the difficulty of disentangling ground reflectance (albedo) from atmospheric attenuation. We present HeMu, a deep learning emulator based on a Temporo-Spatial Vision Transformer (TSViT) architecture. HeMu leverages temporal context to implicitly recover albedo, eliminating the need for static albedo maps or explicit cloud masks. In the challenging terrain of the Swiss Alps, HeMu matches the accuracy of physics-based models while achieving a 10x computational speedup, demonstrating the power of factorized attention in processing multi-spectral satellite sequences.

Introduction

Measuring the solar energy reaching the Earth's surface is critical for grid stability and site selection for renewable energy projects. Satellites provide high-resolution data, but their sensors measure "top-of-atmosphere" reflectance—a composite of light reflected by clouds and the ground. Traditional SSR retrieval algorithms, such as HelioMont, rely on "background reflectance" statistics to isolate the ground's contribution. However, in regions like the Swiss Alps, rapid changes in snow cover make monthly albedo averages unreliable, and the visual similarity between snow and clouds frequently leads to retrieval errors.

Methodology: The HeMu Architecture

HeMu addresses these challenges by treating SSR retrieval as a sequence-to-image task. It utilizes the Temporo-Spatial Vision Transformer (TSViT), which processes a window of past satellite observations to provide the model with necessary temporal context.

The HeMu architecture consists of four main stages: (a) Tokenization of the input cube, (b) a Temporal encoder to process pixel evolution, (c) a Spatial encoder to capture neighborhood relationships, and (d) a Regression head to produce the final SSR map.

Factorized Attention

As shown in the architecture diagram above, the model employs a factorized approach to attention:

  1. Tokenization: Input multi-spectral imagery (MSG-SEVIRI HRV and IR bands) and topographic data are divided into patches and converted into d-dimensional tokens.
  2. Temporal Encoder: The model first applies attention across the temporal dimension for each spatial location. This allows the model to learn the spectral signature's evolution, effectively creating a "soft memory" of what the ground looks like under varying conditions.
  3. Spatial Encoder: Following the temporal processing, a spatial encoder captures the relationships between neighboring pixels, providing the contextual information necessary to distinguish localized cloud patterns from broader terrain features.
  4. Regression Head: Finally, the processed tokens are reassembled and passed through an MLP to generate the SSR estimate for the target timestamp.

Implicit Albedo Recovery

The core innovation of HeMu is Implicit Albedo Recovery. By providing a temporal window (approximately 40 hours), the model learns to exploit the different "inertias" of ground and atmosphere: ground reflectance (even with snow) changes more slowly than cloud cover. This allows the transformer to internally reconstruct the background reflectance without requiring hand-crafted albedo maps.

Datasets and Evaluation

HeMu was trained to emulate HelioMont, a physics-based algorithm tailored for Alpine regions, using data from Switzerland (2015–2020) at a 1.7km resolution. Validation was performed against 87 ground stations from the SwissMetNet network. Input features included multi-spectral satellite bands, NASADEM elevation data, and solar geometry.

Key Results

  • Performance: HeMu achieved a root-mean-square deviation (RMSD) of 134.0 W/m² for instantaneous measurements, dropping to 22.5 W/m² for monthly aggregates. It successfully reduced systematic overestimation biases common in high-altitude, snow-prone areas.
  • Efficiency: By leveraging GPU acceleration, HeMu delivers a 10x speedup over the CPU-based HelioMont pipeline, processing 10 frames per second.
  • Robustness: Ablation studies confirmed that the model is resilient to missing features; while the High-Resolution Visible (HRV) channel is paramount, the attention mechanism effectively compensates for the removal of individual infrared or topographic inputs.

Conclusion

HeMu demonstrates that temporal context can effectively substitute for explicit physical parameters like albedo in SSR retrieval. While currently an emulator tied to the domain of its training data, the TSViT-based approach offers a scalable, high-speed alternative to traditional physics-based pipelines in complex, dynamic environments. Future work will explore the model's transferability to different geographic regions and climate zones.

Summary updated

BibTeX
@misc{frischholz2025retrieval,
  author = {Yael Frischholz and Devis Tuia and Michael Lehning},
  title = {{Retrieval of Surface Solar Radiation through Implicit Albedo Recovery from Temporal Context}},
  year = {2025},
  eprint = {2506.10174},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.10174},
}
Figure