On the Impact of Non-Stationary Spectral Noise in Satellite Image Restoration: A Spatial-Spectral Attention Benchmark

Abstract

Multispectral satellite imagery from sensors like Sentinel-2 is essential for global earth observation but suffers from complex, non-stationary noise that varies across spatial coordinates and spectral channels. Traditional deep learning architectures frequently process multispectral bands uniformly, failing to capture inter-band dependencies and propagating raw noise through skip connections. In this work, we present a mathematical modeling framework for non-stationary spectral noise and propose the Residual Spatial-Spectral Attention Module (R-SSAM) integrated into a U-Net backbone. Our benchmark demonstrates that while quantitative PSNR performance remains comparable (~24.26 dB), the proposed architecture improves spectral fidelity (reducing SAM from 13.84 degrees to 13.51 degrees) and completely eliminates severe vertical striping artifacts present in standard convolutional baselines. Note: This preprint serves as a foundational Proof of Concept (PoC) and reproducible synthetic benchmark. It establishes the baseline architecture for our ongoing research into spatial-spectral attention mechanisms and State-Space Models (SSMs) applied to real-world L1C/L2A Sentinel-2 datasets. Code Availability: The complete PyTorch implementation and benchmarking notebooks are available at: https://github.com/toliveiradev/paper1-spatial-spectral-denoising