Neural Network Models for Improving Satellite Image Quality Based on Pixel-Level Domain Adaptation
Alexander B. Murynin, A. K. Shved, В. А. Козуб · Pattern Recognition and Image Analysis · 2026
Abstract The relevance of this work stems from the problem of domain divergence: images of the same area acquired by different satellite sensors exhibit different characteristics (spatial resolution, noise, contrast, brightness distribution), causing models trained on data from one sensor to lose effectiveness when applied to another. This paper considers the task of jointly improving the quality and spatial resolution of images from a source satellite sensor to a target level, which reduces the domain shift between data from various Earth remote sensing (ERS) space systems and simplifies subsequent processing unification. We propose a neural network domain adaptation algorithm that integrates style transfer and superresolution within a single style-guided RDN-FiLM architecture. The architecture is based on a modified residual dense network (RDN) with feature modulation using the feature-wise linear modulation (FiLM) scheme, and is trained in an unpaired setting on unpaired image sets from two sensors: style alignment is performed on real unpaired scenes, while synthetically degraded pairs serve as an auxiliary source for pretraining and for monitoring quantitative quality metrics. This approach combines the rigor of reference-based metrics on synthetic data with realistic domain diversity on real data. Experimental evaluations on real and synthetic data show that the proposed method outperforms the basic superresolution model without domain adaptation: it more effectively reduces the domain shift between sensors, enhances the sharpness and contrast of fine details, improves the visual consistency of textures, and yields a more uniform stylistic representation of scenes, beneficial for subsequent automatic segmentation, detection, and analysis of satellite imagery. Future development directions include further reduction of artifacts in complex heterogeneous scenes, utilizing the panchromatic channel as a high-frequency guide, and integrating more sophisticated generative components while preserving the computational efficiency of the algorithm.