Deep learning-based Sentinel-2 super-resolution via channel attention and high-frequency feature enhancement

Abstract

Introduction: High-resolution satellite imagery is essential for environmental monitoring, land-use assessment, and disaster management, particularly in Southeast Asia—a region marked by ecological diversity, rapid urbanization, and climate vulnerability. However, the limited spatial resolution of several key spectral bands in widely used platforms such as Sentinel-2 constrains fine-scale analysis, especially in resource-limited contexts. Methods: To overcome these limitations, we develop an enhanced deep learning–based super-resolution framework that extends the DSen2 architecture through two dedicated components: a High-Pass Frequency (HPF) enhancement layer designed to better recover fine spatial details, and a Channel Attention (CA) mechanism that adaptively prioritizes the most informative spectral bands. The model is trained and evaluated on a geographically diverse Sentinel-2 dataset covering 30 regions across Vietnam, serving as a representative case study for Southeast Asian landscapes. Results: Quantitative evaluation using Root Mean Square Error (RMSE) shows that the proposed framework consistently outperforms bicubic interpolation and the original DSen2 model. The most substantial improvements are observed in the red-edge and shortwave infrared (SWIR) bands, which are critical for vegetation and land-surface analysis. Discussion: The performance gains achieved by the proposed model translate into more accurate and operationally useful high-resolution imagery for downstream applications, including vegetation health monitoring, water resource assessment, and urban change detection. Overall, the method provides a scalable and computationally efficient approach for enhancing Sentinel-2 data quality, with Vietnam serving as a practical benchmark for broader deployment across Southeast Asia.