Deep Learning-Based Sentinel-2 Super-Resolution via Channel Attention and High-Frequency Feature Enhancement

Khang Vi NguyenAnh Duc VuBao Quoc BuiNidal Kamel^*

• VinUniversity (VinUni), Hanoi, Vietnam

Abstract. High-resolution satellite imagery plays a critical role in environmental monitoring, land-use assessment, and disaster management, particularly across Southeast Asia—a region characterized by ecological diversity, rapid urbanization, and climate vulnerability. However, the limited spatial resolution of key spectral bands in widely accessible platforms such as Sentinel-2 restricts fine-scale analysis, especially in resource-constrained settings. To address this limitation, we propose an enhanced deep learning–based super-resolution framework that extends the DSen2 architecture with two targeted modules: a High-Pass Frequency (HPF) enhancement layer to recover fine spatial details, and a Channel Attention (CA) mechanism to adaptively emphasize the most informative spectral bands. The model is trained and evaluated on a geographically diverse Sentinel-2 dataset encompassing 30 regions across Vietnam, providing a representative case study for Southeast Asian landscapes. Quantitative evaluations using Root Mean Square Error (RMSE) demonstrate that the proposed framework consistently outperforms both bicubic interpolation and the original DSen2 baseline, with notable improvements in red-edge and shortwave infrared (SWIR) bands. These gains translate into more accurate and actionable high-resolution imagery for downstream applications such as vegetation health monitoring, water resource mapping, and urban change detection. Overall, the proposed approach offers a scalable and computationally efficient solution for enhancing Sentinel-2 data quality, with Vietnam serving as a benchmark for broader deployment across Southeast Asia.