HybridATNet: Multi-Scale Attention and Hybrid Feature Refinement Network for Remote Sensing Image Super-Resolution

Remote sensing image super-resolution faces significant challenges, including detail loss, sensor noise, and atmospheric interference that compromise image quality for critical applications. Current deep learning approaches struggle with three key limitations: insufficient multi-scale feature extraction, inadequate attention to crucial image regions, and poor preservation of refined spatial details during reconstruction. We propose HybridATNet, a novel CNN-Transformer architecture that addresses these limitations through three primary innovations. First, our shallow feature extraction integrates Efficient Channel Attention (ECA), Multi-scale Spatial Attention (MSA), and Residual Atrous Spatial Pyramid Pooling (RASPP) to capture features across multiple scales in both channel and spatial domains. Second, we employ an Activated Sparsely Sub-Pixel Transformer (ASSPT) for deep feature extraction that efficiently captures long-range dependencies while maintaining computational efficiency through a novel sparse attention mechanism that reduces computational complexity from O(n2) to approximately O(n log n). Third, we introduce a dual-domain approach that combines multi-scale spatial refinement with frequency-domain processing to preserve structural details while simultaneously reducing high-frequency noise. Extensive experiments on the UCMerced and AID datasets demonstrate that our model consistently outperforms state-of-the-art methods across multiple scale factors (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times 2$ </tex-math></inline-formula>, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times 3$ </tex-math></inline-formula>, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times 4$ </tex-math></inline-formula>), achieving PSNR improvements of up to 0.26dB on the UCMerced dataset and 0.13dB on the larger AID dataset. Visual comparisons confirm superior preservation of fine details and improved noise reduction, particularly in complex structural regions. HybridATNet delivers high-quality reconstructions while maintaining computational efficiency, requiring approximately 15% fewer parameters and 18% fewer floating-point operations (FLOPs) than other transformer-based models, making it well-suited for practical remote sensing applications.