Attention-Enhanced Feature Refinement Framework for Super-Resolution of Lunar DEMs

Abstract

Lunar Digital Elevation Model images are essential for analyzing geological features, yet their low resolution limits the detection of sub-kilometer craters. To address this, we propose FREDSR-Net, a lightweight network integrating High-Order Attention for Digital Elevation Model super-resolution. Unlike existing methods, FREDSR-Net employs cascaded Feature Refinement Residual Blocks, where a residual branch extracts multi-scale terrain features, and a refinement branch applies High-Order Attention to amplify high-frequency details through adaptive statistical modeling. Evaluated on Lunar Reconnaissance Orbiter datasets, FREDSR-Net achieves a peak signal-to-noise ratio of 46.72 decibels and a structural similarity index of 0.9909, outperforming both neural models. Ablation studies confirm the necessity of High-Order Attention, with K=3 achieving a 0.21 dB gain over K=1. The model generalizes to unseen lunar regions, improving crater detection accuracy by 12% in YOLOv7-based pipelines. Requiring only 1.01M parameters, FREDSR-Net enables real-time DEM enhancement on low-cost GPUs, advancing high-resolution lunar topographic analysis for future missions.