LersGAN: A GAN-Based Model for Low-Light Remote Sensing Image Enhancement

Enhancement of low-light remote sensing imagery captured by drones presents substantial scientific and practical significance. Suboptimal illumination conditions frequently compromise image acquisition quality, particularly under device and environmental constraints. Such degradation not only diminishes visual perception but also impedes the accuracy of computer vision applications. To address these challenges, we propose the LersGAN - an innovative unsupervised enhancement framework integrating cross-attention mechanisms with multiscale discriminators. The architecture incorporates three novel components: 1) A Cross-Enhanced Attention Module combining channel-wise and spatial additive attention for improved feature extraction; 2) Multi-scale discriminators for hierarchical feature evaluation; 3) A composite loss function, which consists of Discrete Cosine Similarity Loss, Fourier High-Frequency Loss, and Enhanced Image Entropy Difference Loss. Experimental validation shows that our model outperforms state-of-the-art methods in complex enhancement scenarios. We conduct ablation studies and comparative analyses to verify these results. The model is applied to UAV remote sensing applications. It reduces persistent noise, chromatic distortion, and illumination deficiencies while preserving textural details. Quantitative metrics and visual assessments confirm its technical advancement and operational superiority in low-light image restoration tasks. Our codes are available at: https://github.com/TianqiLi11/LersGAN.