Cloud Removal from Satellite Imagery using a Lightweight Gated Attention U-Net

Cloud cover in optical satellite images reduces their usefulness for many applications such as land use analysis, vegetation monitoring, and disaster assessment. Traditional methods like cloud masking and image compositing often fail to recover fine spatial details, especially in heavily occluded regions. Recent deep learning approaches, mainly based on U-Net architectures, have shown good performance for cloud removal. However, many of these models struggle to focus on cloud-affected regions while also keeping the computation efficient. This work presents a lightweight Gated Attention U-Net model for satellite image cloud removal. The model is built on a Modified U-Net backbone and uses gating-based attention in the skip connections to highlight important spatial features in cloud-covered areas. Instead of proposing a new attention design, this study systematically evaluates three attention settings, namely No Attention, Light Attention, and Full Gated Attention, under the same training and assessment conditions. This allows a fair analysis of reconstruction accuracy, stability, and computational efficiency. Experiments on the Remote Sensing Image Cloud Removal (RICE1) dataset show that the attention-based configurations achieve stable reconstruction performance across three random seeds. The Full Gated Attention configuration achieves a mean Peak Signal-to-Noise Ratio (PSNR) of 34.76 ± 0.51 dB, while the Light Attention configuration provides the highest mean PSNR (35.21 ± 0.13 dB) and lowest Mean Absolute Error (MAE) (0.0167 ± 0.0004) among the attention variants. Additional assessment on the RICE2 dataset using the same Gated Attention U-Net configuration supports the robustness of the architecture across different cloud patterns and scene conditions. The attention-based configurations maintain low inference times, ranging from 0.0086 to 0.0104 seconds per image across the three-seed experiments. This makes it suitable for low-latency satellite image processing applications and deployment in resource-constrained environments. Overall, the results demonstrate that selective attention can provide a favorable balance between reconstruction quality and computational efficiency within the lightweight Modified U-Net framework.