A Cloud Removal Model for High-Mountain Canyon Terrain Under Weighted Mask Conditions

Cloud cover significantly impedes optical remote sensing research, especially in southeastern Tibet where the average cloud cover reaches 62%, drastically reducing the availability of optical remote sensing images. Compared to temporal cloud removal methods, single-image cloud removal offers more immediate results, though it is more challenging. Existing cloud opacity estimation methods struggle to recover surface information in areas with thick cloud cover, and generative adversarial network approaches often produce artifacts and lack interpretability. To address these issues, this study introduces a weighted mask cloud removal model tailored for the complex alpine climate conditions. Methods: Targeting the stratocumulus phenomenon in remote sensing images, the model establishes assumptions about cloud opacity and brightness compensation. It incorporates an improved Transformer module based on these principles, along with an image restoration strategy, to accurately repair surface pixels obscured by clouds. The model enhances the multi-head attention mechanism by integrating a weighted mask based on cloud opacity and introduces a progressive mask updating strategy. This iterative process begins at the cloud edges, gradually restoring the remote sensing images. The model effectively combines thin and thick cloud removal capabilities, merging the strengths of both in repairing cloud-covered surface information. Results: The model's effectiveness was evaluated under various cloud cover scenarios, comparing its performance with existing models and assessing robustness in alpine regions with snow interference. The results show significant advantages in texture preservation and minimal artifact occurrence when cloud cover is below 20%. The model outperforms comparative methods for cloud cover between 20%-30% and exhibits fewer distortions above 30% cloud cover, generating only minimal artifacts in the densest cloud areas. It maintains stability under snow-covered conditions. The model's accuracy was tested on four metrics, showing MAE=0.025 6, RMSE=0.035 6, PSNR=30.185 1, SSIM=0.899 6, with approximately 61.6 million parameters and a computational complexity of 7.73 GMac. Conclusions: This research innovatively combines cloud opacity estimation and image generation techniques to develop a weighted mask cloud removal model for challenging alpine climate conditions. Utilizing generated cloud opacity and brightness compensation, the model simulates cloud layers and generates weighted masks. It refines the multi-head attention mechanism within the improved Transformer structure and uses weighted masks to regulate neuronal activation, effectively constraining the model's response under complex cloud conditions. A sliding window-based weighted mask updating strategy was introduced to optimize the training process, allowing the weighted mask to gradually diminish and eventually disappear. This study is the first to integrate thin and thick cloud removal strategies, offering a new perspective for the advancement of cloud removal technology and significantly enhancing the usability of optical images in complex alpine climates, thus providing valuable technical support for mountainous scientific research.