Satellite Image Inpainting With Edge-Conditional Expectation Attention

Satellite images often suffer from data loss and corruption due to various factors, including sensor malfunctions and atmospheric interference, leading to incomplete and degraded imagery. In satellite images, long-range dependencies are particularly significant due to irregular and widely distributed geomorphological edges, such as rivers, mountains, and urban structures. Traditional convolutional neural network (CNN)- based inpainting methods face challenges due to their fixed receptive fields and parameter sharing, limiting their ability to effectively capture long-range dependencies and differentiate between corrupted and uncorrupted areas. To address these limitations, we propose a deep learning approach based on an Edge-Conditional Expectation Attention (ECEA) module, which conditions the attention mechanism on edge information to enhance the model's focus on high-frequency edge details. This enables the network to capture critical structures within the image better. Additionally, we apply Chebyshev's inequality within the attention mechanism to constrain the expectation of attention outputs, reducing excessive deviations and stabilizing the reconstruction process. Experimental results demonstrate that our approach outperforms several state-of-the-art methods in restoring missing regions and reconnecting geomorphological features.