Bad strip reconstruction method suitable for multispectral infrared remote sensing satellite data

Abstract

The invention relates to the technical field of remote sensing satellite data processing, in particular to a bad strip reconstruction method suitable for multispectral infrared remote sensing satellite data, which comprises the following steps: S1, preprocessing a registered multispectral infrared image, including image segmentation, standardization and dimension conversion; s2, constructing a Trans U-Net model which is fused with a Unet structure and a Transform structure, and performing training in combination with SSIM-Loss and Smooth-L1 Loss; s3, performing weighted fusion on a model prediction result and a quadratic fitting result, and performing standardized smoothing processing on invalid strips; and S4, carrying out compression optimization on the training model and deploying the training model to a satellite platform to realize real-time reconstruction of an invalid region. According to the method, through fusion of deep learning prediction and physical fitting, standardized smoothing processing and model compression deployment, high-precision, lightweight and real-time reconstruction of the invalid stripe region in the multispectral infrared remote sensing image is realized.

Images