Remote sensing image cloud and shadow restoration method based on gated convolution

Abstract

The invention relates to a cloud and shadow restoration method based on gated convolution in combination with multi-temporal data. The method comprises the following steps: constructing a sample library by utilizing existing remote sensing images; secondly, designing a multi-temporal cloud and shadow restoration network based on gated convolution for training, and deeply mining the correlation andfeatures between a cloudy remote sensing image multi-temporal cloudless image and a pure pixel region containing missing information; then, for a real cloud-containing image, using an existing clouddetection method firstly for carrying out pixel-level semantic segmentation, and selecting a result with a high recall rate; and finally, using the trained cloud and shadow restoration network, a highrecall rate detection result, a real cloud-containing image and corresponding multi-temporal data to perform restoration work, and rebuilding missing information with high fidelity on the basis of ensuring spectral information. The invention has the advantages that the robustness is high, and the method can adapt to detection and restoration of remote sensing image clouds and shadows of differentscales.

Description

technical field [0001] The invention relates to a method for cloud and shadow restoration based on gated convolution, which can be used for reconstruction of missing information such as clouds and shadows in combination with multi-temporal cloud-free remote sensing images, and can effectively perform spatial dimensioning for damaged remote sensing images. and spectral dimension information reconstruction, improving the utilization rate of remote sensing images obtained by sensors under different weather conditions. Background technique [0002] As one of the most important earth observation techniques to obtain geometric and physical information of the earth's surface, remote sensing technology has received more and more attention and has been widely used. However, due to the constraints of atmospheric conditions, the daily global average cloud coverage reaches about 67%, and the invalid pixels with low or even no use value that are blocked by clouds and cloud shadows in rem...

AI Technical Summary

Problems solved by technology

[0005] However, although many methods of cloud removal through deep learning have appeared recently, there are still many shortcomings in some aspects: 1) All the methods of cloud restoration based on deep learning take into account the differences between cloudy areas and The distinction of cloud-free areas, but in feature extraction, the use of ordinary convolutional layers cannot distinguish the features of different areas of missing areas; 2) A single loss function is very commonly used in semantic segmentation and object detection, but in cloud restoration , it is necessary to constrain the abstract features (spectral features), and its ability is insufficient; at the same time, in the traditional cloud repair method, there is no distinction between the loss of the reconstruction area and the cloud-free area, and the texture consistency between the reconstruction area and the neighborhood is lacking. sexual restraint

Images