Coarse-grained label-guided deep learning method for cloud region detection in pixel-level remote sensing images

Abstract

Based on the practical problem that the deep learning-based cloud region detection method needs to cost a lot of labeling, the present invention discloses a pixel-level remote sensing image cloud region detection method based on coarse-grained label-guided deep learning. Image block-level labels, under the constraints of remote sensing image block datasets with coarse-grained image block-level labels, combined with local pooling layer pruning strategies and global convolutional pooling layers, a robust deep network model is trained and cloud generation The activation map is then thresholded to obtain the final cloud mask map. The invention can greatly reduce labeling work, and at the same time realize pixel-level accurate cloud area detection of remote sensing images, and can effectively improve the efficiency and performance of remote sensing image cloud area detection.

Description

technical field [0001] The invention belongs to the technical field of remote sensing and artificial intelligence, and relates to a remote sensing image cloud area detection method based on deep learning, in particular to a pixel-level remote sensing image cloud area detection method guided by coarse-grained labels and deep learning. Background technique [0002] Cloud detection is a key issue in remote sensing image interpretation and application. A large number of cloud cover will affect the availability of remote sensing image data and increase the difficulty of remote sensing image interpretation. Cloud area detection uses various methods in the field of remote sensing and computer vision to detect cloudy areas in remote sensing images. At the level of on-board applications, by not sending images with more clouds, it saves transmission bandwidth and storage space, and reduces resource consumption. Waste; at the level of ground applications, it provides data preparation f...

AI Technical Summary

Problems solved by technology

However, the superior performance of deep learning relies on a large number of accurate pixel-level labels, and the labeling work is quite time-consuming and labor-intensive

Considering that different types of satellites often have great differences in spectrum and spatial resolution, for each remote sensing satellite image, the method based on deep learning requires a corresponding pixel-level annotation data set to operate, which in turn requires a lot of manpower to label

Images