Cloud detection algorithm of remote sensing image based on DenseNet and attention mechanism

The cloud detection of remote sensing images is the first step in the process of remote sensing image processing. To address the problem that the traditional cloud detection algorithm has a poor effect on the detection of small and thin clouds, this paper proposes a cloud detection method for densely connected network remote sensing images based on the attention mechanism. First, cloud vectors are manually checked from the images provided by the Land Satellite Remote Sensing Application Center of the Ministry of Natural Resources and cloud labels are made, and the images are preprocessed by sequential clipping, color jitter, rotation, etc. to enlarge the sample size. Then, the pre-processed remote sensing images and their labels are fed into a neural network with DenseNet as the encoder and decoder, and a cascaded atrous convolution module is added between the encoder and decoder to increase the receptive field, and a dual attention mechanism and a global context modeling module are added to suppress some irrelevant detailed information. Finally, the experimental results showed that the accuracy rate could reach 95% and the intersection over union could reach 91%, which are big improvements over the traditional cloud detection algorithm, and this method performs well in extracting small and thin clouds.