Cloud Detection Algorithm And System For Multi-spectral Remote Sensing Images

Remote sensing technologies and satellite imagery have been widely used nowadays,which playing a pivotal role in battlefield environmental reconnaissance,weapon precision guidance,geological disaster prediction and daily weather forecasting.However,nearly 2/3 of the earth’s surface is covered by cloud and cloud shadow,the quality of remote sensing images will be affected rapidly.Therefore,cloud pixel detection(abbr.cloud detection)and removing cloud have become an important step of remote sensing image optimization.Due to the irregular shape of clouds and unclear boundaries,the cloud detection problem still faces many challenges.Cloud detection belongs to the semantic segmentation task in the field of computer vision.The deep learning network could explore useful information for multi-channel data during the training phase,which is suitable for the classification scene of remote sensing images.However,some established deep neural networks can not obtain satisfying detection performance.Based on the current status of the existing classical semantic segmentation models and cloud detection technology,we propose a new network,called Ac UNet,combining UNet with atrous convolution and traditional physical methods.The contributions of this work is summarized as follows.(1)Ac UNet replaces the deep pooling operation with atrous convolution so it obtain a large receptive field and reduced information loss.(2)Ac UNet uses padding(Padding)convolution instead of Valid convolution,in order to ensure the network’s two-stage-feature-map be the same size;Ac UNet achieves 92.73% for the overall accuracy on the multispectral remote sensing dataset of Himawari.In addition,it obtains the accuracy of 90.63% on the domestic satellite GF-1 dataset.Ac UNet shows better results compared with other baselines,such as the original UNet network,in the condition of the same parameter amount and computing resources.Our ablation experimental results demonstrate that physical methods for channel fusion can make cloud pixels brighter at the visible light level,furthermore,atrous convolution can further expand the receptive field and enable the model to capture more contextual informationFurthermore,after the design and implementation of the above algorithm model,we developed a remote sensing image data analysis system to integrate the algorithm for actual application.