Remote Sensing Images Cloud Detection Based On Multi-Scale Deep Convolution Neural Network

Cloud detection is the premise of accurate application of optical satellite image,and it is an important preprocessing step of cloud removal.There are many research fields of processing satellite images.With the help of satellite images,we can observe the natural environment monitoring on the ground（pollution,forest or river）,improve the traffic conditions and respond to disasters in time.The usefulness of satellite images is largely limited by cloud pollution.The blurring caused by cloud makes it unstable to monitor the situation on the ground with a visible light camera.Therefore,the existence of cloud and cloud shadow affect the availability of satellite data.Before the accurate application of satellite images,cloud and cloud shadow detection is essential.It is very important to extract cloud from cloud image accurately for reducing the negative impact of cloud on image application.According to the thickness of clouds,the existing cloud removal methods are divided into thin cloud removal,thick cloud and cloud shadow removal.Different cloud removal methods require accurate detection of cloud thickness in the cloud detection phase.The multi-scale characteristics reflected by clouds and different positions of clouds are challenges to accurately detect cloud thickness.Thanks to the extraction of cloud features,the recently developed deep learning method has achieved great success in cloud detection.However,most of the existing methods cannot make full use of the multi-scale information contained in the remote sensing image,which makes imcomplete feature extraction and insufficient fusion of detailed and coarser features.Therefore,they often lead to overdetection or under-detection,unable to detect the specific thickness of clouds.In addition,the performance of the existing methods in detecting the cloud boundary is not satisfactory.In this paper,we fully explore the multi-scale features of clouds in remote sensing image,and propose a method of cloud detection based on multi-scale deep convolution neural network（MSDCN）.The network is used to detect clouds from different sensors’ remote sensing images and output the region together with the thickness of clouds.In order to make feature extraction more sufficient,we enrich and enhance the input data.By using multi-spectral image data（all bands of remote sensing images）and multi-temporal data（6different time periods）,the method greatly reduces the difficulty of data acquisition and processing.Experiments on a large number of data show that the algorithm is effective in cloud detection and the detection is accurate in terms of both coarse structure and fine details.Traditional methods can detect most of the clouds in satellite images,but it is easy to mistake the bright surface for a cloud.Compared with it,depth model can deal with the bright surface more effectively.On this basis,the cloud detection accuracy of MSDCN has been significantly improved,and the detection results on bright surfaces such as bright water surface and bright snow surface are better.And MSDCN can accurately detect small cloud boundary areas,and will not ignore small cloud areas,resulting in missed detection.MSDCN also shows good performance in cloud details and boundary detection.MSDCN can detect more accurate cloud mask for two reasons.On the one hand,MSDCN does better in the fusion feature level,combining shallow information with multiscale information.On the other hand,the fusion of multi-scale convolution features makes MSDCN have better target feature extraction ability.Therefore,it can better distinguish between cloud layer and clear and cloudless area.We also added residual learning module to MSDCN.On the premise of not losing the detail information,residual learning makes the model training have better optimal convergence,so as to improve the accuracy of the model.In addition,we train and validate MSDCN on two satellite datasets with different attributes.The training data are very diverse,including different time periods,different surface features and different spectral information.These training data with complex characteristics make MSDCN have a strong ability to deal with complex surfaces.In the final stage of the experiment,we analyzed and discussed the multi-scale,multi-spectral and multi-temporal information of MSDCN,and compared the result of cloud removal experiments.The results of these supplementary experiments further illustrate the advantages of the MSDCN method.