Deep Learning Based Cloud Detection Of Remote Sensing Images

Cloud detection in remote sensing image processing is a key step before other remote sensing image processing.It is important for weather forecast and military affairs.The current traditional cloud detection algorithm is a point-to-point detection algorithm based on the physical model.This kind of algorithm relies on the radiation transmission model.And a large number of the artificial thresholds are used in the cloud detection.The selection of the thresholds has a great influence on the accuracy of cloud detection.The traditional detection algorithm which is a point-to-point algorithm can only get channel features of the single pixel instead of spatial texture features.Therefore,we atempted to develop an algorithm which can automatically extract the features and detect the cloud in remote sensing images.Deep learning is a major achievement in the field of artificial intelligence in recent years and has achieved remarkable results in the fields of image,audio and text.Convolution neural network is a widly used deep learning model.It has a good ability to express two-dimensional images and it can automatically learn more complex non-linear representation to represent an essential features of samples.Therefore,we use convolution neural network to achieve the cloud detection of satellite remote sensing images.It is expected to improve the performance of cloud detection.In the processing of developing the cloud detection algorithm,we build the training dataset and the testing dataset by Himawari 8 satellite remote sensing data.We designed four different convolutional neural networks(FCN original network Cloud＿FCN 、 the modified FCN network Cloud＿M-FCN、Res Net network Cloud＿Resnet、the simplified Res Net network Cloud＿S-Resnet)for cloud detection of satellite remote sensing images based on principle of convolutional feature and multi-scale feature.The convolutional layers,activation layers,max pooling layers and batchnorm layers are used in networks to extract the features of remote sensing images.Then the feature maps of different scales are upsampled and fused.Finally,the networks achieve the pixel-level segmentation.Cloud＿FCN achieves the cloud detection segmentation by a previous study which is a network for segmenting common images.Cloud＿M-FCN adds more levels of scale features based on the Cloud＿FCN.Cloud＿Resnet replaces a convolutional layer with the more complex features extraction unit to increase the features extraction capabilities.Cloud＿S-Resnet is a simplification of Cloud＿Resnet.It deletes some layers to reduce the receptive field of the convolutional neural network.But Cloud＿S-Resnet achieves the same cloud detection capabilities as Cloud＿Resnet.In order to evaluate the performance of the proposed algorithm,we experiment on Caffe.Finally,we get a high detection accuracy rate and the detection efficiency.In the stage of result analysis,comparing with the results of different networks,it is concluded that adding local scale feature information,reducing receptive field,eliminating interference classes conducive to convolutional neural network for satellite remote sensing cloud detection.