Cloud Detection In Remote Sensing Images Based On Multiscale Features-Convolutional Neural Network

In order to meet the needs of earth observation,resource investigation,natural disaster prediction,environmental pollution monitoring and other fields,accurate cloud detection of remote sensing image is of great significance for the application of remote sensing image.However,due to the diversity of cloud types and the complexity of underlying surface,there are still various difficulties in accurate cloud detection of remote sensing images.The existing cloud detection methods either rely on cloud feature calculation and threshold setting,or rely on the location matching of cloud and cloud shadow,or combine feature extraction and simple classification model to achieve the cloud detection task of remote sensing image.However,due to the influence of the thickness of the cloud and the complexity of the underlying surface,the applicability of the existing methods,the missed detection and false detection of the classification model,make it difficult to accurately detect the cloud in the remote sensing image.Therefore,the paper combines the Fmask algorithm and machine learning detection method with the theoretical knowledge of cloud spectral and height features,and takes them as the theoretical basis of remote sensing images cloud detection.Then,the paper uses the single image represented by Landsat-8 remote sensing images and the multi angle images represented by ZY-3 remote sensing images as the basic data source,and uses the spectral information and height information to realize the cloud detection task of different remote sensing images.Experiments show that the multiscale featuresconvolutional neural network(MF-CNN)model can effectively detect cloud areas in a variety of remote sensing images,and improve the detection accuracy of thin cloud while ensuring the detection accuracy of thick cloud.In addition,considering the difficulty of cloud detection caused by the lack of spectral bands in ZY-3 multi angle remote sensing images,the paper combines the height information of cloud,and successfully uses the above method to realize the detection of thin and thick cloud in multi angle remote sensing images.The innovation of this paper includes the following two aspects:(1)In view of the shortcomings of the existing methods,this paper proposes a multiscale features-convolutional neural network model for remote sensing image cloud detection.In this method,the multiscale global features of input data are learned by constructing model,and the high-level semantic information obtained in the feature learning process is combined with the low-level spatial information to realize the thin and thick cloud detection of single remote sensing image.(2)Because the short infrared band and the thermal infrared band can help to identify the cloud areas in the remote sensing images,considering the difficulty of the less spectral band in the ZY-3 high resolution satellite images,this paper combines the height information with the spectral information of the multi angle remote sensing images,and uses the multiscale features-convolutional neural network to realize the application of this model in the thin and thick cloud detection of the multi angle remote sensing images.In order to verify the effectiveness of the multiscale features-convolutional neural network model in cloud detection,firstly,the multi band data of Landsat-8 remote sensing image is used as the data source.The paper explores the detection effect on thick and thin cloud of various commonly cloud detection methods from the qualitative aspect,including support vector machine(SVM),random forest(RF),BP neural network(BPNN),full convolution neural network(FCN)and Fmask algorithm.In addition to qualitative display of test results,the paper first introduces three evaluation indexes,namely,the Precision,Recall and F-Score,to quantitatively evaluate the detection performance of each method on thin and thick cloud,then introduces four indexes,namely,the right rate(RR),error rate(ER),false alarm(FAR),and the ratio of right rate and error rate(RER),to comprehensively evaluate the detection effect of each detection method in the overall cloud detection.Finally,taking ZY-3 multi angle remote sensing image and corresponding area digital elevation model(DEM)as the data source,the paper calculates the corresponding digital height model(DHM)of image area.Through the integration of the four spectral bands data and height feature,the paper combines with multiscale features-convolutional neural network model to realize the thin and thick cloud detection of multi angle remote sensing images,which further verifies the effectiveness of the method.