Research On Cloud Detection Algorithm Based On Deep Learning For Landsat 8 Remote Sensing Image

Landsat 8 Operational Land Imager(OLI)data is widely used in ecological environment monitoring,owing to effective improvements in bands and spatial resolution.However,due to the influence of cloud coverage,its application in land information extraction is greatly limited.Achieving high-precision cloud detection is one of the key links in data processing of Landsat 8 OLI satellite.Due to the complexity of the surface structure and the diversity of clouds,the existing cloud detection methods are difficult to extract cloud pixels with high accuracy,especially for those covered by thin and fragmented clouds and those over high brightness surface.When the spectral information is limited to the red,green and blue bands,the automatic cloud identification using traditional cloud detection methods is more difficult.Because of its self-learning function and data analysis capabilities,the deep learning method can mine the band information in depth,thereby greatly improving the accuracy of satellite data cloud detection by using limited bands,and thus has been well applied.However,due to the limitations of the number,quality and representativeness of the trained samples,and the differences in the applicability of different networks to different learning sources,the accuracy and stability of the cloud detection using traditional deep learning algorithms still have larger limitations.Aiming at the existing problems in the cloud detection method of deep learning for Landsat 8 OLI,a cloud detection algorithm based on deep learning combined with different network structures was proposed in this paper,and the method of multi-feature fusion was proposed simultaneously to improve the accuracy of cloud detection using deep learning technology.The main research work of this study is as follows:(1)Combined Application of FCN-8s and U-net in Cloud Detection for Landsat 8 OLI with Fully Convolutional Neural Network(FCNEL).The FENCL method combined FCN-8s and U-net model to improve the accuracy of cloud detection based on FCN.The main strategy is to use FCN-8s model and U-net structure to learn Landsat 8 OLI data features respectively.By introducing the ensemble learning after network training and using the voting strategy to integrate the results from FCN-8s and U-net model,the final cloud classification map is obtained.Different surface types and different types of clouds were selected to carry out the detection experiments.The results show that it has good scalability and adaptability to various cloud types and diverse underlying surface environments.Remote sensing interpretation cloud results,cloud detection results from FCN-8s U-net model and Fmask algorithm were used for accuracy verification and comparison.The results show that the overall average accuracy of cloud detection is over 91.68%,and the average accuracy of producers is 98.52%,which is higher than that of the Fmask algorithm and the whole convolution neural network cloud detection results of a single method of FCN-8s or U-net;and the method has good scalability and adaptability to various cloud types and different surface environments.(2)Cloud detection based on deep learning combining muti-feature.Due to the lack of spatial features,the traditional cloud detection methods have low detection accuracy in the complicated surface environment(including bright objects that are confused with cloud,such as white buildings,airplanes,etc.)or thin clouds on bright surface.Based on the cloud detection method combined FCN-8s and U-net model using FCN,considering the multi-feature of cloud,a multi-feature fusion deep learning cloud detection method(MDL)is proposed.To effectively utilize the spectral information and spatial feature of the cloud,the deep neural network(DNN)and the FCN-8s model are used to extract the spectral and spatial features of the image respectively,and these features are re-learned as input to another DNN,and the image is also used as the input of the DNN as information supplement.Finally,the joint features obtained through relearning are classified by Support Vector Machines(SVM).The method fully utilizes the spectral and spatial information of the image to comprehensively detect the cloud.The method chooses to use three visible bands as the basis for cloud detection.The Landsat 8 OLI datasets containing a variety of underlying surface is used for comparative experiments.The results show that the MDL method performs well and it can detect cloud efficiently under the condition of using only three visible bands.Compared to a single neural network algorithm and the Fmask algorithm for multiple input,the cloud detection accuracy,especially for the detection accuracy of thin clouds,has been significantly improved.