Research On On-orbit Intelligent Cloud Detection Technology Of Satellite Remote Sensing Image Based On Deep Learning

In recent years,the use of low-orbit small satellite constellation networking has set off a research boom in military and civilian fields at home and abroad.The space technology represented by on-orbit processing technology is the basis of the low-orbit constellation Internet.However,due to limited computing power,power consumption and storage resources,massive real-time remote sensing information cannot be fully utilized,and a large number of cloud-covered remote sensing images have created tremendous pressure on the system’s transmission channel.At the same time,the occlusion of clouds in satellite remote sensing images seriously interferes with the detection and recognition of sensitive targets at high resolution.Aiming at the urgent need of satellite for remote sensing image on-orbit intelligent cloud detection system,considering the complicated process and lack of real-time performance of existing remote sensing image cloud detection algorithm,this paper proposes the application of low-orbit small satellite on-orbit intelligent real-time processing as the application background.We plan to introduce the deep learning algorithm into the cloud detection of satellite remote sensing image,and transplant the remote sensing image cloud detection algorithm model to the high-performance embedded platform to build a remote sensing image on-orbit intelligent cloud detection system.Focus on the following:(1)Research on remote sensing image cloud detection technology based on fully convolutional network(FCN).Aiming at the irregular distribution of cloud targets in remote sensing images,a fully convolutional network is used to realize cloud detection of remote sensing images from the perspective of image semantic segmentation.The method adopts data enhancement methods such as horizontal flipping,rotation and scaling,and adaptive gradient descent optimization algorithm to make the training task of remote sensing image cloud detection model converge quickly and achieve better training results.The feasibility of remote sensing image cloud detection based on fully convolutional network is verified by experimental comparison.(2)Building a cloud detection model acceleration engine based on the inference optimizer TensorRT.Aiming at the high real-time requirements of the spaceborne remote sensing image cloud detection system,a method based on TensorRT for remote sensing image cloud detection model optimization is proposed.The necessity of the optimization of the deep learning model and the advantages of the inference optimizer TensorRT are discussed.The basic flow,custom layer design and optimization measures of the hardware acceleration model of the remote sensing image cloud detection model based on TensorRT are analyzed.Experimental tests show that the remote sensing image cloud detection hardware acceleration engine allows cloud detection accuracy to drop by 0.744%,and the detection speed is 1.23 times higher than that without the inference acceleration.(3)Verification and Realization of Real-time Intelligent Cloud Detection Algorithm for Remote Sensing Images.Aiming at the on-board application requirements of remote sensing image intelligent cloud detection algorithm,a remote sensing image real-time intelligent cloud detection system based on embedded platform Jetson-TX2 was built.The hardware architecture and software system of the embedded deep learning platform Jetson-TX2 are analyzed,and the remote sensing image on-orbit real-time intelligent cloud detection system is designed based on the platform.Experimental tests show that the fully convolutional network remote sensing image cloud detection model trained on the server is transplanted to the embedded platform Jetson-TX2,which can complete remote sensing image cloud detection on a resource-constrained embedded platform,and the average processing time of a single input image is 0.129 seconds,which provides the basis for the on-orbit real-time cloud detection system for subsequent projects.