Research On Multi-spectral Image Cloud Detection Algorithm Based On Tensor And Real-time Processing In Orbit

Remote sensing image processing technology is widely used in the field of military defense and civil economic construction,benefiting from the rapid development of space remote sensing Earth observation technology.The remote sensing images of today's optical imaging satellites have the advantages of high resolution,wide coverage,and rich details.These advantages also mean that the amount of data being imaged is very large,and the massive data will cause greater data transmission pressure,which is greatly delayed.Clouds usually cover nearly two-thirds of the Earth's area,and the oceans have more clouds than land.The electromagnetic wave that blocks the ground features from the cloud layer reaches the sensor system,which seriously hinders the optical remote sensing satellite from obtaining effective information,resulting in redundancy of the on-board storage resources and increasing the digital transmission pressure.With the rapid development of the integrated chip industry,artificial intelligence,deep learning and real-time recognition based on embedded systems have been implemented in terrestrial applications.An embedded real-time processing camera with artificial intelligence recognition is an important direction for the development of future on-orbit processing cameras.The smart camera can perform on-track real-time processing of data,such as real-time detection of on-board targets,eliminating invalid information such as thick clouds or calm seas,and extracting effective military target information such as ships or aircraft.In a data transmission window,the amount of redundant information is greatly reduced,the data usage efficiency is improved,and the digital transmission pressure is reduced,thereby achieving the following purposes: real-time reconnaissance of large military targets,tracking and monitoring of their activities,and mastering the deployment and target trends of large-scale military forces at sea in sensitive areas.Therefore,realizing on-orbit real-time cloud detection processing has important practical application value in military or commercial satellite fields.In order to improve the speed and accuracy of on-orbit remote sensing image automatic cloud detection processing,ease the pressure of in-orbit transmission and storage data,and ensure the timeliness of remote sensing information,this paper focuses on the nature of multi-spectral remote sensing image and cloud detection.According to the imaging characteristics of remote sensing images,a set of tensor-based on-orbit realtime cloud detection processing methods and an on-orbit processing hardware platform based on programmable logic array(FPGA)and multi-core digital signal processor(DSP)are designed.This paper describes the key technologies involved in the algorithm flow track real-time cloud detection process,including superpixel segmentation,selecting the region feature objectives,tensor-based machine learning method,the highefficiency computing architecture on the hardware platform improves the slow operation speed,low detection rate and poor versatility of the cloud detection algorithm in multi-spectral image remote sensing images in complex background.The main research contents are summarized as follows:Multi-spectral images can be viewed as three-dimensional data composed of height,width,and spectrum.Usually,multi-spectral images are expanded into multiple twodimensional matrices for arithmetic processing.Flattening or unfolding operations destroy the structure between internal data,and the structure acts as a multi-dimensional.The reflection of information in space is very important,and tensors can describe data structures with three dimensions at least.This paper introduces and analyzes the problems of tensor operations and support tensors.Since the weight function supporting tensors is obtained by projection,this method cannot capture the coupling between multi-spectral image data structures.In order to maintain the integrity of the data structure,this paper proposes a support specification tensor training machine(SCTTM).The experimental results show that the algorithm can maintain the tensor data structure information,and obtain the efficient identification classification result under the training set of small samples.In the orbit cloud detection processing,the image segmentation step is first needed.Combined with the advantages of parallel processing on the on-orbit hardware platform and the precise segmentation target and background requirements,this paper analyzes the current mainstream superpixel segmentation methods and their advantages and disadvantages.Since superpixel segmentation is mainly for color(RGB)images,when superpixel segmentation is used for multispectral remote sensing images,spectral segment information and rich grayscale information of multispectral images cannot be fully utilized.In this paper,a superpixel segmentation method based on feature weighted wavelet fusion is proposed,which makes full use of all spectral segment information and grayscale information of multispectral.Experiments show that the proposed method can accurately segment oceans and land,ships and oceans and clouds,and superpixel segmentation of complex multi-spectral images compared with today's super-pixel segmentation methods.After the super-pixel segmentation based on feature-weighted wavelet fusion is proposed,the remote sensing image is clustered into several super-pixel sub-slices according to the scene.In this paper,a tensor-based Gabor energy cloud detection method is designed,and a new feature is highlighted.The Gabor texture energy feature map combined with the spectral information feature map constitutes a third-order tensor,which is brought into the training tensor training machine for training recognition.The method is not limited to the imaging parameters of the image,and the cloud detection can be performed on multi-spectral images with different orbital heights,different resolutions and different correction levels,and the obtained result has higher accuracy than the existing cloud detection method and lower false positive rate.According to the requirements of on-orbit processing and the tensor operation of the algorithm flow in this paper,this paper proposes a fast tensor computing architecture based on FPGA,and designs an FPGA-based tensor inner product,outer product and core tensor computing architecture.Experiments show that the proposed method has obvious advantages in computing efficiency on FPGA compared with CPU.When the data size is increased,it has obvious advantages over GPU,which provides guarantee for on-orbit real-time processing in subsequent projects.The on-orbit multi-spectral camera imaging module generates a very impressive data rate transmission for the on-orbit processing module,and designs an FPGA-based ultrahigh speed Cameralink interface under the project requirements of the Cameralink interface.Ultra-high-speed Cameralink transmission is realized by using only the onchip resources of the imaging module and the on-orbit processing module,which not only gets rid of the dependence on the Cameralink dedicated conversion chip,but also significantly breaks the Cameralink transmission of the dedicated conversion chip under the condition of the same data format.The rate bottleneck greatly reduces the wiring board wiring requirements and area,and reduces the hardware design cost.Finally,this paper builds an FPGA and multi-core DSP on-orbit processing platform,and optimizes the cloud detection algorithm and the embedded system.While ensuring the versatility and scalability of the system,the processing efficiency of the algorithm is improved,and the transmission pressure of the remote sensing data on the star is reduced,which lays a foundation for the on-orbit target detection.