Target intelligent identification method using remote sensing information processing system

Abstract

The invention discloses a target intelligent identification method using a remote sensing information processing system. The method comprises the following steps: step 1, finely extracting hyperspectral ground feature elements based on mixed pixel decomposition and spectrum purification; 2, judging the working state of the operation target based on target thermal radiation modeling and multi-temporal analysis; step 3, mining massive SAR data features based on the sparse deep belief network; 4, ship target trajectory fusion based on multi-source information; and 5, carrying out software plug-in architecture and operation unit multi-source load data interpretation. Through integrated construction of the system, data organization management, multi-source data interpretation, data comprehensive reorganization and business management are realized, and the accuracy and production efficiency of remote sensing data of an operation unit are greatly improved.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing surveying, and in particular to a target intelligent recognition method using a remote sensing information processing system. Background Art

[0002] With the rapid development of high-altitude survey technology, the sensor resources available to operating units are becoming increasingly abundant. The massive data provided by the sensor series will greatly enhance the operating unit's ability to survey a wide range of marine ship targets, and will inevitably make high-altitude survey an important means to improve the information-based combat capabilities of operating units. Driven by the visible light, infrared, and hyperspectral data application capabilities of a certain sensor, the results of the construction of the main operating environment sensor series are integrated, and combined with the comprehensive needs of operating units for sensor information applications, the data processing capabilities of ocean observation sensors are improved to solve the problems of insufficient processing capabilities of massive remote sensing data and weak data organization and management capabilities of operating units. Modern high-performance computing technology, massive data storage and management technology, and parallel remote sensing image processing algorithms and software are fully utilized to build a scientific, reasonable, and seamlessly scalable system architecture to achieve a remote sensing information processing platform that integrates flexible and effective data storage management and high-performance computing.

[0003] Based on the experience and foundation of high-altitude survey and observation research and development in the main operating environment, according to the characteristics of marine remote sensing data and the requirements of the operating environment, the processing and fusion of optical, SAR, infrared, hyperspectral and other image data and survey data are realized to make full use of domestic, civilian and foreign commercial sensor resources, form a comprehensive guarantee capability for nearshore and offshore targets based on sensor information, and meet the operating units' requirements for high timeliness and high precision of remote sensing data in the new operating environment.

[0004] In recent years, with the gradual maturity of my country's remote sensing sensor network, remote sensing sensors have accumulated massive amounts of remote sensing data in hot spots. At the same time, other remote sensing data sources such as commercial remote sensing sensors have also accumulated considerable amounts of data. The existing remote sensing data processing systems are lacking in the ability to process massive amounts of remote sensing data and lack effective data organization and management models. The information mining capabilities of multi-source heterogeneous data still need to be improved, which has become a bottleneck problem currently faced by operating units in remote sensing information processing and cannot meet the needs of combat data extraction in the operating environment under the new situation. Therefore, construction is needed in the following aspects.

[0005] 1. According to the characteristics of image data, survey data and multimedia data, it provides two functions: multi-source data fusion and comprehensive display application, and provides comprehensive display for multi-source high-altitude remote sensing image data and data;

[0006] 2. Manage the visible light, SAR, infrared, hyperspectral data and technical survey text data connected to the system, and realize the organization and management of multi-source massive remote sensing data materials and data products;

[0007] 3. Realize the information management of data production business, complete the statistical analysis of received and generated data, monitor and count the operation tasks, generate statistical reports, etc.;

[0008] 4. In view of the characteristics of hyperspectral, thermal infrared, visible light and SAR image data, it provides three functions of remote sensing image processing, image data interpretation and remote sensing data compilation, providing technical support for the transformation of images into data products. Summary of the invention

[0009] The purpose of the present invention is to provide a target intelligent recognition method using a remote sensing information processing system to solve the problems existing in the prior art.

[0010] To achieve the above object, the present invention provides a method for intelligently identifying a target using a remote sensing information processing system, wherein the remote sensing information processing system comprises:

[0011] (1) Multi-source data fusion software; (2) Data organization and management software; (3) Business information management software; (4) Image comprehensive judgment software;

[0012] System external interface: the remote sensing information processing system external relations include interfaces with data distribution system and image text system;

[0013] System internal interface, the remote sensing information processing system internal interface is the interface between multi-source data fusion software, data organization management software, business information management software and image comprehensive judgment software;

[0014] According to the business demand analysis, the remote sensing information processing system includes two major processes: fixed target observation and marine ship target observation;

[0015] The target intelligent recognition method using the remote sensing information processing system comprises:

[0016] Step 1: Refined extraction of hyperspectral features based on mixed pixel decomposition and spectral purification;

[0017] Step 2: Determine the working status of the operation target based on target thermal radiation modeling and multi-temporal analysis;

[0018] Step 3: Feature mining of massive SAR data based on sparse deep belief network;

[0019] Step 4: Fusion of ship target trajectory based on multi-source information;

[0020] Step 5: Software plug-in architecture and multi-source load data interpretation of operating units.

[0021] Furthermore, the multi-source data fusion software provides support for the comprehensive display of multi-source high-altitude remote sensing image data and data provided by image data, survey data and multimedia data, and is composed of data information fusion and comprehensive display application components; the data organization and management software manages the visible light, SAR, infrared, hyperspectral data and technical survey text data connected to the system, realizes the organization and management of multi-source massive remote sensing data materials and data products, and is composed of four components: basic data management, target characteristic data management, data material and data product management, and system management and maintenance; the business information management software realizes the informatization management of data production business, completes the statistical analysis of data, the monitoring and statistics of operation tasks, and the generation of statistical reports, and is composed of three parts: task management component, data statistics component, and user management; the image comprehensive judgment software processes hyperspectral, thermal infrared, visible light and SAR image data, and is composed of three components: remote sensing image processing, image data interpretation, and remote sensing data compilation.

[0022] Furthermore, the fixed target observation process is aimed at operating units and ports, using the second-generation distribution data packet as input to determine the number and type of ship targets in the port; the marine ship target observation process is aimed at marine ship targets, using sensor visible light wide-band data and ScanSAR data, combined with survey data, distribution data, text data, and integrated historical data database and knowledge base data to determine large marine ship targets.

[0023] Furthermore, the step 1 further includes: (1) fast extraction of end members based on the FGDA algorithm, fast extraction of end members based on the end member extraction method of the Cramer determinant, and then using a linear mixed model to decompose the mixed pixels into components, and finally in the abundance composition obtained by the decomposition, the main abundance is preferred and combined with the matched filtering method to perform fine extraction and analysis of battlefield elements; (2) spectral feature extraction based on HOSVD, using the correlation between data for hyperspectral data, in order to reduce the dimension of the data while taking into account the efficiency of calculation, constructing a feature extraction based on high-order singular value decomposition.

[0024] Furthermore, step 2 further includes: (1) imaging condition division, establishing feature models under multiple imaging conditions respectively, wherein the imaging conditions for establishing the feature models include imaging time, longitude and latitude, spatial resolution, temperature resolution, target orientation, target state, and imaging band; (2) target infrared feature calculation, wherein the features used for feature extraction include infrared radiation mean and standard deviation, infrared radiation mean difference between the target and the surrounding background, and statistical quantity of mean contrast to analyze the overall infrared radiation characteristics of the target; (3) feature model generation and update, after feature extraction of all samples under specific imaging conditions, statistical analysis is performed on them, and the mean and standard deviation are calculated. The mean and standard deviation are used to represent the feature distribution under the condition, and the mean and standard deviation are used as the feature model under the imaging condition.

[0025] Furthermore, step 3 further includes: (1) feature learning based on hybrid structure sparse constraints; (2) fine-tuning of model parameters, by training multiple restricted Boltzmann machines layer by layer to obtain high-level abstract features, and then using the back propagation algorithm and conjugate gradient descent method to fine-tune the parameters of the entire network; (3) unsupervised feature extraction, using a batch training method to randomly and evenly divide the training image into multiple data subsets, each image belongs to only one of the subsets, and using an unsupervised pre-training method to train four restricted Boltzmann machines layer by layer, wherein the output of the hidden layer of the previous Boltzmann machine is used as the input of the visible layer of the next Boltzmann machine. After the pre-training is completed, a 5-layer neural network is initialized using the weights and other parameters of the four restricted Boltzmann machines.

[0026] Furthermore, step 4 further includes: (1) fast association based on index, a method for retrieving radar targets based on R*-tree index theory is proposed, the attribute characteristics of the radar target are mapped to a point in the multidimensional space, and then the index-RR-Tree is constructed according to MBR; (2) target information fusion, first position fusion, and then parameter fusion; (3) trajectory compression based on curvature segmentation dynamic threshold; (4) trajectory similarity calculation based on target prior information, similarity matching of position and trajectory is performed based on target prior information, and the confidence probability of each data source is calculated to reduce the impact of noise and optimize the solution efficiency.

[0027] Further, the index is defined as follows: ①. Range query: given a certain error range ε, a query point q is given so that the error between the query result r and q is less than ε; ②. KNN query: given a query point q, query the N points closest to q; ③. Radar target vector: the location information and technical feature parameter data of the radar target are formed into a multidimensional array, called the radar target vector; ④. Radar target vector space: the d items of the radar target vector are used to form a d-dimensional space for description; ⑤. Radar target space distance; ⑥. MBR: the minimum bounding rectangle containing n radar vector points.

[0028] Furthermore, step 5 further includes: (1) a data production process framework, which unifies the data production process framework, unifies data standards, operation modes and data output, standardizes the processing flow, and optimizes the working mode and processing links; (2) a "microkernel + plug-in" software system architecture, in which the microkernel serves as the core layer of the software to realize the most basic data reading, algorithm scheduling, and function attachment; the application plug-in realizes the functions of dedicated interpretation and analysis, geographic spatial data presentation, target query retrieval and proprietary data compilation and output, and the software modules are flexibly split and customized according to usage needs to achieve scalability in data types, scalability in functional tools and scalability in equipment scale.

[0029] Furthermore, the processing logic steps of the process framework are as follows: ①. Task initiation: automatically access the second-generation distribution data, decompress the image compression package, and obtain the image file; ②. Image information extraction: read the image and obtain relevant information of the image, including the image payload, imaging time, target position, etc., store the information in the database, and create a target automatic extraction interface file; ③. Automatic target detection and recognition: automatically extract the target from the image, and write the slice or detection result file into the task file; ④. Target manual identification confirmation and association: the user displays the image, and loads the automatic detection result file to assist in the interpretation, and at the same time uses the target knowledge base to realize the interpretation and extraction of the target information; ④. Target manual identification confirmation and association: the user displays the image, and loads the automatic detection result file to assist in the interpretation, and at the same time uses the target knowledge base to realize the interpretation and extraction of the target information.

[0030] The method of the present invention has the following advantages:

[0031] The present invention forms a ground data processing system for operating units that is compatible with sensor resources, meets the urgent needs of infrared and hyperspectral sensor application processing, and greatly improves the multi-dimensional feature extraction and analysis and recognition capabilities of targets. Through system integration construction, data organization management, multi-source data interpretation, data comprehensive compilation and business management are realized, greatly improving the accuracy and production efficiency of remote sensing data of operating units.

[0032] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the following specifically cites a preferred embodiment and describes it in detail with the accompanying drawings as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the fixed target observation process;

[0034] Figure 2 This is the internal interface diagram of the system;

[0035] Figure 3 This is a schematic diagram of the comparison of the computational efficiency of FGDA end member extraction;

[0036] Figure 4 Schematic diagram of the first 10 bands obtained by HOSVD processing;

[0037] Figure 5 Schematic diagram of feature information obtained by HOSVD processing;

[0038] Figure 6 Schematic diagram of target thermal radiation modeling analysis;

[0039] Figure 7 This is a diagram of the target feature learning system based on deep belief network;

[0040] Figure 8 Schematic diagram for pre-training and model parameter initialization;

[0041] Fig. 9 It is a schematic diagram of position and error ellipse;

[0042] Fig.10 Schematic diagram of trajectory data compression (compression ratio on the right);

[0043] Fig.11 This is the software architecture diagram of "microkernel + plug-in". DETAILED DESCRIPTION

[0044] The technical scheme of the present invention will be clearly and completely described below in conjunction with specific embodiments, but those skilled in the art should understand that the embodiments described below are only used to illustrate the present invention and should not be regarded as limiting the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Based on the design principles, the present invention establishes a processing platform for processing, interpreting and fusion identification of multi-source remote sensing survey images such as visible light, infrared, hyperspectral, SAR, and various types of information, realizing data organization management, multi-source data interpretation, data comprehensive compilation and business process management, so as to enhance the comprehensive guarantee capability of multi-source remote sensing data fusion processing and data product production of the operating unit.

[0046] A remote sensing information processing system, the system comprises:

[0047] Aiming at the urgent need of sensor survey data security for operating units, solving the contradiction between the rapidly growing data volume and the bottleneck of data storage, management and information mining technology, comprehensively utilizing various existing sensor remote sensing data and survey data, building a "remote sensing information processing system", realizing the unified and organic management of massive remote sensing data and other operating environment data, multi-dimensional and multi-mode presentation of multi-source data, and marine operating environment data decoration, distribution and application services based on sensor survey information. The system consists of four software: multi-source data fusion software, data organization management software, business information management software, and image comprehensive judgment software.

[0048] (1) Multi-source data fusion software

[0049] The multi-source data fusion software provides support for the comprehensive display of multi-source high-altitude remote sensing image data and data based on the characteristics of image data, survey data and multimedia data.

[0050] Data information fusion is based on the target information extracted from panchromatic, multispectral, SAR, infrared images, spectral images, as well as survey information, historical data information and other data information to achieve fusion recognition, auxiliary decision-making, and form comprehensive target recognition capabilities;

[0051] The comprehensive display application provides comprehensive display of multi-source high-altitude remote sensing data, realizes three-dimensional display of image data, survey data, multimedia data and other auxiliary information, and has the ability to visualize the operating status of our on-orbit survey sensors. It can simulate and analyze on-orbit sensor resources according to the survey area, survey time, etc., and provide technical support for the formulation of survey needs and the selection of collaborative survey strategies.

[0052] Multi-source data fusion software consists of data information fusion and comprehensive display application components.

[0053] (2) Data organization and management software

[0054] The data organization and management software manages the visible light, SAR, infrared, hyperspectral data and technical survey text data connected to the system, and realizes the organization and management of multi-source massive remote sensing data materials and data products.

[0055] Basic data management functions include standardized storage, modification and deletion functions, and multi-condition compound retrieval of signal data, hyperspectral data, and infrared data products;

[0056] The target characteristic data management function includes the standardized storage, modification and deletion functions, and multi-condition compound retrieval of hyperspectral spectral characteristic data and infrared characteristic library data;

[0057] System management and maintenance functions include management and maintenance of users, permissions, logs, as well as storage, database storage monitoring, etc.

[0058] The data organization and management software consists of four components: basic data management, target characteristic data management, data material and data product management, and system management and maintenance.

[0059] (3) Business information management software

[0060] Business information management software mainly realizes the information management of data production business, completes data production, data reception, statistical analysis of generated data, monitoring and statistics of job tasks, statistical report generation, etc. In terms of data task production, the software needs to provide full life cycle management functions for data production tasks; in terms of data task statistics, the software needs to provide statistical functions based on data acquisition, data product distribution, workload, etc.; in terms of data production life cycle, the software needs to provide each subsystem user with permission to access the business information management software.

[0061] Business information management software consists of three parts: task management component, data statistics component, and user management component.

[0062] (4) Image comprehensive judgment software

[0063] The image comprehensive interpretation software provides three functions of remote sensing image processing, image data interpretation and remote sensing data compilation according to the characteristics of hyperspectral, thermal infrared, visible light and SAR image data, and provides technical support for the transformation of images into data products, specifically including the following three aspects.

[0064] Remote sensing image processing For hyperspectral images, we build general processing tools for hyperspectral images, such as spectrum picking and spectrum envelope removal; for thermal infrared images, we build general infrared processing tools, such as infrared feature measurement and infrared pseudo-color display; for visible light and SAR image data, we provide common basic processing functions such as color scale adjustment, Lee filtering, and geometric correction, providing basic processing support for subsequent image data interpretation;

[0065] Image data interpretation provides basic browsing, measurement and annotation functions. At the same time, the software extracts thematic information from hyperspectral and thermal infrared image products, including state changes, operating environment elements, etc., and uses effective target detection algorithms to detect and extract large ship targets, and combines knowledge bases to realize auxiliary recognition and analysis of large ship targets. Provide technical support for the transformation of images into data materials;

[0066] Remote sensing data compilation is based on the results of image interpretation and analysis. According to standard data product specifications, it realizes the production and standardized output of data products through functions such as annotation, finishing, comprehensive display, and output, thereby ultimately realizing the transformation of data materials into data products.

[0067] The image comprehensive judgment software consists of three components: remote sensing image processing, image data interpretation, and remote sensing data compilation.

[0068] System Process

[0069] Based on business needs analysis, the remote sensing information system of the operating unit includes two major processes: fixed target observation and offshore ship target observation.

[0070] (1) Fixed target observation process

[0071] The fixed target observation process is aimed at operating units and ports, using the second-generation distribution data package as input to determine the number and type of ship targets in the port.

[0072] The system first decompresses the second-generation distribution data package, obtains visible light, SAR, thermal infrared, and hyperspectral images, creates thumbnails, obtains image information, and initiates the task process based on the target information contained in the image. If it is a nearshore port interpretation based on high-resolution visible light images, the nearshore port ship target automatic extraction process based on the knowledge base is initiated. The user logs in to the task management software, seizes the interpretation task and opens the target image. If there is an automatic detection result, it is automatically loaded to assist the user in interpretation. The user uses the knowledge base, historical data materials and products to realize the interpretation and confirmation of the target, and on this basis, performs data compilation and processing, and finally closes the data storage loop. The whole process is like Figure 1 shown.

[0073] (2) Observation process of marine ship targets

[0074] The marine ship target observation process targets marine ship targets and utilizes sensor visible light wide-band data and ScanSAR data, combined with survey data, distribution data, text data, and integrates historical data database and knowledge base data to realize the determination of large marine ship targets.

[0075] The system first decompresses the second-generation distribution data packet, obtains the sensor wide-band visible light data and ScanSAR data, obtains the image information to initiate the operation task process, and simultaneously initiates the rapid detection task of marine targets. The system simultaneously connects the sensor survey data, distribution data, and text data and stores them in the database. The user logs in to the operation task management software, seizes the interpretation task and opens the target image, uses the ship target detection result confirmation tool to confirm the ship target extracted from the image, and then pushes the slice information to the multi-source data fusion software, and comprehensively confirms and interprets the ship target in combination with the survey data, distribution data, and text data information. Finally, the data is compiled and output based on the interpretation data information, and the data is finally stored in the database to realize the task closed loop.

[0076] System external interface

[0077] The external relations of this system mainly include interfaces with data distribution system and image and text system.

[0078] (1) Interface with the data distribution system

[0079] This system accesses the image data and distribution package data of the data distribution system. The input information is the series of sensor survey data and distribution package data released by the data distribution terminal, which are sent to the directory monitored by the data access related function module of the business information management software of this system through network transmission. The business information management software will automatically decompress the data packet, obtain the data information and initiate the operation task.

[0080] (2) Interface with image text system

[0081] The interface between this system and the image text system is the input interface. The input information is the target image data and text data generated by the image text system, which are sent to the remote sensing image target information database subsystem of this system through the local area network and included in the integrated database for unified management. The data to be released is sent to the data product standardization processing station of this system through the internal local area network for standardization processing before data release.

[0082] System internal interface

[0083] The internal interfaces of the remote sensing information processing system of the operating unit are mainly the interfaces between the multi-source data fusion software, data organization and management software, business information management software and image comprehensive judgment software. Figure 2 shown.

[0084] The internal interface of the remote sensing information processing system of the operating unit adopts TCP / IP message mechanism or files to transmit instructions, messages and data.

[0085] The present invention provides a target intelligent recognition method using a remote sensing information processing system, comprising:

[0086] Step 1: Fine-grained extraction of hyperspectral features based on mixed pixel decomposition and spectral purification.

[0087] The present invention constructs access to the sensor's hyperspectral payload image data, and uses hyperspectral image data to extract ground features in a refined manner, which is an important application of sensor hyperspectral data. However, due to the low spatial resolution and high spectral resolution of hyperspectral data, the phenomenon of "mixed pixels" is more significant, which directly affects applications such as battlefield feature refinement and target material determination. At the same time, there are many bands, high dimensions, and large amounts of relevant information between bands, which also affects the amount of computation and accuracy of processing.

[0088] (1) Rapid endmember extraction based on FGDA algorithm

[0089] To address the problem of mixed pixels, this system constructs an endmember extraction method based on the FastGram Determinantbased Algorithm (FGDA) to quickly extract endmembers, and then uses a linear mixed model to decompose the mixed pixels into components. Finally, in the abundance composition obtained by the decomposition, the main abundance is selected and combined with the matched filter (MF) method to perform fine extraction and analysis of battlefield elements.

[0090] The key to the processing is the rapid extraction of endmembers. When performing mixed pixel decomposition, the endmembers in the actual scene are often unknown, and the amount of calculation required to directly search for endmembers is often large. Therefore, a rapid endmember extraction algorithm based on the FGDA method was developed during the implementation of the system.

[0091] First, we introduce a new high-order simplex volume (SV) V(e1,...,e p ) as shown in formula (4-39):

[0092]

[0093] Among them is A p =[e2-e1,e3-e1,...,e p -e1] end member vector, when looking for the Kth end member, let:

[0094]

[0095] Then for K≥2, the expanded relationship is as shown in formula (4-41):

[0096]

[0097] Comparison found B k+1 With B k There exists a recursive relationship as shown in formula (4-42):

[0098]

[0099] Symmetric matrix B k The determinant of is called the Cramer determinant, and according to the block matrix determinant calculation rule, it is easy to get formula (4-43):

[0100]

[0101] Assuming that the Kth end member has been obtained, the K+1th end member can be obtained by finding the maximum det(Bk+1). It is a constant for all pixels, so formula (4-44) is used to calculate and determine which pixel is the K+1th endmember. Therefore, the endmember score index (ESI) is defined to define this indicator.

[0102]

[0103] It can be seen that the K+1th endmember can be obtained by calculating ESI alone, which greatly reduces the amount of calculation for endmember search. In addition, the application of data has proved that the endmembers found by the above method are effective. In terms of computational complexity, let L be the number of bands and N be the total number of pixels. The computational complexity of N-FINDR is σp η+1 N, the computational complexity of SGA is The computational complexity of MVHT is The complexity of NVFS is σ(p η+1 +p 3 L-2p 2 L+pL)N, the computational complexity of OBA is 3pLN, where 2.3≤η≤2.9, 3≤σ≤5. For FGDA, extracting a new end member only requires calculating ESI, and in each cycle, It only needs to be calculated once, and the complexity of calculating ESI for each pixel is k 2 +L, adding a new end member requires a cycle (k 2 + L)N times, so the computational complexity of FGDA is like Figure 3 As shown in the figure, the computational efficiency of FGDA is generally higher than that of endmember extraction algorithms such as N-FINDR, SGA, and MVHT.

[0104] After obtaining the end members, the linear mixed model is used to decompose the mixed pixels. The formula is shown in formula (4-45):

[0105]

[0106] Where r is the actual observed spectrum vector, x is the mixed spectrum, n is the noise, E = [e1, e2, ..., e p ] is the end member vector, c i For each end member component.

[0107] In this way, by decomposing the mixed pixels, the abundance of each component can be obtained. Assuming that the first abundance component of a pixel is x, the MF is constructed by combining the target spectrum x0 to achieve matching enhancement of the extracted military target. The relevant transformation can be described as Equation (4-46)-Equation (4-48):

[0108] y=H T x (4-46)

[0109] min(H T ΣH)stH T x0=1 (4-47)

[0110]

[0111] Where y is the matching result and Σ is the covariance matrix of the input image vector.

[0112] (2) Spectral feature extraction based on HOSVD

[0113] Hyperspectral data has high dimensionality, and direct calculation requires a lot of computation. In addition, there is a large correlation between the data. In order to reduce the dimensionality of the data while taking into account the efficiency of calculation, a feature extraction method based on Higher Order Singular Value Decomposition (HOSVD) was constructed.

[0114] Use tensors to describe this process. Assume that the tensor where I1,I2,...,I N ∈N is the high-dimensional boundary of the tensor, then for the hyperspectral data B={r1,...,r M}, then its m-order tensor can be described as formula (4-49):

[0115]

[0116] Where M is the number of pixels, r i is the image spectrum vector, then R can be regarded as the outer product of the spectrum, that is Similar to PCA operation, the data is first centralized, that is:

[0117]

[0118] In formula (4-50), is the spectral mean vector, F is the whitening matrix, so for the tensor A , define its HOSVD operation as:

[0119]

[0120] In formula (4-51) S is an I1×I2×…×I N The core tensor of rank, is a singular value matrix, and similar to SVD, S The subtensors of are also orthogonal, that is, Assumption A (n) (S(n) )for A ( S ) is the expanded form, then A (n) The SVD decomposition of can be expressed as:

[0121]

[0122] In formula (4-52), let U (1) =U (2) =…=U (N) =U, then A (n) Expressed as:

[0123]

[0124] And there is || S 1||≥|| S 2||≥…≥|| S L || is the root of the eigenvalue of the matrix U. So far, we have implemented the HOSVD operation on the original data and realized the dimensionality reduction operation on the data.

[0125] In this way, by optimizing the principal components, the dimensionality reduction of the original image is achieved. Experiments have shown that for hyperspectral data, selecting the first 10 principal components can represent the main information of the original scene image. This can be seen from Figure 4 and Figure 5 Verified in.

[0126] By adopting the above algorithm, the decomposition and feature extraction of mixed pixels of hyperspectral data are realized. Further using classifiers such as support vector machines and random forests, the extraction of ground features can be realized. To prove the effectiveness of the algorithm, we selected 20 standard scenes of hyperspectral images containing typical scenes for testing, and used manual interaction to evaluate the performance of the extraction results. The experiment showed that the system's refined extraction accuracy of battlefield environmental elements is higher than 90%, which is significantly better than the extraction results of commercial software such as ENVI.

[0127] Step 2: Determine the working status of the operation target based on target thermal radiation modeling and multi-temporal analysis.

[0128] Extracting the target's thermal radiation information from thermal infrared images will help determine and analyze the state of the operating target. Infrared remote sensing reflects the target's thermal radiation in a quantitative form, so the target's infrared feature model must also be quantitative and must reflect the target's infrared radiation, rather than just using the features that reflect the target's geometric shape commonly used in visible light imaging. In order to reflect the role of specific heating components in the target on target interpretation, the target feature model needs to include features such as the location of the target's hot or cold components and their mutual relationship. In addition, the sensor's infrared imaging resolution is low, and the target shows a small number of pixels in the image, making it difficult to extract statistical features (such as texture, etc.) that require a certain number of pixels to be more accurate.

[0129] (1) Classification of imaging conditions

[0130] Since the infrared radiation of the target under different imaging conditions shows great differences, it is necessary to establish feature models under multiple imaging conditions. The target radiation characteristics under each imaging condition are relatively stable, which is conducive to establishing a stable feature model. The imaging conditions that need to be considered in establishing the feature model include imaging time, longitude and latitude, spatial resolution, temperature resolution, target orientation, target state, imaging band, etc. At present, the system only considers time (month and time), spatial resolution, temperature resolution and imaging band, and other conditions are not distinguished. Since the spatial resolution, temperature resolution and imaging band are given for the same sensor, the time division is mainly considered. The month is divided into 12 intervals, and each natural month is a monthly interval; the time is divided into 24 intervals, and the model of the hour is established. The scope of each moment model includes the first half hour and the second half hour of the moment. In order to enhance the scalability of the model, the following two treatments are performed: ① Although the current sensor spatial resolution, temperature resolution, and imaging band are fixed, there is still room for expansion, so that after obtaining sensor data with other spatial resolutions, temperature resolutions, and imaging bands, it can still be accurately modeled; ② Although other imaging conditions are not distinguished, these imaging conditions, if they exist, are also added to the database to provide scalability for providing more detailed models.

[0131] (2) Target infrared signature calculation

[0132] Compared with visible light images, infrared images, especially thermal infrared images, mainly reflect the thermal radiation characteristics of the target, so feature extraction needs to take into account the characteristics of infrared images. The main features used in this project include infrared radiation mean, standard deviation, infrared radiation mean difference between the target and the surrounding background, mean contrast and other statistical quantities to analyze the overall infrared radiation characteristics of the target. The calculation formula is as follows:

[0133] The target infrared radiation mean is as follows:

[0134]

[0135] Where (x, y) is the target area R O The pixels within, N O is the number of pixels in the target area;

[0136] The standard deviation of target infrared radiation is as follows:

[0137]

[0138] The mean difference of infrared radiation between the target and the surrounding background is as follows:

[0139] μ=μ O -μ B (4-56)

[0140] in R B is the background area, N B is the number of pixels in the background area.

[0141] The infrared mean contrast between the target and the surrounding background is expressed as (4-57):

[0142]

[0143] in

[0144] As for the infrared radiation characteristics inside the target, we analyze them from the aspects of absolute position and relative position relationship of special points such as the strongest point, second strongest point, weakest point, second weakest point, etc., so as to avoid statistical characteristics that require a certain number of statistical sample points to obtain.

[0145] (3) Feature model generation and update

[0146] The characteristic model of the target is established under different conditions, and the infrared characteristic model of the target under various conditions is generated. After the features of all samples under specific imaging conditions are extracted, statistical analysis is performed on them, and the mean and standard deviation are calculated. The mean and standard deviation are used to represent the characteristic distribution under this condition. The mean and standard deviation are used as the characteristic model under this imaging condition. Since the characteristic model can be used for target detection tasks, the parameters in the method of constant detection rate are given.

[0147] If there is no data in the feature model under the imaging condition corresponding to the feature to be stored, the feature can be added directly. If the feature model under the imaging condition has been established, the feature model needs to be updated. Assuming that there are N samples, their mean and variance are μ and σ respectively, and the feature to be stored is F, the updated mean and variance are respectively as equation (4-58) and equation (4-59):

[0148]

[0149] This system has built a target-based thermal radiation modeling and multi-time comparison analysis plug-in, realizing the thermal radiation modeling tool for targets such as ships, oil tanks, and fixed target facilities. Figure 6 As shown, the modeling and display of the thermal radiation of a certain target changing with time is realized.

[0150] Step 3: Feature mining of massive SAR data based on sparse deep belief network.

[0151] This system construction is connected to SAR payload data, and it is of great significance to interpret and interpret ship target information from SAR image data.

[0152] In view of the characteristics of massive SAR data, this project uses restricted Boltzmann machines to perform layer-by-layer unsupervised feature learning on a large amount of training data. The learning process uses the contrast divergence algorithm based on Gibbs sampling to solve the parameters. Then the parameters solved by the multi-layer restricted Boltzmann machine are used as the initial values ​​of the multi-layer neural network. The model parameters of the multi-layer neural network are then fine-tuned using the labeled training data and the multinomial logistic regression model is used as the final classification model. The fine-tuning process uses the backpropagation algorithm and the conjugate gradient descent method. Finally, the new data is forward propagated through the network to obtain the final feature representation result.

[0153] (1) Feature learning based on hybrid structure sparse constraints

[0154] In high-resolution SAR image recognition, an image contains not only the target of interest to the observer, but also a large amount of complex background. These interference information will affect the performance of the recognition method. A common method is to suppress useless or noise information in the features through function constraints or regularization to improve the generalization performance of the system. In this regard, the present invention proposes an unsupervised training method of a restricted Boltzmann machine based on hybrid structure sparse constraints. Specifically, the optimization loss function is defined as shown in formula (4-60):

[0155] L=-logP(v)+μ||P(h=1|v)||1+λP(h=1|v)|| 1,2 +η||W||2 (4-60)

[0156] Among them, μ represents the hidden layer node activation probability, L1 norm constraint partition coefficient, λ represents the hidden layer node activation probability L 1,2 The partition coefficient of the mixed norm constraint, η represents the partition coefficient of the L2 norm constraint of the weight matrix W.

[0157] The second term on the right side of formula (4-60) is the hidden layer node sparse constraint based on the L1 norm, which is defined as follows:

[0158]

[0159] According to sparse coding theory, the optimization problem with norm constraints is equivalent to the optimization problem with L0 norm constraints, and the latter can make most of the optimized target parameters 0, achieving the purpose of parameter sparseness. Here, adding sparse constraints can make the activation probability of most hidden layer nodes 0, thereby achieving the purpose of encoding the input data features with a small number of hidden layer nodes.

[0160] The third term on the right side of formula (4-60) is based on L 1,2 Sparse constraint of the mixed norm. Divide the hidden layer nodes of the restricted Boltzmann machine into m sets, where the L of the hidden nodes is defined 1,2 The mixed norm is as shown in formula (4-62):

[0161]

[0162] Among them G m Represents the mth hidden layer node set. By optimizing and adding L 1,2 The objective function of the mixed norm is that the sum of the squares of the node activation values ​​of most hidden layer node sets is close to zero, which makes the activation values ​​of most nodes in the set also close to zero, thus achieving an effect similar to the L1 norm constraint. Therefore, when data is input to the visible layer nodes, only a small number of hidden layer nodes in the set are activated, thus achieving the purpose of sparse constraint. By adding the L1 norm of the hidden layer node activation value and the L 1,2 Mixed norm constraints can give different types of constraints to the activation probabilities of hidden layer nodes.

[0163] The fourth term on the right side of formula (4-60) is the L2 norm constraint that limits the Boltzmann machine network weight matrix. Its specific form is shown in formula (4-63):

[0164]

[0165] Adding constraints on the weight matrix during optimization can make the model smoother and reduce the possibility of overfitting by suppressing excessive weights. At the same time, this constraint can suppress redundant weights in the weight matrix and enhance the generalization performance of the model.

[0166] Optimizing energy models usually relies on stochastic gradient descent, which requires the use of algorithms such as Gibbs sampling to estimate the parameters of the model and thus calculate the gradient. However, repeatedly performing Gibbs sampling on undirected graph models such as Boltzmann machines is not only computationally intensive but also very time-consuming. The present invention uses the contrast divergence algorithm to perform approximate calculations. This approximate calculation can obtain similar optimization results and greatly reduces the amount of calculation. Therefore, this method is widely used in the actual gradient solution process.

[0167] (2) Model parameter fine-tuning

[0168] By training multiple restricted Boltzmann machines layer by layer, we can obtain high-level abstract features, that is, achieve the purpose of unsupervised feature learning. The parameters of these restricted Boltzmann machines will be used as the initial values ​​of the multi-layer neural network parameters. Then we use the back propagation algorithm and conjugate gradient descent method to fine-tune the parameters of the entire network. This process is a supervised training process and will be trained with labeled data.

[0169] In the optimization problem of high-dimensional feature space, in order to avoid the possible local optimum, the model has more optimization paths to choose from, which makes it easier to find the global optimal solution for the optimization problem in high-dimensional space than in low-dimensional space. In other words, the more hidden layer nodes there are in the network, the lower the possibility of the optimization process falling into the local extreme value, and the more parameters there are, the stronger the feature expression ability of the model will be. However, the increase in model parameters will bring another more serious problem, that is, the possibility of overfitting is greatly increased, which greatly reduces the generalization performance of the model. In the relevant theories of machine learning, the most commonly used method to avoid overfitting is regularization, that is, adding regularization terms in the optimization objective function to suppress overtraining of parameters. In the present invention, we will use sparsity constraints to improve the back propagation algorithm used for multi-layer neural network optimization to improve the generalization performance of the model.

[0170] When the activation probability of a hidden layer node is close to 1, the node is considered to be activated, and when it is close to 0, the node is suppressed. When the average activation probability of a node in the hidden layer for all input data is small, it can be considered that the node is not activated most of the time, and the hidden layer node can be considered to be in a sparse state. The average activation value of the jth hidden node in the hidden layer is defined as formula (4-64):

[0171]

[0172] where a j (x (i) ) represents the activation probability of the jth hidden layer node when the i-th data is input to the visual layer, and M represents the total number of input data. When the value of is a small value, it can be considered that the hidden layer nodes have reached a sparse state. Close to a smaller value, the following sparsity constraint function is defined here as shown in formula (4-65):

[0173]

[0174] Add this constraint function to the loss function of the back propagation algorithm, and define it as formula (4-66):

[0175]

[0176] Where β represents the partition coefficient of the sparse constraint term, and J(W,b;x,y) represents the loss function of the back propagation algorithm, which is in the form of formula (4-67):

[0177]

[0178] where h w,b (x) is the predicted value of the input data x through the multi-layer neural network, and y is the manually labeled value of the data. The loss function is also optimized using the stochastic gradient descent method, and the specific method is shown in equations (4-68) and (4-69):

[0179]

[0180] The partial derivative calculation method of the model is as shown in equations (4-70) and (4-71):

[0181]

[0182] in represents the back-propagated residual of the i-th node in the k+1-th hidden layer of the neural network. In this model, the residual function can be derived:

[0183]

[0184] In formula (4-72), f'(x i ) represents the derivative of the activation function. The ρ value is set to a value close to 0. Through the above iterative optimization, a multi-layer neural network model with sparse hidden layer nodes is finally output.

[0185] (3) Unsupervised feature extraction

[0186] In the model training stage, a batch training method is used to randomly and evenly divide the training images into multiple data subsets. Each image belongs to only one of the subsets. The unsupervised pre-training method introduced above is used to train four restricted Boltzmann machines layer by layer, where the output of the hidden layer of the previous Boltzmann machine is used as the input of the visible layer of the next Boltzmann machine. After the pre-training is completed, the weights and other parameters of the four restricted Boltzmann machines are used to initialize a 5-layer neural network. The parameter initialization settings are as follows: Figure 7 and Figure 8 shown.

[0187] After the parameters are initialized, a supervised layer is added to the top layer of the neural network. The parameters of this layer are initialized by a Gaussian distribution random number matrix with a mean of 0 and a standard deviation of 0.01. Then we use the supervised fine-tuning algorithm introduced earlier and the labeled data to perform supervised training of the neural network. The classification model of the supervised layer is defined as follows:

[0188]

[0189] After the network's unsupervised pre-training and supervised fine-tuning training are completed, a final model is obtained. When new data is obtained, the new data only needs to be simply processed and input into the network's visual layer, and the neural network is forward propagated layer by layer to finally learn the feature vector.

[0190] Step 4: Ship target trajectory fusion based on multi-source information.

[0191] During the construction of the system of the present invention, the focus is on introducing load slicing and point information. When the target capacity is not too large and the processing speed requirement is not high, the target memory management is usually based on the target object data linked list for management. When searching for the target, the target linked list needs to be traversed sequentially, and the time complexity is O(n). If the traditional method is used for processing, the target data information cannot be processed in real time. If the data is backlogged, it will bring great risks to the stable operation of the system.

[0192] Therefore, a fast association retrieval method must be used to achieve fast association retrieval of large-capacity targets. This method can be used for fast target retrieval in the target association fusion process, and can also be used for fast filtering of important targets to further improve the processing speed.

[0193] In addition, there are inconsistencies between sensors in positioning accuracy and time resolution, so effective measures need to be taken to achieve correlation fusion analysis of different points and trajectories.

[0194] (1) Fast association based on index

[0195] With the increasing amount of radar target data, a better index structure is needed to process radar targets more effectively, so that target retrieval can be quickly realized based on the location information and characteristic parameter information of radar targets. However, the conventional one-dimensional parameter retrieval method cannot solve this problem well.

[0196] In the multi-dimensional spatial data retrieval technology, the R*-tree index theory is well-known internationally. The starting point of the theoretical research is to achieve fast retrieval of multi-dimensional spatial data. After more than 20 years of development, it has basically matured and has been put into practical use by well-known commercial database manufacturers.

[0197] Based on the R*-tree index theory, we proposed a new method for retrieving radar targets. We mapped the attribute features of radar targets to a point in multidimensional space, and then constructed an index - RR-Tree based on MBR. When retrieving data, we pruned the RR-tree according to the effective filtering method of the index, which can quickly approximate the query results. The RR-tree can update the index while completing the retrieval process. Experiments have shown that R*-tree is a very effective index structure. The basic definition of the index is as follows:

[0198] ①. Range query: Given a certain error range ε, a query point q is given so that the error between the query results r and q is less than ε;

[0199] ②.KNN query: given a query point q, query the N points closest to q;

[0200] ③. Radar target vector: The position information and technical characteristic parameter data of the radar target are formed into a multidimensional array, which is called the radar target vector;

[0201] ④. Radar target vector space: It is described by forming a d-dimensional space with d items of radar target vector;

[0202] ⑤. Radar target space distance: Define the distance between two points p1(m0,m1…md-1) and p2(n0,n1…nd-1) in radar target vector space as:

[0203]

[0204] ⑥.MBR: The minimum bounding rectangle containing n radar vector points.

[0205] In R*-tree, node definitions include leaf nodes and non-leaf nodes. For an RR-tree with N index nodes, if the target number of each node is at most m, the height of the tree is at most logmN–1. Leaf nodes are represented by (I, tuple-identifier), where tuple-identifier corresponds to a tuple in the target database, and I represents a rectangle or point in n-dimensional space, that is, I = (I0, I1, …, In-1), where n is the dimension of the space, Ii represents a closed interval in the i-dimensional space, and represents the expansion of a spatial data in the i-dimensional space; non-leaf nodes are represented by (I, child-pointer), where child-pointer is the pointer to the child node of the node in the index tree, and I is the large rectangle containing all child nodes.

[0206] To describe query algorithms such as range query, KNN query, and NN query, the following concepts are introduced:

[0207] (a) Expansion queue (ItemQueue): A priority queue formed by sorting the MINDIST size, storing the nodes to be expanded first;

[0208] (b) ResultQueue: It is a linked list of size K, storing the best k nodes so far. The result queue is formed by sorting according to MINMAXDIST;

[0209] (c) K distance (kthdist): MINMAXDIST corresponding to the first k items in the result queue so far, or infinity if there are fewer than k items. It is an upper bound on the K nearest neighbors.

[0210] The query algorithm can be described as:

[0211] ①. Use MINDIST to control the extended nodes;

[0212] ②. Use the nodes of ResultQueue to reduce MBR and nodes;

[0213] ③. When ResultQueue is the target node, the search ends.

[0214] (2) Target Information Fusion

[0215] The purpose of target information fusion is: first, to improve the accuracy of data processing, that is, to improve the accuracy of parameters or positions through fusion; second, to improve the credibility of data processing, by continuously tracking the target or accumulating parameters to improve the credibility of the target; third, to reduce the increase in target batches during data processing, by integrating the correlation characteristics of the detected targets through fusion technology, merging the same target information, and reducing the increase in target batches.

[0216] The first is position fusion. In position fusion, Fig. 9 The positions of the two radars and their corresponding error ellipses are shown. The fused radar position is given by equation (4-75):

[0217]

[0218] The error ellipse of the corresponding fusion result can be calculated by the following method:

[0219] Using two position error parameters (major axis, minor axis, and major axis direction), the variance matrices of the two position measurements can be given respectively, as shown in equations (4-76) and (4-77):

[0220]

[0221] The variance matrix after fusion is shown in formula (4-78):

[0222]

[0223] Then the major axis of the corresponding error ellipse is as shown in formula (4-79):

[0224]

[0225] The short axis is shown in formula (4-80):

[0226]

[0227] The major axis direction is shown in formula (4-81):

[0228]

[0229] Then comes parameter fusion. Parameter fusion performs a similar one-dimensional variance weighted fusion. For example, for RF parameters, the RF features are first fused to merge the RF features of the two. For related RF parameters, the difference between the highest frequency and the lowest frequency is used as the variance, and the frequency center is used as the measurement mean to perform one-dimensional variance weighted fusion to obtain the fused frequency center and variance, and thus the fused highest frequency and lowest frequency. The fusion of other parameters is similar.

[0230] Targeted processing is carried out for the problem of weak sensor observation signals and large errors in some parameter strategies. For example, for small pulse width processing near the sensitivity, the pulse width grouping and large strategy is adopted, which can effectively improve the accuracy of the pulse width.

[0231] (3) Curvature-based segmented dynamic threshold trajectory compression

[0232] Aiming at the problems of low efficiency and information loss of traditional compression methods in cleaning massive high-noise data, a data compression method based on curvature segmented dynamic threshold algorithm is proposed. Fig.10 As shown in the figure, this method can avoid the crack phenomenon caused by noise and the significant loss of useful information. The compression ratio of this method can be as high as 83.3%, which is far higher than similar methods.

[0233] (4) Trajectory similarity calculation based on target prior information

[0234] Finally, an efficient and widely applicable trajectory similarity calculation algorithm is proposed. This method matches the similarity of position and trajectory based on the target prior information and calculates the confidence probability of each data source. Based on this strategy and method, the influence of noise can be significantly reduced and the solution efficiency can be optimized.

[0235] Step 5: Software plug-in architecture and multi-source load data interpretation of operating units.

[0236] The remote sensing information processing system of the operating unit has the characteristics of multiple types of access loads, integrated and specialized data processing, etc. In addition, it faces difficulties such as large data volume, strong timeliness, and high requirements for comprehensive data processing capabilities during data processing. Therefore, it is necessary to establish an efficient data production line to realize the comprehensive application and analysis of different sensor data and give full play to the comprehensive interpretation benefits of multi-load observation data.

[0237] Establishing a unified data interpretation process can support multi-person and multi-task interpretation, realize batch production of data, and improve comprehensive interpretation efficiency.

[0238] In addition, the subsystem software architecture is required to be able to support functional expansion and updating well, have strong flexible expansion capabilities, and realize the access and analysis and processing of multi-source payload data, so that the system has stronger multi-source data comprehensive analysis capabilities and improves image interpretation and compilation efficiency.

[0239] (1) Data production process framework

[0240] For multiple types of payloads such as visible light, SAR, thermal infrared, and hyperspectral, it provides advanced interpretation process management and comprehensive interpretation compilation.

[0241] The present invention realizes a unified data production process framework, unifies data standards, operation modes and data output, standardizes the processing flow, and optimizes the working mode and processing links. For the processing process that takes a long time and can be undertaken by the computer, it is completed automatically; for the links that must be participated by humans, a variety of semi-automatic auxiliary tools are provided, with process processing as the main line to supplement the interpretation link, which significantly improves the data processing efficiency.

[0242] The processing logic steps of the process framework are as follows:

[0243] ①.Task initiation: automatically access the second-generation distribution data, decompress the image compression package, and obtain the image file;

[0244] ②. Image information extraction: read the image and obtain relevant information of the image, including the image payload, imaging time, target location, etc., store the information in the database, and create a target automatic extraction interface file;

[0245] ③. Automatic target detection and recognition: automatically extract targets from images and write slices or detection result files into task files;

[0246] ④. Target manual identification, confirmation and association: The user displays the image and loads the result file of automatic detection to assist in interpretation. At the same time, the target information is interpreted and extracted with the help of the target knowledge base;

[0247] ⑤. Data compilation, output and storage: Based on the data information obtained, the data is compiled and output to form data products, the products are stored in the warehouse, and the next interpretation task is carried out.

[0248] (2) “Microkernel + plug-in” software system architecture

[0249] The system needs to access, process and analyze multiple payload data such as visible light, SAR, hyperspectral, thermal infrared, etc. To ensure the processing and analysis of multi-source data, the system proposes a "microkernel + application plug-in" platform design in the software architecture design, such as Fig.11 shown.

[0250] This is a flexible system design structure that realizes the sharing of information, software, computing, storage and other resources, and can provide information, software, computing and storage resources on demand, supporting the on-demand distribution of information. It standardizes the data transmission format and information code standards, provides a universal interface, and has unified information exchange and computing capabilities. It adopts virtualization design ideas and flexible and scalable software architecture design to solve integration and expansion problems.

[0251] The architectural features of this "microkernel + application plug-in" are:

[0252] Microkernel: As the core layer of the software, it implements the most basic data reading, algorithm scheduling, function hooking, etc.

[0253] Application plug-ins: It realizes functions such as dedicated interpretation and analysis, geospatial data presentation, target query retrieval and proprietary data compilation and output. The software modules can be flexibly split and customized according to usage needs, achieving scalability in data types, functional tools and equipment scale.

[0254] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for intelligent target recognition using a remote sensing information processing system, wherein: The remote sensing information processing system comprises: (1) Multi-source data fusion software; (2) Data organization and management software; (3) Business information management software; (4) Image comprehensive judgment software; System external interface: the remote sensing information processing system external relations include interfaces with data distribution system and image text system; System internal interface, the remote sensing information processing system internal interface is the interface between multi-source data fusion software, data organization management software, business information management software and image comprehensive judgment software; According to the business demand analysis, the remote sensing information processing system includes two major processes: fixed target observation and marine ship target observation; The target intelligent recognition method using the remote sensing information processing system comprises: Step 1: Refined extraction of hyperspectral features based on mixed pixel decomposition and spectral purification; Step 2: Determine the working status of the operation target based on target thermal radiation modeling and multi-temporal analysis; Step 3: Feature mining of massive SAR data based on sparse deep belief network; Step 4: Fusion of ship target trajectory based on multi-source information; Step 5: Software plug-in architecture and multi-source load data interpretation of operating units.

2. The target intelligent identification method using a remote sensing information processing system according to claim 1, characterized in that: in, The multi-source data fusion software provides support for the comprehensive display of multi-source high-altitude remote sensing image data and data provided by image data, survey data and multimedia data, and is composed of data information fusion and comprehensive display application components; the data organization and management software manages the visible light, SAR, infrared, hyperspectral data and technical survey text data connected to the system, realizes the organization and management of multi-source massive remote sensing data materials and data products, and is composed of four components: basic data management, target characteristic data management, data material and data product management, and system management and maintenance; the business information management software realizes the informatization management of data production business, completes data statistical analysis, task monitoring and statistics, and statistical report generation, and is composed of three parts: task management component, data statistics component, and user management; the image comprehensive judgment software processes hyperspectral, thermal infrared, visible light and SAR image data, and is composed of three components: remote sensing image processing, image data interpretation, and remote sensing data compilation.

3. The target intelligent identification method using a remote sensing information processing system according to claim 1, characterized in that: The fixed target observation process is aimed at operating units and ports, using the second-generation distribution data packet as input to realize the interpretation of the number and model of ship targets in the port; the marine ship target observation process is aimed at marine ship targets, using sensor visible light wide-band data and ScanSAR data, combined with survey data, distribution data, text data, and integrated historical data database and knowledge base data to realize the determination of large marine ship targets.

4. The target intelligent identification method using a remote sensing information processing system according to claim 1, characterized in that: The step 1 further includes: (1) fast extraction of end members based on the FGDA algorithm, fast extraction of end members based on the Cramer determinant fast calculation end member extraction method, and then using a linear mixed model to decompose the mixed pixels into components, and finally in the abundance composition obtained by decomposition, the main abundance is preferred and combined with the matched filtering method to perform fine extraction and analysis of battlefield elements; (2) spectral feature extraction based on HOSVD, using the correlation between data for hyperspectral data, in order to reduce the dimension of the data while taking into account the efficiency of calculation, constructing a feature extraction based on high-order singular value decomposition.

5. The target intelligent identification method using a remote sensing information processing system according to claim 1, characterized in that: The step 2 further includes: (1) imaging condition division, establishing feature models under multiple imaging conditions respectively, wherein the imaging conditions for establishing the feature models include imaging time, longitude and latitude, spatial resolution, temperature resolution, target orientation, target state, and imaging band; (2) target infrared feature calculation, wherein the features used for feature extraction include infrared radiation mean and standard deviation, infrared radiation mean difference between the target and the surrounding background, and statistical quantity of mean contrast to analyze the overall infrared radiation characteristics of the target; (3) feature model generation and update, after feature extraction of all samples under specific imaging conditions, statistical analysis is performed on them, and the mean and standard deviation are calculated. The mean and standard deviation are used to represent the feature distribution under the condition, and the mean and standard deviation are used as the feature model under the imaging condition.

6. The target intelligent identification method using a remote sensing information processing system according to claim 1, characterized in that: The step 3 further includes: (1) feature learning based on hybrid structure sparse constraints; (2) fine-tuning of model parameters, by training multiple restricted Boltzmann machines layer by layer to obtain high-level abstract features, and then using the back propagation algorithm and conjugate gradient descent method to fine-tune the parameters of the entire network; (3) unsupervised feature extraction, using a batch training method to randomly and evenly divide the training image into multiple data subsets, each image belongs to only one of the subsets, and using an unsupervised pre-training method to train four restricted Boltzmann machines layer by layer, wherein the output of the hidden layer of the previous Boltzmann machine is used as the input of the visible layer of the next Boltzmann machine. After the pre-training is completed, a 5-layer neural network is initialized using the weights and other parameters of the four restricted Boltzmann machines.

7. The target intelligent identification method using a remote sensing information processing system according to claim 1, characterized in that: The step 4 further includes: (1) fast association based on index, a method for retrieving radar targets based on R*-tree index theory is proposed, the attribute characteristics of the radar target are mapped to a point in multidimensional space, and then the index - RR-Tree is constructed according to MBR; (2) target information fusion, first position fusion, then parameter fusion; (3) trajectory compression based on curvature segmentation dynamic threshold; (4) trajectory similarity calculation based on target prior information, similarity matching of position and trajectory is performed based on target prior information, and the confidence probability of each data source is calculated to reduce the impact of noise and optimize the solution efficiency.

8. The target intelligent identification method using a remote sensing information processing system according to claim 7, characterized in that: in, The index is defined as follows: ①. Range query: given a certain error range ε, a query point q is given so that the error between the query result r and q is less than ε; ②. KNN query: given a query point q, query the N points closest to q; ③. Radar target vector: the position information and technical characteristic parameter data of the radar target are formed into a multidimensional array, called the radar target vector; ④. Radar target vector space: the d items of the radar target vector are used to form a d-dimensional space for description; ⑤. Radar target space distance; ⑥. MBR: the minimum bounding rectangle containing n radar vector points.

9. The target intelligent identification method using a remote sensing information processing system according to claim 1, characterized in that: The step 5 further includes: (1) a data production process framework, which unifies the data production process framework, unifies data standards, operation modes and data output, standardizes the processing flow, and optimizes the working mode and processing links; (2) a "microkernel + plug-in" software system architecture, in which the microkernel is used as the software core layer to realize the most basic data reading, algorithm scheduling, and function attachment; the application plug-in realizes the functions of special interpretation and analysis, geographic spatial data presentation, target query retrieval and proprietary data compilation and output, and the software modules are flexibly split and customized according to the needs of use, so as to realize the scalability of data types, functional tools and equipment scale.

10. The target intelligent identification method using a remote sensing information processing system according to claim 9, characterized in that: The processing logic steps of the process framework are as follows: ①. Task initiation: automatically access the second-generation distribution data, decompress the image compression package, and obtain the image file; ②. Image information extraction: read the image and obtain relevant information of the image, including the image payload, imaging time, target location, etc., store the information in the database, and create a target automatic extraction interface file; ③. Automatic target detection and recognition: automatically extract the target from the image, and write the slice or detection result file into the task file; ④. Target manual identification, confirmation and association: The user displays the image and loads the result file of automatic detection to assist in interpretation. At the same time, the target information is interpreted and extracted with the help of the target knowledge base; 5. Manual recognition, confirmation and association of targets: The user displays the image and loads the result file of automatic detection to assist in interpretation, while using the target knowledge base to interpret and extract target information.

Images