Remote sensing image information extraction system and method of case-based reasoning coupling knowledge graph

Abstract

The invention belongs to the technical field of remote sensing, and discloses a remote sensing image information extraction system and method of a case-based reasoning coupling knowledge graph. According to the invention, a high-resolution image natural resource element knowledge graph and case-based reasoning coupling intelligent interpretation normal form is established. The natural resource element knowledge graph and case-based reasoning coupling intelligent interpretation paradigm established by the invention not only can fully utilize interpretation business knowledge, but also can automatically expand the case library, adaptively adjust the case feature weight and continuously improve the automatic interpretation precision, is one of the characteristics and innovation of the invention, and has a wide application prospect. A Python + PyQGIS framework is used for writing codes to develop prototype system software, a natural resource element sample multi-scale case expression model and an intelligent learning model based on case reasoning and knowledge driving are established, and rapid extraction and intelligent interpretation of natural resource elements under man-machine collaborative guidance are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing technology, and in particular, relates to a remote sensing image information extraction system and method based on case reasoning coupled with knowledge graph. Background Art

[0002] Natural resource elements refer to natural resource entities that exist in material form, are visible and tangible, and can be investigated, monitored and controlled. They cover land and sea, above ground and underground, including land, minerals, forests, grasslands, water, wetlands, sea areas and islands, etc., and have multiple basic attributes such as natural, economic, social, ecological, and spatial three-dimensional. The "Overall Plan for the Construction of the Natural Resources Survey and Monitoring System" released in 2020 clearly proposed to carry out unified surveys and monitoring of natural resources to accurately grasp the status and changes of various natural resources. The interpretation of natural resource elements based on remote sensing images is the most basic, most widely used, most labor-intensive and most technically difficult work in survey and monitoring. The "Outline of the Natural Resources Science and Technology Innovation Development Plan" issued by the Ministry of Natural Resources in 2018 proposed: "We must strengthen the ability to quickly extract and intelligently interpret all elements of natural resources based on multi-source surveys and monitoring results." How to achieve the rapid extraction and intelligent interpretation of natural resource elements from remote sensing images is a key scientific and technological issue that needs to be broken through.

[0003] my country has diverse topography, rich element types, and fragmented surface structures. Existing intelligent interpretation methods face many challenges, such as few interpretable categories, inaccurate and incomplete element boundaries, and low reliability of category attributes. The reasons are that existing methods have the following three problems: ① Incomplete category representation, lack of "feature-mode-category" representation and higher-level feature mining, which makes it difficult to achieve layered deconstruction and accurate cognition of land objects; ② Insufficient knowledge application, the interpretation process lacks business rule constraints and knowledge reasoning, resulting in inaccurate land interpretation and boundaries that cannot meet business needs; ③ Insufficient algorithm reasoning and self-learning capabilities, the interpretation model algorithm is not closely integrated with application requirements and application processes, resulting in low model algorithm reusability, weak migration capabilities, and poor scene adaptability, making it difficult to support high-resolution remote sensing intelligent interpretation of complex natural scenes.

[0004] 1.2 Analysis of current research status and development trends at home and abroad

[0005] (1) Machine Learning Methods

[0006] Among various machine learning methods, various algorithms such as decision trees, random forests, support vector machines, artificial neural networks, and convolutional neural networks have been widely used in remote sensing image interpretation research, especially deep learning methods, which are the hot spots in the current high-resolution image remote sensing interpretation research because of their good feature learning ability. Convolutional neural network (CNN) is the main method used in deep learning remote sensing interpretation. Early remote sensing interpretation research mainly used general image classification CNNs such as VGGNet and AlexNet for classification, and generally used image segmentation or feature reuse methods to complete surface classification. After the emergence of semantic segmentation deep neural networks such as FCN, adding a skip layer structure or feature extraction layer on the basis of FCN, or fusing CNN with MLP, can greatly improve the classification accuracy of CNN. CNN can achieve good results in extracting surface coverage information from high-resolution remote sensing images, but there are still some shortcomings, such as the extremely high number of training samples required, and the surface coverage on high-resolution remote sensing images has an unbalanced spatial distribution, and the annotation of training samples requires a lot of manpower and time. Due to the black box characteristics of neural networks, network features are not easy to understand and lack rigorous mathematical theoretical explanations.

[0007] (2) Cognitive reasoning method

[0008] The knowledge graph is essentially a semantic network. It usually uses a triple such as "head entity-relationship-tail entity" to express the attributes of things and the semantic relationship between things. It can describe the concepts, entities and the relationship between the objective world in a structured form, and can express the remote sensing information in the real world in a form closer to human cognition, so as to better organize and manage the expert prior knowledge, remote sensing imaging mechanism, and geographical knowledge attached to remote sensing images. Various types of knowledge graphs have emerged, such as the geological knowledge graph, remote sensing field knowledge graph, surface cover knowledge graph, wetland knowledge graph, etc. The joint interpretation method of knowledge reasoning and data learning has become an important research trend in the intelligent processing of remote sensing big data, such as the collaboration of geographic knowledge graph and deep learning model, coupling knowledge graph and deep learning, etc. Knowledge graphs are helpful for automatic and intelligent acquisition, mining and application of knowledge, but they still have certain limitations, such as the inability to reasonably represent the relationship between numerical values ​​and the unclear semantic association between numerical nodes.

[0009] The idea of ​​case-based reasoning (CBR) was first described by Schank of Yale University in the United States. Its principle is to use the original case to explain or solve a new problem. Later, it gradually became an important problem-solving method in the field of artificial intelligence. It can effectively deal with the fuzziness and uncertainty problems encountered in acquiring knowledge. It is particularly suitable for the field of geoscience where professional knowledge is difficult to be summarized, abstracted and expressed. Case-based reasoning combined with geography to solve spatial problems has become a new method for geoscience data analysis and prediction. It has been widely used in the research of land use classification and surface cover information extraction from radar images, hyperspectral images [high-resolution remote sensing images]. Case-based reasoning does not require the establishment of a rule-based knowledge base. The case base can be reused, which can improve the efficiency of interpretation work. However, current research has not combined multi-sensor spatiotemporal images for analysis, and the use of remote sensing interpretation business knowledge is insufficient. The model of reasoning application needs to be further improved and deepened, and the paradigm of remote sensing intelligent interpretation has not been formed.

[0010] (3) Analysis of the development trends of intelligent interpretation methods for natural resource elements

[0011] Remote sensing image interpretation technology has gone through the stages of manual visual interpretation, computer classification and semi-automatic interpretation with human-computer interaction. With the ever-changing demand for remote sensing applications, the launch of high-resolution remote sensing satellites, the development of artificial intelligence technology, the accumulation of geographical knowledge, and the application of historical accumulated data, it is currently developing towards the direction of automated intelligence with human-computer integration. Knowledge-driven interpretability methods have received increasing attention, and knowledge graphs and case reasoning provide a technical basis for knowledge-driven interpretation methods. Starting from the cognition of natural resource survey and monitoring objects, building a self-learning, interpretable, scalable, and reasonable element interpretation intelligent cognitive framework and intelligent interpretation system is the core driving force for promoting the intelligent development of remote sensing image interpretation. By comprehensively utilizing the advantages of various methods, it is expected to solve the bottleneck problem that has long plagued the automatic interpretation results of remote sensing images in natural resource survey and monitoring business.

[0012] (4) Review of existing research

[0013] Currently, scholars at home and abroad have used various methods in the field of artificial intelligence to study the intelligent interpretation methods of natural resource elements in high-resolution images. However, there are still deficiencies in category representation, knowledge application, algorithm reasoning and self-learning, sample labeling, and interpretation accuracy. There is still a need to continue exploring the intelligent interpretation methods of natural resource elements in high-resolution images.

[0014] Through the above analysis, the problems and defects of the prior art are as follows:

[0015] (1) The existing natural resource survey and monitoring system lacks the ability to quickly extract and intelligently interpret all-factor information.

[0016] (2) High-resolution image feature information is rich. The representative image features of natural resource elements of different types and resolutions in different time and space environments have their own characteristics. The knowledge used in the interpretation business is also different. How to mine the characteristics of different types of natural resource elements to extract their main common features and construct a natural resource element interpretation knowledge graph and sample case model is one of the key scientific issues of the present invention.

[0017] (3) Most of the existing weighting methods are based on the acquisition of domain knowledge. Although they can achieve the goal of improving the learning performance of the system, they are also prone to local optimal problems. Moreover, these methods are actually passive learning methods. Once the weights are determined, they will no longer be adjusted, even if the case library is constantly updated. Therefore, in order to improve the performance indicators of the CBR classifier, how to solve the problem of parameter optimization and seeking the global optimal solution, and to avoid passive learning so that the feature weights are dynamically adjusted as the case library is constantly updated, is also one of the key scientific issues of the present invention. Summary of the invention

[0018] In response to the problems existing in the prior art, the present invention provides a remote sensing image information extraction system and method based on case reasoning coupled with knowledge graph.

[0019] The present invention is implemented as follows: a remote sensing image information extraction method of case-based reasoning coupled with knowledge graph includes:

[0020] Step 1: Collect high-resolution images of the study area, geographical conditions monitoring results, field survey data, land use maps and various socio-economic statistical data, and form a basic data set of the study area through image radiometric calibration, atmospheric correction and geometric correction preprocessing. Collect business knowledge on interpretation of natural resource elements, organize it into a knowledge graph, and then complete and correct the knowledge through knowledge reasoning;

[0021] Step 2: After segmenting the high-resolution image and optimizing the segmentation results, an optimized image object is obtained. Typical natural resource element samples are manually selected as case objects, and natural resource element sample cases are formed after feature selection and weight allocation;

[0022] Step 3: After selecting a certain number of natural resource element cases from different sensors with different spatiotemporal environments and different types, integrate them into the interpretation knowledge graph, organize various sample cases based on the improved GHSOM algorithm to form a spatiotemporal case library of natural resource elements, and maintain the case library;

[0023] Step 4: The high-resolution image to be interpreted is segmented and feature-assigned to form a target case, which is then searched and matched with the case library. The retrieved case type is the natural resource element type corresponding to the target case, and is added to the case library as a new case through case learning. The extracted result is the intelligent interpretation result of the natural resource element after accuracy verification.

[0024] Furthermore, the case objects include:

[0025] Case number: number each image object case to ensure that the number value is unique;

[0026] Case characteristics: corresponding to the "Problem Description" section, including imaging characteristics and non-imaging characteristics;

[0027] Case category: Corresponding to the "Solution" section, the natural resource element classification system refers to the land cover classification system in the "Basic Geographic National Condition Monitoring Content and Indicators (CHT 9029-2019)" to classify the natural resource elements of the image objects;

[0028] Case effect: corresponds to the "Solution effect" part, describing the role and effect of the case in the process of obtaining the "Solution" from the "Problem Description", including whether it is a representative case, the number of case retrievals, and the number of successful matches;

[0029] The case model consists of case matching features, case retrieval features, case categories and case effects. A specific case can be expressed as formula (1):

[0030] Case = {match 1，2，...，n , retrieval 1，2，...，8 ; class, subclass; effect 1，2，3} (1)

[0031] In the formula, Case represents a specific case, match is the case matching feature, retrieval is the case retrieval feature, class is the surface coverage classification code to which the case belongs, subclass is the subclass code of the surface coverage classification to which the case belongs, and effect is the case effect.

[0032] Further, the feature selection method is as follows:

[0033] Phase 1: The improved ReliefF algorithm is used to calculate and sort feature weights, and n feature combinations with strong classification capabilities are selected to participate in Wrapper optimization to achieve the purpose of rapid dimensionality reduction and reduce particle search space; based on the ReliefF algorithm, feature weights are calculated based on all training samples, and the total number of samples is set as the number of iterations. The feature weights within the class are first summed and averaged, and then the feature averages of each class are summed and averaged to obtain the final feature weights;

[0034] The improved ReliefF algorithm weight calculation formula is shown in formula (2) and formula (3):

[0035]

[0036] Among them, x is the category, x i is the sample in this category, n is the number of categories, D f (i, j) is the standard Euclidean distance of feature f, D f (x i ,H(x i )) is x i The distance to the nearest similar sample, D f (x i ,M(x i )) is the distance between the nearest samples of different classes, is the number of samples of category x;

[0037] The second stage: Use the PSO encapsulation algorithm to measure the importance of each feature in the pre-selected feature subset, further eliminate redundant and irrelevant features, and retain m features that contribute to classification; PSO is initialized as a group of random particles. Through the collaboration and information sharing between individuals in the group, the movement of the entire group evolves from disorder to order in the problem-solving space, thereby obtaining the optimal solution; the d-dimensional space speed and position update formula of particle i is

[0038] The formula is formula (4) and formula (5):

[0039]

[0040] in, is the d-th dimension component of the flight velocity vector of particle i in the k-th iteration; is the position vector; Pid represents the dth dimension of the individual extreme value of particle i,

[0041] It is also the local optimum; P gd is the best position of the entire particle swarm, indicating the dth dimension of the global optimal solution; r1 and r2 are random functions between 0 and 1, ω is the inertia weight, and c1 and c2 are acceleration coefficients;

[0042] The present invention uses PSO as a search strategy, comprehensively considers the characteristics of the target object and the advantages and disadvantages of each classification algorithm, combines the multi-level segmentation classification model, selects the random forest algorithm and the J48 decision tree algorithm to form the PSO=RF and PSO_J48 encapsulation algorithms respectively, and applies them to the image classification feature selection at the corresponding level;

[0043] The third stage: according to the order of feature importance, the random forest algorithm is combined with the sequential backward selection method SBS. The last feature in the importance order is deleted from the feature set each time, and the operation is iterated repeatedly to determine the relationship between feature dimension and classification accuracy and find the optimal feature subset.

[0044] Furthermore, the weight allocation method is as follows:

[0045] 1) Weight initialization and iterative update of weights

[0046] Let the feature weight in the iteration process be ω(t), and the update formula is formula (6):

[0047] ω(t)=[0, (2r(t1)-1), cumsum(2r(t2)-1),..., cumsum(2r(t) m )-1)] (6)

[0048] Among them, cumsum means calculating the cumulative sum, m means the maximum number of iterations (also m features), t1:t m The maximum number of iterations for each column corresponding to the feature weight is 1 to m, r(t i ) is a random function, defined as formula (7):

[0049]

[0050] Among them, rand is used to generate random numbers with uniform distribution in the interval [0, 1];

[0051] Since the weight is constrained by the boundary range, the feature weight is constrained using formula (8):

[0052]

[0053] Among them, a i is the minimum weight among the initialization weights, b i is the maximum value of the weight in the initialization weight, is the minimum value of the ith feature variable at the tth iteration, is the maximum value of the i-th characteristic variable in the t-th iteration;

[0054] The sum of the jth weight of the tth iteration and the maximum values ​​of all variables in the tth iteration is the maximum value of the ith characteristic variable of the tth iteration. The maximum value of the ith characteristic variable of the tth iteration is defined as follows:

[0055]

[0056] Among them, c t is the minimum value of all feature variables in the tth iteration, d t is the maximum value of all feature variables in the tth iteration, represents the jth feature weight selected at the tth iteration, that is, the optimal feature weight remaining after removing the i-th feature weight in a set of weights;

[0057] In order to obtain the maximum value of the characteristic variable more quickly, the search range of the characteristic variable is continuously narrowed during the iteration process, and the formula (9) is constrained to obtain the formula (10):

[0058] in

[0059] where w is a constant defined based on the current iteration (when t>0.1m, w=2; when t>0.5m, w=3; when t>0.75m, w=4; when t>0.9m, w=5; when t>0.95m, w=6;

[0060] 2) Calculate the fitness function

[0061] The fitness function for evaluating weight optimization is the accuracy of the classifier, as shown in formula (11), where g is the number of correctly classified cases in the test set, and p′ is the total number of cases in the test set;

[0062]

[0063] 3) Selecting feature variables using the roulette method

[0064] In the weight optimization process of the Ant Lion Algorithm, according to the fitness value corresponding to the weight, the roulette method is used to select the feature variable with a high fitness value for the next iteration;

[0065] 4) Update feature weights and obtain optimal feature weights

[0066] In each iteration, an optimal weight is obtained. When the optimal weight appears, the optimal weight is defined as the current value according to formula (12):

[0067]

[0068] in, represents the i-th feature weight selected at the t-th iteration, as shown in formula (13):

[0069]

[0070] in, represents the weight of the feature variable selected by roulette at the tth iteration, represents the optimal value among all weight combinations at the tth iteration;

[0071] When the last iteration is completed, m optimal feature weights are determined, and the feature weight allocation process is completed;

[0072] Success-driven weight introspective learning

[0073] From the perspective of cognitive science, the present invention draws on the principle of introspective learning and adjusts the weight of attributes by monitoring the system's own reasoning process and performance changes to achieve the goal of giving CBR active learning capabilities. In the same case library, suppose there are two cases, C1=(X I ; Y I ) and C II =(X II ; Y II ), where X I =(X I1 , X I2 , ..., X IK , ..., X Im ), X I =(X II1 , X II2 , ..., X IIK , ..., X IIm ), f match To set the threshold, used to determine whether the attribute matches. If the condition 0≤|X is met Ii -X IIi |≤f match When , the i-th attribute of the two cases is called a matching attribute; if it does not meet this condition, it is called a non-matching attribute;

[0074] When a case is successfully matched and correctly classified, the weight value of the matching attribute of the case increases, and the weight value of the unmatched attribute decreases accordingly; when the weight value of the case matching attribute increases, the weight is calculated according to formula (14):

[0075]

[0076]

[0077] In formula (14) and formula (15), w i(t) represents the weight of the tth iteration of the i-th attribute of case; w i (t+1) is the process weight of the t+1th iteration of the attribute of the case; represents the change in weight, and m represents the number of case attributes.

[0078] Further, the case library searches:

[0079] On the basis of improving the GHSOM algorithm to cluster the case library, a case retrieval model based on RBF network is established; the network has n input nodes, representing n characteristic attributes of the case; generally, the characteristic attributes should be normalized so that their input values ​​are between 0 and 1;

[0080] Hidden layer unit: The selection of RBF network center is completed by self-organization method, that is, the case library clustering result of improved GHSOM algorithm is used; the number of hidden layer center nodes is equal to the sum of the number of clusters and the number of valuable outliers;

[0081] Output layer O: has m nodes, representing m different cases. The output of the network corresponds to the index or similarity value of similar cases.

[0082] Retrieval machine structure model based on RBF network;

[0083] Assume that the network has Nr hidden layer nodes. If the output layer transfer function is f(x)=x, then the RBF retrieval machine with n inputs and m outputs can realize the functions as shown in equations (16) and (17):

[0084]

[0085]

[0086] And the network satisfies the minimum sum of squared output errors, as shown in formula (18):

[0087]

[0088] Further, the case study:

[0089] Suppose the new case generated is C * =(F * , FS * ), firstly, we obtain new cases C according to the case retrieval strategy * Similarity with all cases in the case library λ1 and λ2 are thresholds satisfying the relationship of 0.5<λ2<λ1≤1. Their specific values ​​are determined based on actual experience or experts, and can be λ1=0.95 and λ2=0.85;

[0090] If SIM max <λ2, indicating that there is no case library with C* If the case is matched, C * Stored in the cache case library;

[0091] If λ2≤SIM max ≤λ1, indicating that there is a case library with C * A case that matches well, but not very well. Suppose the case library has SIM max The case is C i =(F i , FS i ), calculation case C * With Case C i The similarity SIM(C i , C * ), if SIM(C i , C * )≥λ2, then C i Using C * Replace, otherwise case C * Store in case library; if SIM max ≥λ1, indicating that there is already a case library with C * A very good match, so C * Not to be stored;

[0092] When the number of cases in the cache case library reaches a certain number, the cache case library processing mechanism is triggered, the cache case library is clustered using the adaptive FCM algorithm and noise cases are deleted, and the processed cases are stored in the case library;

[0093] Under the case learning strategy based on the cache mechanism, new cases will be automatically and continuously added to the existing case library as the CBR system runs.

[0094] Another object of the present invention is to provide a remote sensing image information extraction system for case-based reasoning coupled with knowledge graph, comprising:

[0095] The collection module is used to collect high-resolution images of the study area, geographical national conditions monitoring results, field survey data, land use maps and various socio-economic statistical data. After image radiometric calibration, atmospheric correction and geometric correction preprocessing, the basic data set of the study area is formed. The business knowledge of natural resource element interpretation is collected and organized into a knowledge graph, and then knowledge is supplemented and error corrected through knowledge reasoning.

[0096] The optimization processing module is used to segment the high-resolution image and optimize the segmentation results to obtain the optimized image object, manually select typical natural resource element samples as case objects, and form natural resource element sample cases after feature selection and weight allocation;

[0097] The integration module is used to select a certain number of natural resource element cases from different sensors with different spatiotemporal environments and different types, integrate them into the interpretation knowledge graph, organize various sample cases based on the improved GHSOM algorithm to form a spatiotemporal case library of natural resource elements, and maintain the case library;

[0098] The extraction module is used to form a target case after image segmentation and feature assignment of the high-resolution image to be interpreted, and then search and match it with the case library. The case type retrieved is the natural resource element type corresponding to the target case, and is added to the case library as a new case through case learning. The extraction result is the intelligent interpretation result of the natural resource element after accuracy verification.

[0099] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the remote sensing image information extraction method of case reasoning coupled with knowledge graph.

[0100] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the remote sensing image information extraction method of case reasoning coupled with knowledge graph.

[0101] Another object of the present invention is to provide an information data processing terminal, which is used to implement the remote sensing image information extraction system of the case reasoning coupled knowledge graph.

[0102] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0103] First, the present invention applies knowledge graph semantic network and case reasoning theory to study the construction and reasoning of natural resource element interpretation knowledge graph, the construction and maintenance of spatiotemporal case library, and the reasoning and learning methods of cases, and establishes a high-resolution image natural resource element remote sensing intelligent interpretation paradigm that couples knowledge graph and case reasoning, to achieve rapid extraction and intelligent interpretation of natural resource elements under the guidance of human-computer collaboration, and to provide a new solution for realizing the intelligent extraction of high-resolution image information, which has great application value and practical significance for enhancing the ability to obtain spatial information of natural resource elements.

[0104] The present invention is based on the knowledge graph and case reasoning theory in the field of artificial intelligence. It uses the image object formed after high-resolution image segmentation as the interpretation unit, uses the interpretation business knowledge and image features to construct the knowledge graph and spatiotemporal case library, establishes the high-resolution image natural resource element remote sensing intelligent interpretation paradigm that couples the knowledge graph and case reasoning, and realizes the human-computer collaborative intelligent interpretation of natural resource elements under the guidance of human-computer interaction. This method can make full use of the interpretation business knowledge, automatically expand the case library and adaptively adjust the case feature weights under the incremental learning mechanism of case reasoning, reduce the workload of manual annotation of case samples and improve the accuracy of automatic interpretation. It is a useful exploration of the intelligent interpretation method of high-resolution remote sensing images, and is of great significance to ensuring national ecological security and serving the unified monitoring of natural resources.

[0105] (1) A knowledge-driven high-resolution image spatiotemporal case expression model for natural resource elements is constructed. Most existing studies use the image features of remote sensing images themselves, but the high-level features obtained by deep learning methods are not very interpretable. The present invention considers both the image features and non-image features of remote sensing image objects, and uses interpretation business knowledge to construct a knowledge graph, and then constructs a knowledge-driven natural resource element spatiotemporal case expression model, which can more comprehensively express the features of different types of natural resource element objects, which is one of the characteristics and innovations of the present invention.

[0106] (2) A knowledge graph of natural resource elements in high-resolution images and a coupled intelligent interpretation paradigm of case-based reasoning have been established. Existing intelligent extraction methods for natural resource elements in remote sensing images are insufficiently knowledge-driven, have weak reasoning capabilities and interpretability, and require a huge workload for manual labeling of samples. The knowledge graph of natural resource elements and a coupled intelligent interpretation paradigm of case-based reasoning established in the present invention can fully utilize interpretation business knowledge, automatically expand the case library, adaptively adjust case feature weights, and continuously improve automatic interpretation accuracy, which is also one of the characteristics and innovations of the present invention.

[0107] Second, the expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0108] After transformation, the technical solution of the present invention has broad application prospects and huge commercial value in the fields of natural resource monitoring, smart city construction, etc. This technical solution can achieve accurate automatic interpretation of high-resolution images through the organic combination of knowledge graph and machine learning, greatly improving analysis efficiency and decision support capabilities, and has the following main expected benefits and commercial value:

[0109] ① Improve the ability to monitor natural resources: It can accurately identify and extract key elements such as vegetation, water bodies, and buildings in images, greatly improving the timeliness and accuracy of natural resource surveys and monitoring.

[0110] ②Support the construction of smart cities: It can quickly identify the current status of urban land use, road networks, greening conditions, etc., and provide strong support for urban planning, construction and management.

[0111] ③ Enhance emergency response capabilities: It can quickly analyze image information after a disaster occurs, accurately identify the disaster situation, and provide decision-making basis for emergency rescue and post-disaster reconstruction.

[0112] ④Optimize industry operation management: In the fields of forestry, agriculture, water conservancy, etc., this technology can be applied to reduce manual survey costs and improve the timeliness and accuracy of resource surveys.

[0113] ⑤ Drive the development of the industrial chain: This technical solution can promote the innovative development of related industries such as geographic information services and image analysis software, and form a complete ecosystem.

[0114] In general, this technical solution has broad application prospects and good commercial value, and is expected to become the key technical support for the new generation of natural resource remote sensing intelligent services. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] Figure 1 It is a flow chart of a remote sensing image information extraction method of case reasoning coupled with knowledge graph provided in an embodiment of the present invention.

[0116] Figure 2 It is a structural block diagram of a remote sensing image information extraction system of case reasoning coupled with a knowledge graph provided in an embodiment of the present invention.

[0117] Figure 3 It is a diagram of the main research contents and their interrelationships provided by the embodiments of the present invention.

[0118] Figure 4 It is the overall technical roadmap provided by the embodiments of the present invention.

[0119] Figure 5 It is a high-resolution remote sensing image multi-scale segmentation optimization diagram that takes into account global and local optimization, provided by an embodiment of the present invention.

[0120] Figure 6 It is a natural resource element interpretation knowledge graph (taking rivers as an example) provided by an embodiment of the present invention.

[0121] Figure 7 It is a flow chart of the improved ReliefF-PSO three-stage combined feature selection algorithm provided by an embodiment of the present invention.

[0122] Figure 8 It is a case organization chart based on the GHSOM algorithm provided in an embodiment of the present invention.

[0123] Fig. 9 It is a technical roadmap for simplifying the case library provided by an embodiment of the present invention.

[0124] Fig.10 It is a diagram of the RBF case retrieval structure model provided by an embodiment of the present invention.

[0125] Fig.11 It is a case study flow chart based on a cache mechanism provided by an embodiment of the present invention.

[0126] Fig.12 It is a surface cover information extraction verification map based on case reasoning provided by an embodiment of the present invention.

[0127] Fig.13 This is a partial surface cover classification case diagram provided by an embodiment of the present invention.

[0128] Fig.14 This is a case metadata diagram provided by an embodiment of the present invention.

[0129] Fig.15 This is a diagram of a segmentation parameter training interface provided by an embodiment of the present invention.

[0130] Fig.16 It is a case spectrum feature weight curve diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0131] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0132] like Figure 1 As shown, a remote sensing image information extraction method based on case reasoning coupled with a knowledge graph provided by an embodiment of the present invention comprises the following steps:

[0133] S101: Collect high-resolution images of the study area, geographical conditions monitoring results, field survey data, land use maps and various socio-economic statistics, and form a basic data set of the study area through image radiometric calibration, atmospheric correction and geometric correction preprocessing. Collect business knowledge on interpretation of natural resource elements, organize it into a knowledge graph, and then complete and correct the knowledge through knowledge reasoning;

[0134] S102, segmenting the high-resolution image and optimizing the segmentation result to obtain an optimized image object, manually selecting typical natural resource element samples as case objects, and forming natural resource element sample cases after feature selection and weight allocation;

[0135] S103, after selecting a certain number of natural resource element cases from different sensors with different spatiotemporal environments and different types, integrating them into the interpretation knowledge graph, organizing various sample cases based on the improved GHSOM algorithm to form a natural resource element spatiotemporal case library, and maintaining the case library;

[0136] S104, the high-resolution image to be interpreted is subjected to image segmentation and feature assignment to form a target case, which is then searched and matched with the case library. The case type retrieved is the natural resource element type corresponding to the target case, and is added to the case library as a new case through case learning. The extracted result is the intelligent interpretation result of the natural resource element after passing the accuracy verification.

[0137] The first step of the method of the present invention is to collect remote sensing image data and geographic information related to the study area, including high-resolution images, geographical national conditions monitoring results, field survey data, land use maps and socio-economic statistics. These data are pre-processed by image radiation calibration, atmospheric correction and geometric correction to form a basic data set for the study area. In addition, business knowledge related to the interpretation of natural resource elements is collected and organized into a knowledge graph, and knowledge is supplemented and corrected through knowledge reasoning to ensure the integrity and accuracy of the data.

[0138] After the data set is prepared, the high-resolution image is segmented and the segmentation results are optimized to obtain more accurate image objects. Then, typical natural resource elements are manually selected as case samples, and natural resource element sample cases are constructed through feature selection and weight allocation. These sample cases contain image features, spatial features, and unique attributes of specific natural resource elements, laying the foundation for the construction of the subsequent case library.

[0139] A variety of natural resource element cases are selected from different sensors and different time and space environments and integrated into the interpretation knowledge graph. The improved GHSOM algorithm is used to organize these sample cases to build a spatiotemporal case library of natural resource elements covering multiple time and space and multiple types. The case types and attributes in the case library are constantly updated and maintained to ensure that they can reflect the latest data features and interpretation requirements at any time.

[0140] The high-resolution images to be interpreted are segmented and assigned features to form target cases, which are then searched and matched with the case library. Through case matching, the natural resource element type of the target case is determined, and the matching results are added to the case library as new cases to achieve case learning and library updates. Finally, the interpretation results are verified for accuracy, and once they meet the qualified standards, they are the intelligent interpretation results of natural resource elements, providing effective support for the accurate interpretation of remote sensing images.

[0141] The case objects provided by the embodiments of the present invention include:

[0142] Case number: number each image object case to ensure that the number value is unique;

[0143] Case characteristics: corresponding to the "Problem Description" section, including imaging characteristics and non-imaging characteristics;

[0144] Case category: Corresponding to the "Solution" section, the natural resource element classification system refers to the land cover classification system in the "Basic Geographic National Condition Monitoring Content and Indicators (CHT 9029-2019)" to classify the natural resource elements of the image objects;

[0145] Case effect: corresponds to the "Solution effect" part, describing the role and effect of the case in the process of obtaining the "Solution" from the "Problem Description", including whether it is a representative case, the number of case retrievals, and the number of successful matches;

[0146] The case model consists of case matching features, case retrieval features, case categories and case effects. A specific case can be expressed as formula (1):

[0147] Case = {match 1，2，...，n , retrieval 1，2，...，8 ;class,subclass;effect 1，2，3} (1)

[0148] In the formula, Case represents a specific case, match is the case matching feature, retrieval is the case retrieval feature, class is the surface coverage classification code to which the case belongs, subclass is the subclass code of the surface coverage classification to which the case belongs, and effect is the case effect.

[0149] The feature selection method provided by the embodiment of the present invention is as follows:

[0150] Phase 1: The improved ReliefF algorithm is used to calculate and sort feature weights, and n feature combinations with strong classification capabilities are selected to participate in Wrapper optimization to achieve the purpose of rapid dimensionality reduction and reduce particle search space; based on the ReliefF algorithm, feature weights are calculated based on all training samples, and the total number of samples is set as the number of iterations. The feature weights within the class are first summed and averaged, and then the feature averages of each class are summed and averaged to obtain the final feature weights;

[0151] The improved Relief F algorithm weight calculation formula is shown in formula (2) and formula (3):

[0152]

[0153] Among them, x is the category, x iis the sample in this category, n is the number of categories, D f (i, j) is the standard Euclidean distance of feature f, D f (x i ,H(x i )) is x i The distance to the nearest similar sample, D f (x i ,M(x i )) is the distance between the nearest samples of different classes, is the number of samples of category x;

[0154] The second stage: Use the PSO encapsulation algorithm to measure the importance of each feature in the pre-selected feature subset, further eliminate redundant and irrelevant features, and retain m features that contribute to classification; PSO is initialized as a group of random particles. Through the collaboration and information sharing between individuals in the group, the movement of the entire group evolves from disorder to order in the problem-solving space, thereby obtaining the optimal solution; the d-dimensional space speed and position update formula of particle i is

[0155] The formula is formula (4) and formula (5):

[0156]

[0157] in, is the d-th dimension component of the flight velocity vector of particle i in the k-th iteration; is the position vector; P id The dth dimension represents the individual extreme value of particle i, which is also the local optimal value; P gd is the best position of the entire particle swarm, indicating the dth dimension of the global optimal solution; r1 and r2 are random functions between 0 and 1, ω is the inertia weight, and c1 and c2 are acceleration coefficients;

[0158] The present invention uses PSO as a search strategy, comprehensively considers the characteristics of the target object and the advantages and disadvantages of each classification algorithm, combines the multi-level segmentation classification model, selects the random forest algorithm and the J48 decision tree algorithm to form the PSO_RF and PSO_J48 encapsulation algorithms respectively, and applies them to the image classification feature selection at the corresponding level;

[0159] The third stage: according to the order of feature importance, the random forest algorithm is combined with the sequential backward selection method SBS. The last feature in the importance order is deleted from the feature set each time, and the operation is iterated repeatedly to determine the relationship between feature dimension and classification accuracy and find the optimal feature subset.

[0160] The weight allocation method provided by the embodiment of the present invention is as follows:

[0161] 1) Weight initialization and iterative update of weights

[0162] Let the feature weight in the iteration process be ω(t), and the update formula is formula (6):

[0163] ω(t)=[0,(2r(t1)-1), cumsum(2r(t2)-1),..., cumsum(2r(t) m )-1)] (6)

[0164] Among them, cumsum means calculating the cumulative sum, m means the maximum number of iterations (also m features), t1:t m The maximum number of iterations for each column corresponding to the feature weight is 1 to m, r(t i ) is a random function, defined as formula (7):

[0165]

[0166] Among them, rand is used to generate random numbers with uniform distribution in the interval [0, 1];

[0167] Since the weight is constrained by the boundary range, the feature weight is constrained using formula (8):

[0168]

[0169] Among them, a i is the minimum weight among the initialization weights, b i is the maximum value of the weight in the initialization weight, is the minimum value of the ith feature variable at the tth iteration, is the maximum value of the i-th characteristic variable in the t-th iteration;

[0170] The sum of the jth weight of the tth iteration and the maximum values ​​of all variables in the tth iteration is the maximum value of the ith characteristic variable of the tth iteration. The maximum value of the ith characteristic variable of the tth iteration is defined as follows:

[0171]

[0172] Among them, c t is the minimum value of all feature variables in the tth iteration, d t is the maximum value of all feature variables in the tth iteration, represents the jth feature weight selected at the tth iteration, that is, the optimal feature weight remaining after removing the i-th feature weight in a set of weights;

[0173] In order to obtain the maximum value of the characteristic variable more quickly, the search range of the characteristic variable is continuously narrowed during the iteration process, and the formula (9) is constrained to obtain the formula (10):

[0174] in

[0175] Where w is a constant defined based on the current iteration (when t>0.1m, w=2; when t>0.5m, w=3; when t>0.75m, w=4; when t>0.9m, w=5; when t>0.95m, w=6);

[0176] 2) Calculate the fitness function

[0177] The fitness function for evaluating weight optimization is the accuracy of the classifier, as shown in formula (11), where g is the number of correctly classified cases in the test set, and p′ is the total number of cases in the test set;

[0178]

[0179] 3) Selecting feature variables using the roulette method

[0180] In the weight optimization process of the Ant Lion Algorithm, according to the fitness value corresponding to the weight, the roulette method is used to select the feature variable with a high fitness value for the next iteration;

[0181] 4) Update feature weights and obtain optimal feature weights

[0182] In each iteration, an optimal weight is obtained. When the optimal weight appears, the optimal weight is defined as the current value according to formula (12):

[0183]

[0184] in, represents the i-th feature weight selected at the t-th iteration, as shown in formula (13):

[0185]

[0186] in, represents the weight of the feature variable selected by roulette at the tth iteration, represents the optimal value among all weight combinations at the tth iteration;

[0187] When the last iteration is completed, m optimal feature weights are determined, and the feature weight allocation process is completed;

[0188] Success-driven weight introspective learning

[0189] From the perspective of cognitive science, the present invention draws on the principle of introspective learning and adjusts the weight of attributes by monitoring the system's own reasoning process and performance changes to achieve the goal of giving CBR active learning capabilities. In the same case library, suppose there are two cases, CI =(X I ; Y I ) and C II =(X II ; Y II ), where X I =(X I1 , X I2 , ..., X IK , ..., X Im ), X I =(X II1 , X II2 , ..., X IIK , ..., X IIm ), f match To set the threshold, used to determine whether the attribute matches. If the condition 0≤|X is met Ii -X IIi |≤f match When , the i-th attribute of the two cases is called a matching attribute; if it does not meet this condition, it is called a non-matching attribute;

[0190] When a case is successfully matched and correctly classified, the weight value of the matching attribute of the case increases, and the weight value of the unmatched attribute decreases accordingly; when the weight value of the case matching attribute increases, the weight is calculated according to formula (14):

[0191]

[0192]

[0193] In formula (14) and formula (15), w i (t) represents the weight of the tth iteration of the i-th attribute of case; w i (t+1) is the process weight of the t+1th iteration of the attribute of the case; represents the change in weight, and m represents the number of case attributes.

[0194] The case library provided in the embodiment of the present invention is searched:

[0195] On the basis of improving the GHSOM algorithm to cluster the case library, a case retrieval model based on RBF network is established; the network has n input nodes, representing n characteristic attributes of the case; generally, the characteristic attributes should be normalized so that their input values ​​are between 0 and 1;

[0196] Hidden layer unit: The selection of RBF network center is completed by self-organization method, that is, the case library clustering result of improved GHSOM algorithm is used; the number of hidden layer center nodes is equal to the sum of the number of clusters and the number of valuable outliers;

[0197] Output layer O: has m nodes, representing m different cases. The output of the network corresponds to the index or similarity value of similar cases.

[0198] Retrieval machine structure model based on RBF network;

[0199] Assume that the network has N r hidden layer nodes, if the output layer transfer function is taken as f(x) = x, then the RBF retrieval machine with n inputs and m outputs implements the functions as shown in equations (16) and (17):

[0200]

[0201]

[0202] And the network satisfies the minimum sum of squared output errors, as shown in formula (18):

[0203]

[0204] Case study provided by the embodiment of the present invention:

[0205] Suppose the new case generated is C * =(F * , FS * ), firstly, we obtain new cases C according to the case retrieval strategy * Similarity with all cases in the case library λ1 and λ2 are thresholds satisfying the relationship of 0.5<λ2<λ1≤1. Their specific values ​​are determined based on actual experience or experts, and can be λ1=0.95 and λ2=0.85;

[0206] If SIM max <λ2, indicating that there is no case library with C * If the case is matched, C * Store in the cache case library;

[0207] If λ2≤SIM max ≤λ1, indicating that there is a case library with C * A case that matches well, but not very well. Suppose the case library has SIM max The case is C i =(F i , FS i ), calculation case C * With Case C i The similarity SIM(C i , C * ), if SIM(C i , C * )≥λ2, then C i Using C *Replace, otherwise case C * Store in case library; if SIM max ≥λ1, indicating that there is already a case library with C * A very good match, so C * Not to be stored;

[0208] When the number of cases in the cache case library reaches a certain number, the cache case library processing mechanism is triggered, the cache case library is clustered using the adaptive FCM algorithm and noise cases are deleted, and the processed cases are stored in the case library;

[0209] Under the case learning strategy based on the cache mechanism, new cases will be automatically and continuously added to the existing case library as the CBR system runs.

[0210] like Figure 2 As shown, a remote sensing image information extraction system of case reasoning coupled with knowledge graph provided by an embodiment of the present invention includes:

[0211] The collection module is used to collect high-resolution images of the study area, geographical national conditions monitoring results, field survey data, land use maps and various socio-economic statistical data. After image radiometric calibration, atmospheric correction and geometric correction preprocessing, the basic data set of the study area is formed. The business knowledge of natural resource element interpretation is collected and organized into a knowledge graph, and then knowledge is supplemented and error corrected through knowledge reasoning.

[0212] The optimization processing module is used to segment the high-resolution image and optimize the segmentation results to obtain the optimized image object, manually select typical natural resource element samples as case objects, and form natural resource element sample cases after feature selection and weight allocation;

[0213] The integration module is used to select a certain number of natural resource element cases from different sensors with different spatiotemporal environments and different types, integrate them into the interpretation knowledge graph, organize various sample cases based on the improved GHSOM algorithm to form a spatiotemporal case library of natural resource elements, and maintain the case library;

[0214] The extraction module is used to form a target case after image segmentation and feature assignment of the high-resolution image to be interpreted, and then search and match it with the case library. The case type retrieved is the natural resource element type corresponding to the target case, and is added to the case library as a new case through case learning. The extraction result is the intelligent interpretation result of the natural resource element after accuracy verification.

[0215] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the remote sensing image information extraction method of case reasoning coupled with knowledge graph.

[0216] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the remote sensing image information extraction method of case reasoning coupled with knowledge graph.

[0217] Another object of the present invention is to provide an information data processing terminal, which is used to implement the remote sensing image information extraction system of the case reasoning coupled knowledge graph.

[0218] The present invention is specifically implemented:

[0219] 1. Based on case-based reasoning, this invention studies the spatiotemporal case expression model of natural resource element objects from high-resolution images with different spatiotemporal environments from different sensors, constructs a knowledge graph for natural resource element interpretation and a sample spatiotemporal case library, builds a knowledge-driven intelligent learning model based on case-based reasoning, establishes a high-resolution image natural resource element remote sensing intelligent interpretation paradigm that couples knowledge graphs and case-based reasoning, realizes human-computer collaborative intelligent interpretation of natural resource elements under the guidance of human-computer interaction, and verifies the method. The specific research content includes the following three aspects. The relationship between the various parts in the research content is as follows: Figure 3 shown.

[0220] (1) Spatiotemporal case model expression of natural resource element samples

[0221] This part is the basic theoretical research of this invention, mainly studying the case object selection, case structure composition, case feature selection and weight allocation of natural resource elements from high-resolution images of different sensors with different spatiotemporal environments. The specific contents include:

[0222] Case object selection: The study used a multi-scale segmentation result optimization method that takes into account global and local optimality to perform multi-scale segmentation and result optimization on high-resolution images, and manually selected "pure" natural resource element object samples as case objects.

[0223] Case structure composition: Make full use of the image characteristics of natural resource elements and interpretation business knowledge, and use the ternary structure of "problem description-solution-solution effect" to express the natural resource element case model.

[0224] Case feature selection: The study adopted the improved ReliefF-PSO three-stage combined feature selection algorithm and the feature weight allocation strategy based on the ant lion algorithm to determine the case characteristics of different types of natural resource elements.

[0225] (2) Construction of knowledge graph for interpretation of natural resource elements and sample spatiotemporal case library This part is the key technical support of this invention, and mainly studies the construction of knowledge graph for interpretation of natural resource elements and the organization and maintenance of sample spatiotemporal case library. The specific contents include:

[0226] Knowledge graph construction: Collect business knowledge on interpretation of natural resource elements to construct a knowledge graph on interpretation of natural resource elements. Research the use of knowledge reasoning to complete and correct the interpretation business knowledge for the construction of a spatiotemporal case library of natural resource elements.

[0227] Case library organization: The main research is to use the improved dynamic hierarchical self-organizing algorithm (GHSOM) strategy to organize and index the case library, so as to effectively improve the accuracy and efficiency of retrieval.

[0228] Case library maintenance: The main research is on case library streamlining strategies to solve the problem of data redundancy that may occur due to the expansion of case library size, as well as feature weight introspection learning mechanisms to give CBR active learning capabilities, so that the CBR system can dynamically adjust the weights of case features according to changes in reasoning process and performance.

[0229] (3) Implementation of a human-machine collaborative intelligent interpretation system for natural resource elements

[0230] This part of the research is the core goal of this invention, mainly studying case retrieval models, case learning strategies and intelligent interpretation implementation. The specific research contents include:

[0231] Case retrieval model: mainly studies the similar case retrieval model based on the radial basis function (RBF) network to improve the retrieval efficiency and realize adaptive retrieval of similar cases.

[0232] Case learning strategy: mainly studies case learning strategies based on caching mechanisms to automatically expand the scale of existing case libraries and adapt to changes in new cases.

[0233] Implementation of intelligent interpretation: The main research is to realize the integration of natural resource element intelligent interpretation algorithms based on the "Python+PyQGIS" open source system and the human-computer collaborative intelligent interpretation experiment of natural resource elements to verify the effectiveness and practicality of the method of the present invention.

[0234] 2 Research objectives

[0235] The present invention takes high-resolution images such as WorldView-2, GF-2, SWDC-5 and UAV images in the study area as target images, constructs a knowledge-driven intelligent learning model based on the natural resource element interpretation knowledge graph and case reasoning, establishes a high-resolution image natural resource element remote sensing intelligent interpretation paradigm that couples knowledge graph and case reasoning, realizes the intelligent interpretation of natural resource elements under the guidance of human-computer interaction, provides a new solution for realizing the intelligent extraction of high-resolution image information, and promotes the rapid extraction and intelligent interpretation of all-factor information of natural resources in natural resource survey and monitoring results.

[0236] The overall research method of the present invention includes three steps: case model expression, knowledge graph and case library construction, and intelligent interpretation implementation. First, collect high-resolution images of the study area, geographical national conditions monitoring results, field survey materials, land use maps, and various socio-economic statistics. After pre-processing such as image radiation calibration, atmospheric correction, and geometric correction, form a basic data set for the study area. Collect business knowledge on natural resource element interpretation, organize it into a knowledge graph, and then perform knowledge completion and error correction through knowledge reasoning. Then, segment the high-resolution image and optimize the segmentation results to obtain an optimized image object. Manually select typical natural resource element samples as case objects, and form natural resource element sample cases after feature selection and weight allocation. Next, select After a certain number of natural resource element cases with different types and in different spatiotemporal environments from different sensors are integrated into the interpretation knowledge graph, various sample cases are organized based on the improved GHSOM algorithm to form a spatiotemporal case library of natural resource elements, and the case library is maintained; finally, the high-resolution image to be interpreted is formed into a target case after image segmentation and feature assignment, and then searched and matched with the case library. The retrieved case type is the natural resource element type corresponding to the target case, and is added to the case library as a new case through case learning. The extraction result is the intelligent interpretation result of natural resource elements after the accuracy verification. The specific technical route is as follows: Figure 4 shown.

[0237] Guangxi is located in the southeastern edge of the Yunnan-Guizhou Plateau in the second step of China's terrain, and in the western part of the Guangdong-Guangxi hills. It is generally a mountainous and hilly basin landform, known as the "Guangxi Basin". There are many hilly plains in the central and southern parts, and more mountains and less land are the main characteristics of Guangxi's land resources. Considering the influence of topographic factors on the distribution of natural resource elements, especially cultivated land elements, this paper selects Wuming District of Nanning City, Gangbei District of Guigang City, Qinnan District of Qinzhou City and Bobai County of Yulin City, which are located in the central and southern parts of Guangxi and have relatively flat terrain, as the research area.

[0238] Through field investigation, platform download and procurement, the relevant data of the experimental area are systematically collected to form a basic data database. ① Collection of satellite remote sensing images: WV-2 satellite images, GF-2 domestic high-resolution images, SWDC-5 aerial images and drone images in the experimental area are collected. This data is the main data source for feature selection and interpretation of natural resource elements; ② Field investigation and measurement: GPS is used to locate the sample points of natural resource elements in the experimental area and take photos on the spot. This data can be used as case samples and accuracy verification samples; ③ Other data: land use maps of the experimental area in various periods, local conditions monitoring data and various social and economic statistics are collected. These data can assist in the collection of sample data and verify the reliability of the results of this study.

[0239] Since there are objects of different sizes in high-resolution image scenes, it is difficult to achieve global and local optimal segmentation results through conventional optimal segmentation scale algorithms. This project adopts a high-resolution remote sensing image multi-scale segmentation optimization algorithm that takes into account global and local optimality. It is mainly divided into four steps: the first step is to calculate the local variance of each scale segmentation result after multi-scale segmentation to determine the global optimal segmentation scale; the second step is to use the object space statistical index (local Moran index and object variance) based on the global optimal segmentation scale to detect over-segmented and under-segmented objects in the global optimal segmentation result; the third step is to optimize the under-segmented objects and over-segmented objects to obtain the local optimal segmentation result; the fourth step is to fuse the global optimal segmentation result and the local optimal segmentation result to finally obtain the multi-scale segmentation optimization segmentation result. See the specific process for details. Figure 5 .

[0240] Based on the multi-scale segmentation optimization of high-resolution remote sensing images taking into account global and local optimality, the present invention selects typical surface coverage image objects as the case subjects, and adopts the triple structure of "problem description-solution-solution effect" in the object-oriented framework to express the cases. A specific image object case should include four aspects: case number, case characteristics, case category and case effect, as follows:

[0241] Case number: Number each image object case to ensure that the number value is unique.

[0242] Case characteristics: corresponds to the "Problem Description" section, including image characteristics and non-image characteristics.

[0243] The image features of a case include information such as the spectrum, geometry, and texture of the case image object. The non-image features of a case include the terrain features, spatiotemporal features, and metadata features of the case object. Among them, terrain features include attributes such as elevation and slope, spatiotemporal features include attributes such as acquisition month, acquisition year, center longitude, and center latitude, and metadata features include attributes such as case number, sensor type, and spatial resolution. The image features, terrain features, and expert knowledge of a case are mainly used in the case matching stage, and are called case matching features. The spatiotemporal features and metadata features of a case are mainly used in the case retrieval stage, and are called case retrieval features.

[0244] Case category: Corresponding to the "Solution" part, since there is no unified opinion on the classification of natural resources at present, the natural resource element classification system studied in this invention refers to the surface cover classification system in the "Basic Geographic National Conditions Monitoring Content and Indicators (CHT 9029-2019)" to classify the natural resource elements of the image objects.

[0245] Case effect: corresponds to the "Solution effect" part, describing the role and effect of the case in the process of obtaining the "solution" from the "problem description", including whether it is a representative case, the number of case retrievals, the number of successful matches, and other attributes.

[0246] The case model consists of case matching features, case retrieval features, case categories and case effects. A specific case can be expressed as formula (1):

[0247] Case = {match 1，2，...，n , retrieval 1，2，...，8 ;class,subclass;effect 1，2，3} (1)

[0248] In the formula, Case represents a specific case, match is the case matching feature, retrieval is the case retrieval feature, class is the surface coverage classification code to which the case belongs, subclass is the subclass code of the surface coverage classification to which the case belongs, and effect is the case effect.

[0249] 3.2.5 Knowledge Graph Construction

[0250] Collect the business knowledge of natural resource element interpretation, and after data extraction (including concept and entity extraction, relationship extraction), construct a visualized natural resource element interpretation knowledge graph, complete and correct the knowledge graph through knowledge reasoning, and obtain a complete natural resource interpretation knowledge graph, which is integrated into the natural resource element spatiotemporal case library through case library organization, and participates in the judgment of whether the case retrieval results are correct. Due to space limitations, taking rivers as an example, the example of the transformation of natural resource element interpretation knowledge text and knowledge graph triples is shown in Table 1, and the visualization effect of natural resource element interpretation knowledge graph (part) is shown in Figure 6 shown.

[0251] Table 1 Example of transformation between natural resource element interpretation text and knowledge graph triples (taking rivers as an example)

[0252]

[0253] The ReliefF algorithm has the characteristics of simple implementation and high efficiency, and is not restricted by data types and is not affected by noise. PSO is a random search algorithm based on group collaboration developed to simulate the foraging behavior of bird flocks. It has the characteristics of being simple and easy to implement, simple parameters, fast convergence, and strong global optimization capabilities. In order to effectively weaken the influence of small object-oriented samples, high-dimensional feature sets, and uneven samples on the accuracy of image classification, the present invention adopts an improved ReliefF-PSO three-stage combined feature selection method to optimize the construction of image classification feature sets, so as to automatically select feature subsets and parameter combinations that have a more obvious effect on the classification of land objects, realize the optimization construction of the corresponding hierarchical classification feature space, and make the classification algorithm have the best performance.

[0254] Phase 1: The improved ReliefF algorithm is used to calculate and sort the feature weights, and n feature combinations with strong classification capabilities are selected to participate in Wrapper optimization to achieve the purpose of rapid dimensionality reduction and reduce the particle search space. Based on the ReliefF algorithm, the feature weights are calculated based on all training samples, and the total number of samples is set as the number of iterations. The feature weights within the class are first summed and averaged, and then the feature averages of each class are summed and averaged to obtain the final feature weights.

[0255] The improved ReliefF algorithm weight calculation formula is shown in formula (2) and formula (3):

[0256]

[0257] Among them, x is the category, x i is the sample in this category, n is the number of categories, D f (i, j) is the standard Euclidean distance of feature f, D f (x i ,H(x i)) is the distance between xi and the nearest similar sample, D f (x i ,M(x i )) is the distance between the nearest samples of different classes, is the number of samples of category x.

[0258] The second stage: Use the PSO encapsulation algorithm to measure the importance of each feature in the pre-selected feature subset, further eliminate redundant and irrelevant features, and retain m features that contribute to classification. PSO is initialized as a group of random particles. Through the collaboration and information sharing between individuals in the group, the movement of the entire group evolves from disorder to order in the problem-solving space, thereby obtaining the optimal solution. The d-dimensional space speed and position update formula of particle i is:

[0259] The formula is formula (4) and formula (5):

[0260]

[0261] in, is the d-th dimension component of the flight velocity vector of particle i in the k-th iteration; is the position vector; P id The dth dimension represents the individual extreme value of particle i, which is also the local optimal value; P gd is the optimal position of the entire particle swarm, representing the dth dimension of the global optimal solution; r1 and r2 are random functions between 0 and 1, ω is the inertia weight, and c1 and c2 are acceleration coefficients.

[0262] The present invention uses PSO as the search strategy, comprehensively considers the characteristics of the target objects and the advantages and disadvantages of each classification algorithm, combines the multi-level segmentation classification model, selects the random forest algorithm and the J48 decision tree algorithm to form the PSO_RF and PSO_J48 encapsulation algorithms respectively, and applies them to the image classification feature selection at the corresponding level.

[0263] The third stage: according to the order of feature importance, the random forest algorithm is combined with the sequential backward selection method (SBS), and the last feature in the importance order is deleted from the feature set each time. The operation is iterated repeatedly to determine the relationship between feature dimension and classification accuracy and find the optimal feature subset.

[0264] The specific process is as follows Figure 7 .

[0265] The present invention adopts the antlion algorithm to optimize the distribution of feature weights in order to improve the performance index of the CBR classifier. According to the antlion algorithm mechanism, the feature weights are made to correspond to the position coordinates of the antlion, and the optimization distribution of the feature weights includes the following five steps.

[0266] 1) Weight initialization and iterative update of weights

[0267] Let the feature weight in the iteration process be ω(t), and the update formula is formula (6):

[0268] ω(t)=[0,(2r(t1)-1), cumsum(2r(t2)-1),..., cumsum(2r(t) m )-1)] (6)

[0269] Among them, cumsum means calculating the cumulative sum, m means the maximum number of iterations (also m features), t1:t m The maximum number of iterations for each column corresponding to the feature weight is 1 to m, r(t i ) is a random function, defined as formula (7):

[0270]

[0271] Among them, rand is used to generate random numbers with uniform distribution in the interval [0, 1].

[0272] Since the weight is constrained by the boundary range, the feature weight is constrained using formula (8):

[0273]

[0274] Among them, a i is the minimum weight among the initialization weights, b i is the maximum value of the weight in the initialization weight, is the minimum value of the ith feature variable at the tth iteration, is the maximum value of the i-th characteristic variable in the t-th iteration;

[0275] The sum of the jth weight of the tth iteration and the maximum values ​​of all variables in the tth iteration is the maximum value of the ith characteristic variable of the tth iteration. The maximum value of the ith characteristic variable of the tth iteration is defined as follows:

[0276]

[0277] Among them, c t is the minimum value of all feature variables in the tth iteration, d t is the maximum value of all feature variables in the tth iteration, It represents the j-th feature weight selected at the t-th iteration, that is, the optimal feature weight remaining after removing the i-th feature weight in a set of weights.

[0278] In order to obtain the maximum value of the characteristic variable more quickly, the search range of the characteristic variable is continuously narrowed during the iteration process, and the formula (9) is constrained to obtain the formula (10):

[0279]

[0280] Where w is a constant defined based on the current iteration (when t>0.1m, w=2; when t>0.5m, w=3; when t>0.75m, w=4; when t>0.9m, w=5; when t>0.95m, w=6).

[0281] 2) Calculate the fitness function

[0282] The fitness function for evaluating weight optimization is the accuracy of the classifier, as shown in formula (11), where g is the number of correctly classified cases in the test set, and p′ is the total number of cases in the test set.

[0283]

[0284] 3) Selecting feature variables using the roulette method

[0285] In the weight optimization process of the ant lion algorithm, according to the fitness value corresponding to the weight, the roulette method is used to select the feature variable with high fitness value for the next iteration.

[0286] 4) Update feature weights and obtain optimal feature weights

[0287] In each iteration, an optimal weight is obtained. When the optimal weight appears, the optimal weight is defined as the current value according to formula (12):

[0288]

[0289] in, represents the i-th feature weight selected at the t-th iteration, as shown in formula (13):

[0290]

[0291] in, represents the weight of the feature variable selected by roulette at the tth iteration

[0292] Heavy, represents the optimal value among all weight combinations at the tth iteration.

[0293] When the last iteration is completed, m optimal feature weights are determined, and the feature weight allocation process is completed.

[0294] The present invention utilizes an improved dynamic hierarchical self-organizing mapping algorithm to organize and index the case library, and divides the initial case library into several sub-databases. The cases in each sub-case library are more similar. When retrieving a case, the new case is first directed to the corresponding sub-case library, and then the historical cases similar to the new case can be obtained by simply performing a similarity measurement calculation on the new case and the cases in the sub-case library, thereby effectively improving the accuracy and efficiency of the retrieval.

[0295] The improved GHSOM clustering algorithm is used to organize cases. The case base of the CBR system is divided into several different sub-case bases, which are located at different levels. Usually, 4 layers are enough to deal with case clustering of most CBR systems. If the historical cases are too many or too complex, the number of GHSOM map layers can be appropriately increased. The process of building a case base using the improved GHSOM algorithm is as follows: Figure 8 shown.

[0296] From the perspective of cognitive science, this invention draws on the principle of introspective learning and adjusts the weight of attributes by monitoring the system's own reasoning process and performance changes to achieve the goal of giving CBR active learning capabilities. I =(X I ; Y I ) and C II =(X II ; Y II ) Among them, X I =(X I1 , X I2 , ..., X IK , ..., X Im ), X I =(X II1 , X II2 , ..., X IIK , ..., X IIm ), f match To set the threshold, used to determine whether the attribute matches. If the condition 0≤|X is met Ii -X IIi |≤f match When , the i-th attribute of the two cases is called a matching attribute; if it does not meet this condition, it is called a non-matching attribute.

[0297] When a case is successfully matched and correctly classified, the weight value of the matching attribute of the case increases, and the weight value of the unmatched attribute decreases accordingly. When the weight value of the case matching attribute increases, the weight is calculated according to formula (14):

[0298]

[0299]

[0300] In formula (14) and formula (15), w i (t) represents the weight of the tth iteration of the i-th attribute of case; w i (t+1) is the process weight of the t+1th iteration of the i-th attribute of case; represents the change in weight, and m represents the number of case attributes.

[0301] Case library streamlining includes noise case adjustment and redundant case deletion. During the use of the case library, the number of times the case is used and the number of successful matches are recorded. When the number of times it is used is much greater than the number of successful matches, it can be determined that the case is not very helpful for system reasoning and is a typical noise case. After user diagnosis, it will be adjusted to a correct case or deleted directly. Redundant cases are cases in the same case library that are extremely similar to a certain case. Redundant cases can be found by setting a similarity threshold β that is very close to 1.

[0302] The present invention intends to establish a case retrieval model based on RBF network on the basis of improving GHSOM algorithm for case library clustering. The network has n input nodes, representing n characteristic attributes of the case; the characteristic attributes should generally be normalized so that their input values ​​are between 0 and 1.

[0303] Hidden layer unit: The selection of RBF network center is completed by self-organization method, that is, the case library clustering result of improved GHSOM algorithm is used. The number of hidden layer center nodes is equal to the sum of the number of clusters and the number of valuable outliers.

[0304] Output layer O: has m nodes, representing m different cases. The output of the network corresponds to the index or similarity value of similar cases.

[0305] The retrieval machine structure model based on RBF network is as follows Fig.10 shown.

[0306] Assume that the network has Nr hidden layer nodes. If the output layer transfer function is f(x)=x, then the RBF retrieval machine with n inputs and m outputs can realize the functions as shown in equations (16) and (17):

[0307]

[0308] And the network satisfies the minimum sum of squared output errors, as shown in formula (18):

[0309]

[0310] 3.2.12 Case Study Based on Cache Mechanism

[0311] In order to improve retrieval efficiency and information extraction accuracy, the present invention introduces case learning based on a cache mechanism into case storage.

[0312] Suppose the new case generated is C * =(F * ,FS * First, we use the case retrieval strategy to find new cases C * Similarity with all cases in the case library λ1 and λ2 are thresholds that satisfy the relationship 0.5<λ2<λ1≤1. Their specific values ​​are determined based on actual experience or experts, and λ1=0.95 and λ2=0.85 can be taken.

[0313] If SIM max <λ2, indicating that there is no case library with C * If the case is matched, C * Stored in the cache case library.

[0314] If λ2≤SIM max ≤λ1, indicating that there is a case library with C * A case that matches well, but not very well. Suppose the case library has SIM max The case is C i =(F i , FS i ), calculation case C * With Case C i The similarity SIM(C i , C * ), if SIM(C i , C * )≥λ2, then C i Using C * Replace, otherwise case C * Store in case library; if SIM max ≥λ1, indicating that there is already a case library with C * A very good match, so C * Not stored.

[0315] When the number of cases in the cache case library reaches a certain number, the processing mechanism of the cache case library is triggered, the cache case library is clustered using the adaptive FCM algorithm, noise cases are deleted, and the processed cases are stored in the case library.

[0316] Under the case learning strategy based on the cache mechanism, new cases will be automatically and continuously added to the existing case library as the CBR system runs, avoiding the need to spend a lot of manpower to supplement the case library. Fig.11 shown.

[0317] For the target image, firstly, the target image is subjected to multi-scale segmentation processing that takes into account global and local optimality to obtain the optimized segmentation result of the target image, and the target case is obtained after calculating the image features and assigning non-image features; then, the information extraction result is obtained through the case reasoning process, and then the natural resource monitoring and land use results corresponding to the experimental image are used as reference data to verify the accuracy of the natural resource element extraction result obtained by the method of the present invention and calculate the interpretation accuracy. For the specific process, see Fig.12 .

[0318] Remote sensing image interpretation technology is developing towards the direction of automation and intelligence with human-machine integration. The intelligent interpretation method of natural resource elements proposed in this project is a key scientific and technological issue that needs to be broken through in natural resource survey and monitoring. The knowledge graph and case reasoning theory involved in the research content are both from the cognition and reasoning directions in the field of artificial intelligence. They have been successfully applied in the field of remote sensing interpretation and have reliable theoretical support.

[0319] Knowledge graphs are helpful for automatic and intelligent acquisition, mining and application of knowledge. The joint interpretation method of knowledge reasoning and data learning has become an important research trend in intelligent processing of remote sensing big data. Case reasoning can effectively deal with the fuzziness and uncertainty problems encountered in acquiring knowledge, and combined with geography to solve spatial problems has become a new method for geoscience data analysis and prediction. Therefore, the present invention couples knowledge graphs and case reasoning to achieve intelligent interpretation of natural resource elements, and the technical method is feasible.

[0320] The present invention constructs a knowledge-driven high-resolution image spatiotemporal case expression model for natural resource elements. Most existing studies use the image features of remote sensing images themselves, but the high-level features obtained by deep learning methods are not very interpretable. The present invention considers both the image features and non-image features of remote sensing image objects, and uses interpretation business knowledge to construct a knowledge graph, and then constructs a knowledge-driven natural resource element spatiotemporal case expression model, which can more comprehensively express the features of different types of natural resource element objects, which is one of the characteristics and innovations of the present invention.

[0321] The present invention establishes a knowledge graph of natural resource elements in high-resolution images and a coupled intelligent interpretation paradigm of case reasoning. The existing intelligent extraction methods of natural resource elements in remote sensing images are insufficiently knowledge-driven, have weak reasoning ability and interpretability, and require a huge workload for manual labeling of samples. The knowledge graph of natural resource elements and the coupled intelligent interpretation paradigm of case reasoning established by the present invention can fully utilize interpretation business knowledge, automatically expand the case library, adaptively adjust case feature weights, and continuously improve automatic interpretation accuracy, which is also one of the characteristics and innovations of the present invention.

[0322] The specific application fields of the present invention include geographical national conditions monitoring, natural resources investigation, land use / land cover change research and other fields.

[0323] The invention has been successfully applied to the geographical survey and monitoring of the Guangxi Zhuang Autonomous Region, improving work efficiency and saving work costs.

[0324] Due to the complexity of land cover types, according to the land cover classification standards in the national geographic survey, the objects with the same land cover type still vary greatly in remote sensing images. In order to extract information according to the land cover classification standard system, the present invention reclassifies the land cover objects with the same land cover type and different image features on the basis of the land cover classification standard, that is, the original land cover classification code is expanded. After all land classes are extracted, they are converted into land cover classification standard codes using land class coding conversion.

[0325] After selecting the cases, for the cases with the same land cover type, the visual feature pairs of the representative images of the cases are analyzed and reclassified. The cases are reclassified as follows Fig.13 (Based on arable land <0100> for example).

[0326] Fill in case metadata: Enter the scale, acquisition month, acquisition year, sensor type, center point longitude, center point latitude, case matching status and other information of the case image in the case layer attribute table. After filling in, export the attribute table as an Excel spreadsheet file for the next step of feature weight calculation. See the following interface for filling in case metadata information Fig.14 .

[0327] Segmentation parameter training: Train the segmentation parameters of multi-scale segmentation to determine the best segmentation parameters, including segmentation scale, shape factor and compactness. The segmentation scale selection interface is shown in the figure. The segmentation parameter training interface is shown below Fig.15 .

[0328] Feature weight calculation: According to the analysis conclusion of the case feature table, select the appropriate weight calculation method to calculate the weight of each feature of the case. The calculated weight result is displayed in the form of a curve graph and can be exported to an Excel table. The feature weight calculation interface is shown below Fig.16 .

[0329] Case training is to select cases for case matching threshold training, obtain case matching results, and adjust the threshold by superimposing the matching results and images. When the matching accuracy is 100% and the number of successful matches is the largest, the threshold is the best matching threshold. The best threshold is selected to determine the best threshold. There are three ways of case training: manual training with weights considered, manual training without weights considered, and automatic training without weights considered.

[0330] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0331] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A remote sensing image information extraction method based on case-based reasoning coupled with knowledge graph, characterized in that: The following steps are involved: Step 1: Collect high-resolution images of the study area, geographical conditions monitoring results, field survey data, land use maps and various socio-economic statistical data, and form a basic data set of the study area through image radiometric calibration, atmospheric correction and geometric correction preprocessing. Collect business knowledge on interpretation of natural resource elements, organize it into a knowledge graph, and then complete and correct the knowledge through knowledge reasoning; Step 2: After segmenting the high-resolution image and optimizing the segmentation results, an optimized image object is obtained. Typical natural resource element samples are manually selected as case objects, and natural resource element sample cases are formed after feature selection and weight allocation; Step 3: After selecting a certain number of natural resource element cases from different sensors with different spatiotemporal environments and different types, integrate them into the interpretation knowledge graph, organize various sample cases based on the improved GHSOM algorithm to form a spatiotemporal case library of natural resource elements, and maintain the case library; Step 4: The high-resolution image to be interpreted is segmented and feature-assigned to form a target case, which is then searched and matched with the case library. The retrieved case type is the natural resource element type corresponding to the target case, and is added to the case library as a new case through case learning. The extracted result is the intelligent interpretation result of the natural resource element after accuracy verification.

2. The remote sensing image information extraction method based on case-based reasoning coupled with knowledge graph as claimed in claim 1 is characterized in that: The case objects include: Case number: number each image object case to ensure that the number value is unique; Case characteristics: corresponding to the "Problem Description" section, including imaging characteristics and non-imaging characteristics; Case category: Corresponding to the "Solution" section, the natural resource element classification system refers to the land cover classification system in the "Basic Geographic National Condition Monitoring Content and Indicators (CHT 9029-2019)" to classify the natural resource elements of the image objects; Case effect: corresponds to the "Solution effect" part, describing the role and effect of the case in the process of obtaining the "Solution" from the "Problem Description", including whether it is a representative case, the number of case retrievals, and the number of successful matches; The case model consists of case matching features, case retrieval features, case categories and case effects. A specific case is expressed as formula (1): Case={match 1，2，...，n ,retrieval 1，2，...，8 ;class,subclass;effect 1，2，3 } (1) In the formula, Case represents a specific case, match is the case matching feature, retrieval is the case retrieval feature, class is the surface coverage classification code to which the case belongs, subclass is the subclass code of the surface coverage classification to which the case belongs, and effect is the case effect.

3. The remote sensing image information extraction method based on case-based reasoning coupled with knowledge graph as claimed in claim 1 is characterized in that: The feature selection method is as follows: Phase 1: The improved ReliefF algorithm is used to calculate and sort feature weights, and n feature combinations with strong classification capabilities are selected to participate in Wrapper optimization to achieve the purpose of rapid dimensionality reduction and reduce particle search space; based on the ReliefF algorithm, feature weights are calculated based on all training samples, and the total number of samples is set as the number of iterations. The feature weights within the class are first summed and averaged, and then the feature averages of each class are summed and averaged to obtain the final feature weights; The improved ReliefF algorithm weight calculation formula is shown in formula (2) and formula (3): Among them, x is the category, x i is the sample in this category, n is the number of categories, D f (i, j) is the standard Euclidean distance of feature f, D f (x i ,H(x i )) is x i The distance to the nearest similar sample, D f (x i ,M(x i )) is the distance between the nearest samples of different classes, m class (x i ) is the number of samples of category x; The second stage: the PSO encapsulation algorithm is used to measure the importance of each feature in the pre-selected feature subset, further eliminate redundant and irrelevant features, and retain m features that contribute to classification; PSO is initialized as a group of random particles. Through the collaboration and information sharing between individuals in the group, the movement of the entire group produces an evolution process from disorder to order in the problem-solving space, thereby obtaining the optimal solution; the d-dimensional space speed and position update formula of particle i are formula (4) and formula (5): in, is the d-th dimension component of the flight velocity vector of particle i in the k-th iteration; is the position vector; P id The dth dimension represents the individual extreme value of particle i, which is also the local optimal value; P gd is the best position of the entire particle swarm, indicating the dth dimension of the global optimal solution; r1 and r2 are random functions between 0 and 1, ω is the inertia weight, and c1 and c2 are acceleration coefficients; PSO is used as a search strategy, the characteristics of the target objects and the advantages and disadvantages of each classification algorithm are comprehensively considered, and the multi-level segmentation classification model is combined. The random forest algorithm and the J48 decision tree algorithm are selected to form the PSO_RF and PSO_J48 encapsulation algorithms respectively, which are applied to the image classification feature selection at the corresponding level; The third stage: according to the order of feature importance, the random forest algorithm is combined with the sequential backward selection method SBS. The last feature in the importance order is deleted from the feature set each time, and the operation is iterated repeatedly to determine the relationship between feature dimension and classification accuracy and find the optimal feature subset.

4. The remote sensing image information extraction method based on case-based reasoning coupled with knowledge graph as claimed in claim 1, characterized in that: The weight allocation method is as follows: 1) Weight initialization and iterative update of weights Let the feature weight in the iteration process be ω(t), and the update formula is formula (6): ω(t)=[0,(2r(t1)-1),sum(2r(t2)-1),...,sum(2r(t m )-)] (6) Among them, cumsum means calculating the cumulative sum, m means the maximum number of iterations, that is, m features, t1:t m The maximum number of iterations for each column corresponding to the feature weight is 1 to m, r(t i ) is a random function, defined as formula (7): Among them, rand is used to generate random numbers with uniform distribution in the interval [0, 1]; Since the weight is constrained by the boundary range, the feature weight is constrained using formula (8): Among them, ai is the minimum weight in the initialization weight, bi is the maximum weight in the initialization weight, is the minimum value of the ith feature variable at the tth iteration, is the maximum value of the i-th characteristic variable in the t-th iteration; The sum of the jth weight of the tth iteration and the maximum values ​​of all variables in the tth iteration is the maximum value of the ith characteristic variable of the tth iteration. The maximum value of the ith characteristic variable of the tth iteration is defined as follows: Among them, c t is the minimum value of all feature variables in the tth iteration, d t is the maximum value of all feature variables in the tth iteration, represents the jth feature weight selected at the tth iteration, that is, the optimal feature weight remaining after removing the i-th feature weight in a set of weights; In the iterative process, the search range of the feature variables is continuously narrowed, and the formula (9) is constrained to obtain the formula (10): Where w is a constant defined based on the current iteration. When t>0.1m, w=2; when t>0.5m, w=3; when t>0.75m, w=4; when t>0.9m, w=5; when t>0.95m, w=6; 2) Calculate the fitness function The fitness function for evaluating weight optimization is the accuracy of the classifier, as shown in formula (11): Where g is the number of correctly classified cases in the test set, and p′ is the total number of cases in the test set; 3) Selecting feature variables using the roulette method In the weight optimization process of the Ant Lion Algorithm, according to the fitness value corresponding to the weight, the roulette method is used to select the feature variable with a high fitness value for the next iteration; 4) Update feature weights and obtain optimal feature weights In each iteration, an optimal weight is obtained. When the optimal weight appears, the optimal weight is defined as the current value according to formula (12): in, represents the i-th feature weight selected at the t-th iteration, as shown in formula (13): in, represents the weight of the feature variable selected by roulette at the tth iteration, represents the optimal value among all weight combinations at the tth iteration; When the last iteration is completed, m optimal feature weights are determined, and the feature weight allocation process is completed; Success-driven weight introspective learning In the same case library, suppose there are two cases, C I =(X I ; Y I ) and C I =(X II ; Y II ), where X I =(X I1 , X I2 , ..., X IK , ..., X Im ), X I =(X II1 , X II2 , ..., X IIK , ..., X IIm ), f match To set the threshold, used to determine whether the attribute matches. If the condition 0≤|X is met Ii -X IIi |≤f match When , the i-th attribute of the two cases is called a matching attribute; if it does not meet this condition, it is called a non-matching attribute; When a case is successfully matched and correctly classified, the weight value of the matching attribute of the case increases, and the weight value of the unmatched attribute decreases accordingly; when the weight value of the case matching attribute increases, the weight is calculated according to formula (14): In formula (14) and formula (15), w i (t) represents the weight of the tth iteration of the i-th attribute of case; w i (t+1) is the process weight of the t+1th iteration of the attribute of the case; represents the change in weight, and m represents the number of case attributes.

5. The remote sensing image information extraction method based on case-based reasoning coupled with knowledge graph as claimed in claim 1, characterized in that: The case library is searched: On the basis of improving GHSOM algorithm to cluster case database, a case retrieval model based on RBF network is established; the network has n input nodes, representing n characteristic attributes of the case; Generally, the feature attributes should be normalized so that their input values ​​are between 0 and 1; Hidden layer unit: The selection of RBF network center is completed by self-organization method, that is, the case library clustering result of improved GHSOM algorithm is used; the number of hidden layer center nodes is equal to the sum of the number of clusters and the number of valuable outliers; Output layer O: has m nodes, representing m different cases. The output of the network corresponds to the index or similarity value of similar cases. Retrieval machine structure model based on RBF network; Assume that the network has Nr hidden layer nodes. If the output layer transfer function is f(x)=x, then the RBF retrieval machine with n inputs and m outputs can realize the functions as shown in equations (16) and (17): And the network satisfies the minimum sum of squared output errors, as shown in formula (18):

6. The remote sensing image information extraction method based on case-based reasoning coupled with knowledge graph as claimed in claim 1, characterized in that: The case study described: Suppose the new case generated is C * =(F * , FS * ), firstly, we obtain new cases C according to the case retrieval strategy * Similarity with all cases in the case library λ1 and λ2 are thresholds satisfying the relationship of 0.5<λ2<λ1≤1. Their specific values ​​are determined based on actual experience or experts, and λ1=0.95 and λ2=0.85; If SIM max <λ2, indicating that there is no case library with C * If the case is matched, C * Store in the cache case library; If λ2≤SIM max ≤λ1, indicating that there is a case library with C * A case that matches well, but not very well. Suppose the case library has SIM max The case is C i =(F i , FS i ), calculation case C * With Case C i The similarity SIM(C i , C * ), if SIM(C i , C * )≥λ2, then C i Using C * Replace, otherwise case C * Store in case library; if SIM max ≥λ1, indicating that there is already a case library with C * A very good match, so C * Not to be stored; When the number of cases in the cache case library reaches a certain number, the cache case library processing mechanism is triggered, the cache case library is clustered using the adaptive FCM algorithm and noise cases are deleted, and the processed cases are stored in the case library; Under the case learning strategy based on the cache mechanism, new cases will be automatically and continuously added to the existing case library as the CBR system runs.

7. A remote sensing image information extraction system using case-based reasoning coupled with a knowledge graph that implements the remote sensing image information extraction method using case-based reasoning coupled with a knowledge graph as described in any one of claims 1 to 6, characterized in that: include: The collection module is used to collect high-resolution images of the study area, geographical national conditions monitoring results, field survey data, land use maps and various socio-economic statistical data. After image radiometric calibration, atmospheric correction and geometric correction preprocessing, the basic data set of the study area is formed. The business knowledge of natural resource element interpretation is collected and organized into a knowledge graph, and then knowledge is supplemented and error corrected through knowledge reasoning. The optimization processing module is used to segment the high-resolution image and optimize the segmentation results to obtain the optimized image object, manually select typical natural resource element samples as case objects, and form natural resource element sample cases after feature selection and weight allocation; The integration module is used to select a certain number of natural resource element cases from different sensors with different spatiotemporal environments and different types, integrate them into the interpretation knowledge graph, organize various sample cases based on the improved GHSOM algorithm to form a spatiotemporal case library of natural resource elements, and maintain the case library; The extraction module is used to form a target case after image segmentation and feature assignment of the high-resolution image to be interpreted, and then search and match it with the case library. The case type retrieved is the natural resource element type corresponding to the target case, and is added to the case library as a new case through case learning. The extraction result is the intelligent interpretation result of the natural resource element after accuracy verification.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the remote sensing image information extraction method of case reasoning coupled with knowledge graph as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the remote sensing image information extraction method of case reasoning coupled with knowledge graph as described in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the remote sensing image information extraction system of case reasoning coupled with knowledge graph as described in claim 7.

Images