Interpretation sign driven remote sensing image sample automatic production method and system

Abstract

This invention relates to an automated production method and system for remote sensing image samples based on interpretation markers, specifically in the field of intelligent remote sensing interpretation. This solution transforms discrete samples and domain knowledge into a computable and reasonable structured semantic network through a constructed heterogeneous knowledge graph. It encapsulates task semantics and preferences using dynamic query vectors, driving a graph walk strategy anchored by a set of key interpretation markers and incorporating transition probabilities. Under real-time feedback from an online evaluation function, it intelligently explores and iteratively optimizes the subset of samples with the highest comprehensive utility score. When coverage is insufficient, it initiates procedural sample synthesis to form an enhanced sample set. Simultaneously, it utilizes performance feedback from model training to dynamically update the weight parameters of the online evaluation function and the utility weights of sample nodes in the graph, achieving a closed loop from passive retrieval to active compilation and self-optimization. This significantly improves the ability to accurately select and synthesize and enhance massive amounts of samples for complex tasks.

Description

Technical Field

[0001] This invention relates to the field of intelligent remote sensing interpretation, and more specifically, to a method and system for automated production of remote sensing image samples based on interpretation markers. Background Technology

[0002] With the rapid development of Earth observation technology, remote sensing image data has experienced explosive growth. Deep learning-based automatic target interpretation technology has become a key means of extracting information from massive amounts of imagery. However, the performance of deep learning models is highly dependent on a large number of high-quality and diverse labeled samples. To this end, leading geographic information service organizations in the industry typically invest heavily in establishing enterprise-level standardized remote sensing image sample product libraries, which systematically store massive sample data classified by target category, acquisition sensor, imaging time, geographic region, and manual annotation accuracy level. When faced with diverse practical business needs such as emergency monitoring of natural disasters, reconnaissance of specific military targets, and dynamic inspection of illegal buildings in cities, it is necessary to quickly train a dedicated and efficient deep learning model for each new and specific interpretation task, such as "identification of helicopters and rescue vehicles in post-earthquake disaster areas" or "investigation of illegal marine facilities." This requires the ability to quickly and accurately select and combine an optimal training sample set suitable for the specific task from hundreds of millions of samples to support the rapid development and deployment of the model.

[0003] However, existing database-managed sample repository technologies are insufficient to effectively support the aforementioned intelligent construction requirements for task-oriented training sets. The fundamental technical problem lies in the lack of a mechanism capable of understanding complex task semantics and performing intelligent recommendation and synthesis based on deep sample features and related knowledge. Specifically, existing technologies have the following limitations: First, sample retrieval capabilities remain at a coarse-grained filtering stage based on simple metadata tags, unable to respond to complex semantic queries that "include multi-target interaction relationships" or "have specific scene contexts," making it difficult to directly obtain candidate sample sets highly relevant to the task. Second, the construction of training sets is essentially a process involving... The problem of multi-objective optimization, including diversity, category balance, difficulty distribution, and scene coverage, is addressed by current trial-and-error combinatorial methods that rely on engineers' experience. These methods lack theoretical guidance and quantitative evaluation, and cannot guarantee that the subsets extracted from massive samples are globally near-optimal. This not only results in low efficiency but may also introduce model bias due to improper training set construction. Finally, although the professional interpretation markers containing target characteristics and scene patterns formed during sample production are not effectively utilized, this structured knowledge has not been deeply integrated with the sample selection logic. This fails to drive creative sample synthesis and adaptation for new task requirements, thus preventing the full release of the potential of the sample library. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an automated production method and system for remote sensing image samples based on interpretation markers, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: an automated production method for remote sensing image samples based on interpretation markers, specifically including the following steps: Step S1: Obtain sample metadata, annotation information and associated original images from the sample library, and extract sample visual features; simultaneously obtain predefined target interpretation tags from the interpretation tag library; based on this, construct a heterogeneous knowledge graph containing sample nodes, target category nodes, interpretation tag nodes and scene nodes, and form a structured semantic network. Step S2: Parse the target task description text, identify target entities and scene constraints, and map them to corresponding nodes in the heterogeneous knowledge graph to form a dynamic query vector containing a set of target entities, a set of key interpretation markers, constraints, and task preference parameters. Step S3: Using the dynamic query vector as input, and the key interpretation marker set as the initial anchor point in the heterogeneous knowledge graph, a walking strategy combining transition probabilities is executed to dynamically explore sample nodes and construct candidate sample subsets; through an online evaluation function, the comprehensive utility score of the candidate sample subset is calculated in real time, and the walking strategy is iteratively adjusted based on the feedback of the comprehensive utility score until the optimal sample subset that meets the conditions is found. Step S4: If the coverage of the optimal sample subset for a specified combination or a specified scene combination in the key interpretation tag set is lower than a preset threshold, then procedural sample synthesis is triggered to generate a synthetic image; pseudo-labels are generated for the synthetic image that passes quality verification and added to the optimal sample subset as synthetic samples to obtain an enhanced sample set for model training; training performance feedback is collected, and the weight parameters of the online evaluation function and the utility weights of the sample nodes in the heterogeneous knowledge graph are updated.

[0006] In a preferred embodiment, step S1 involves extracting visual features of the sample, specifically as follows: Using a deep convolutional neural network model pre-trained on a large-scale remote sensing image dataset, forward computation is performed on the corresponding region of the original image associated with each sample in the sample library. The feature map output by the last convolutional layer of the deep convolutional neural network model is extracted, and global average pooling is performed on the feature map in the spatial dimension to obtain a fixed-dimensional multi-dimensional real vector that can represent the visual content of the sample. This multi-dimensional real vector is defined as the visual feature vector.

[0007] In a preferred embodiment, a heterogeneous knowledge graph is constructed, comprising sample nodes, target category nodes, interpretation marker nodes, and scene nodes, specifically including: Create a sample node for each sample in the sample library and assign a corresponding visual feature vector to the sample node; Create a target interpretation flag node for each target interpretation flag in the interpretation flag library; create target category nodes for all known target categories; create scene nodes for all known scene types; Establish directed edges representing the category affiliation relationship between sample nodes and target category nodes; Establish directed edges representing the relationship between target category nodes and target interpretation marker nodes; Establish directed edges representing the relationship between the source scene and the sample nodes; Establish an attribute edge with quantifiable weights between the sample node and the target interpretation marker node; The calculation process of attribute edge weights is as follows: First, a trainable attention mechanism network is used to process the visual feature vectors of sample nodes in order to focus on the feature components related to the description of the target interpreter node. Simultaneously, a trainable projection mapping function is used to convert the text or feature descriptions of the associated target interpretation marker nodes into vector representations of the same dimension; Then, the element-wise product of the feature vector output by the attention mechanism network and the vector output by the projection mapping function is calculated, and then mapped to a real value between zero and one through a learnable linear transformation layer and a non-linear activation function. This real value is the association strength. The structured semantic network is composed of sample nodes, target category nodes, target interpretation marker nodes, scene nodes, and directed edges connecting these nodes and attribute edges with association strength.

[0008] In a preferred embodiment, step S2, parsing the target task description text and identifying target entities and scene constraints, specifically includes: Pre-trained natural language processing models are used to perform sequence labeling and dependency parsing on target task description text; Lexical units representing specific geographical features in the target task description text are extracted through sequence labeling operations and identified as target entities. By using dependency parsing to extract lexical units that modify the target entity and represent its environment and time, these units are identified as scene constraints; ultimately, a structured list of target entities and a list of scene constraints are obtained.

[0009] In a preferred embodiment, mapping to corresponding nodes in a heterogeneous knowledge graph and forming a dynamic query vector specifically includes: Each target entity in the identified target entity list is semantically matched with a predefined target category node in the heterogeneous knowledge graph, and the successfully matched target category node is added to the target entity set. Each scene constraint in the identified scene constraint list is semantically similar to a predefined scene node in the heterogeneous knowledge graph, and the successfully matched scene nodes are added to the scene constraint set. For each target category node in the target entity set, query all target interpretation marker nodes that are directly connected to it through relation edges in the heterogeneous knowledge graph, and filter out the key target interpretation marker nodes according to preset rules to form a key interpretation marker set; By combining preset task preference parameters, the target entity set, key interpretation flag set, scene constraint set, and task preference parameters are encapsulated together to form a dynamic query vector.

[0010] In a preferred embodiment, in step S3, a walk strategy incorporating transition probabilities is executed to dynamically explore sample nodes and construct a subset of candidate samples. The specific process is as follows: Using the target interpretation marker nodes in the key interpretation marker set as initial anchors, locate all sample nodes in the heterogeneous knowledge graph that are directly connected to these initial anchors through attribute edges, use these sample nodes as the initial set to be explored, and initialize an empty subset of candidate samples. The transition probability in the walk strategy is determined by three factors: the first factor is the geometric mean of the association strength between the sample node to be evaluated and all target interpreter node in the key interpreter set; The second factor is the matching degree between the sample node to be evaluated and the set of scene constraints in the dynamic query vector. If the sample node to be evaluated is connected to any scene node in the set of scene constraints through a directed edge representing the relationship of the source scene, the factor takes a value of one; otherwise, it takes a very small positive number close to zero. The third factor is a subtraction of the cosine similarity between the visual feature vector of the sample node to be evaluated and the average visual feature vector of all sample nodes in the current candidate sample subset. The final transition probability is the product of the first factor, the second factor, and the third factor, after normalization. Normalization ensures that the sum of the probabilities of all possible transition directions starting from the current sample node is one. Each time a new sample node is selected to be added to or replaced in the candidate sample subset based on the transition probability, the overall utility score of the current candidate sample subset is recalculated. If the new overall utility score is higher than the highest score in history, then the adjustment will be accepted and the highest score record will be updated. If a predetermined number of consecutive walks and adjustments fail to produce a higher overall utility score, a walk strategy adjustment is triggered. This involves either increasing the weight of the third factor in the transition probability calculation or randomly selecting a sample node from the highest-scoring candidate sample subset in the historical records as a new walk starting point to change the exploration direction.

[0011] In a preferred embodiment, a comprehensive utility score for a subset of candidate samples is calculated in real time using an online evaluation function, wherein the online evaluation function is a weighted sum of four utility measures, the four dimensions being: Diversity measures, difficulty measures, coverage measures, and redundancy measures; Diversity measures are characterized by calculating the information entropy of the distribution of visual feature vectors of all sample nodes in a subset of candidate samples in the vector space. The difficulty metric evaluates each sample node in the candidate sample subset using a pre-trained neural network model that takes visual feature vectors as input and outputs predicted difficulty values, and takes the arithmetic mean of the predicted difficulty values ​​of all sample nodes. Coverage is measured by how many target interpreter nodes in the key interpreter set are covered by sample nodes in the candidate sample subset. A target interpreter node is considered covered if and only if there is at least one sample node in the candidate sample subset with a correlation strength greater than a preset coverage threshold. Redundancy is measured by calculating the arithmetic mean of the cosine similarity of the visual feature vectors between all unordered sample node pairs in the candidate sample subset. The weight coefficients of the four dimensions—diversity measure, difficulty measure, coverage measure, and redundancy measure—are provided by the task preference parameters in the dynamic query vector. The search process continues until the preset maximum number of total steps is reached, or the improvement of the highest comprehensive utility score in multiple consecutive iterations is less than a preset small positive threshold. At this point, the candidate sample subset corresponding to the highest historical comprehensive utility score is determined as the optimal sample subset that meets the conditions.

[0012] In a preferred embodiment, step S4, determining whether the coverage of the optimal sample subset for a specified combination or specified scene combination in the key interpretation flag set is lower than a preset threshold, specifically includes: For each combination of pairs of target interpretation marker nodes in the key interpretation marker set and each combination of pairs of scene nodes in the scene constraint set in the dynamic query vector, calculate the coverage state of the optimal sample subset for each combination. A specified combination is considered covered if and only if there exists at least one sample node in the optimal sample subset such that the association strength between that sample node and every node in the combination is greater than a preset coverage strength threshold. Subsequently, the proportion of the number of covered specified combinations to the total number of all specified combinations is calculated to obtain the combination coverage. The calculated combination coverage is compared with a preset combination coverage threshold. If it is lower than the combination coverage threshold, it is determined that there is insufficient coverage, and all the specified uncovered combinations are identified as targets to trigger programmatic sample synthesis.

[0013] In a preferred embodiment, the weight parameters of the online evaluation function for programmatic sample synthesis and subsequent updates, and the utility weights of sample nodes in the heterogeneous knowledge graph, specifically include: For each uncovered specified combination, based on the semantic and structural constraint descriptions of each node in the combination in the interpretation tag library, high-quality sample image patches that satisfy the constraints of a single node are retrieved and segmented from the sample library. Based on physical laws and scene context knowledge, a constraint satisfaction optimization model is used to calculate the optimal spatial layout of each high-quality sample image patch on a selected background image. The goal of the constraint satisfaction optimization model is to maximize the satisfaction of the spatial relationship constraints derived from the interpretation label library. Using a pre-trained generative adversarial network, high-quality sample image patches are processed for illumination, shadow and edge fusion based on the calculated optimal spatial layout to generate synthetic images. The quality of the synthetic image is verified by a realism discrimination neural network. The image is considered to pass only when the realism confidence score output by the realism discrimination neural network is higher than a preset strict quality threshold. To automatically generate pseudo-labels for the verified synthetic images, synthetic samples are formed and added to the optimal sample subset to obtain the enhanced sample set; The system collects performance metrics obtained from training the model using augmented sample sets. Based on these performance metrics, it adjusts the weight parameters of each utility metric in the online evaluation function using an optimization algorithm. It also updates the utility weight attributes of the corresponding sample nodes in the heterogeneous knowledge graph based on the gradient contribution of each sample node in the augmented sample set during the model training process.

[0014] This application also provides an automated remote sensing image sample production system based on interpretation markers, specifically including: The knowledge graph construction module extracts visual features of samples and obtains predefined interpretation symbols based on the sample library and interpretation symbol library. It then constructs a heterogeneous knowledge graph containing sample nodes, target category nodes, interpretation symbol nodes and scene nodes to form a structured semantic network. Query generation module: Based on task description text and heterogeneous knowledge graph, it parses and identifies target entities and scene constraints, completes the mapping with graph nodes, and generates dynamic query vectors by encapsulating target entity set, key interpretation mark set, constraint conditions and task preference parameters; Intelligent retrieval module: Taking dynamic query vectors as input, starting from the set of key interpretation markers in the heterogeneous knowledge graph, it performs a walk and online evaluation that combines transition probabilities, it iteratively adjusts the walk strategy and calculates the comprehensive utility score of the candidate sample subset, and searches for and determines the optimal sample subset; Enhancement and optimization module: When the coverage of the optimal sample subset for the specified interpretation tag combination or scene combination is lower than the preset threshold, the programmatic sample synthesis and quality verification are triggered. An enhanced sample set is obtained by generating and adding synthetic samples for model training, and training performance feedback is collected to update the weight parameters of the online evaluation function and the utility weights of sample nodes in the heterogeneous knowledge graph.

[0015] The beneficial effects of this invention are as follows: By constructing a heterogeneous knowledge graph, discrete samples and domain knowledge are transformed into a computable and reasonable structured semantic network; by using dynamic query vectors to encapsulate task semantics and preferences, a graph walking strategy combined with transition probabilities, anchored by a set of key interpretation markers, is driven to intelligently explore and iteratively optimize the subset of samples with the highest comprehensive utility score under real-time feedback from the online evaluation function; when coverage is insufficient, procedural sample synthesis is initiated to form an enhanced sample set; simultaneously, by utilizing the performance feedback of model training, the weight parameters of the online evaluation function and the utility weights of sample nodes in the graph are dynamically updated, realizing a closed loop from passive retrieval to active compilation and then to self-optimization, significantly improving the ability to accurately select and synthesize and enhance massive samples for complex tasks. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application. Example 1

[0020] This embodiment provides, for example Figure 1 The automated production method for remote sensing image samples based on interpretation markers, as shown, specifically includes the following steps: Step S1: Obtain sample metadata, annotation information and associated original images from the sample library, and extract sample visual features; simultaneously obtain predefined target interpretation tags from the interpretation tag library; based on this, construct a heterogeneous knowledge graph containing sample nodes, target category nodes, interpretation tag nodes and scene nodes, and form a structured semantic network. Step S2: Parse the target task description text, identify target entities and scene constraints, and map them to corresponding nodes in the heterogeneous knowledge graph to form a dynamic query vector containing a set of target entities, a set of key interpretation markers, constraints, and task preference parameters. Step S3: Using the dynamic query vector as input, and the key interpretation marker set as the initial anchor point in the heterogeneous knowledge graph, a walking strategy combining transition probabilities is executed to dynamically explore sample nodes and construct candidate sample subsets; through an online evaluation function, the comprehensive utility score of the candidate sample subset is calculated in real time, and the walking strategy is iteratively adjusted based on the feedback of the comprehensive utility score until the optimal sample subset that meets the conditions is found. Step S4: If the coverage of the optimal sample subset for a specified combination or a specified scene combination in the key interpretation tag set is lower than a preset threshold, then procedural sample synthesis is triggered to generate a synthetic image; pseudo-labels are generated for the synthetic image that passes quality verification and added to the optimal sample subset as synthetic samples to obtain an enhanced sample set for model training; training performance feedback is collected, and the weight parameters of the online evaluation function and the utility weights of the sample nodes in the heterogeneous knowledge graph are updated.

[0021] In this embodiment, it is particularly important to explain step S1, which involves extracting the visual features of the sample. The specific process is as follows: Using a deep convolutional neural network model pre-trained on a large-scale remote sensing image dataset, forward computation is performed on the corresponding region of the original image associated with each sample in the sample library. "Large-scale remote sensing image dataset" refers to a collection of publicly available or self-built remote sensing images that include multiple sensor sources, cover different regions, seasons, and weather conditions, and contain a variety of typical ground features. For example, it can cover satellite images such as WorldView, GeoEye, Gaofen series, and Sentinel. The data volume is usually in the range of hundreds of thousands to millions of scenes to ensure that the trained deep convolutional neural network model has a strong generalization feature extraction capability. The feature map output from the last convolutional layer of a deep convolutional neural network model is extracted, and global average pooling is performed on the feature map in the spatial dimension to obtain a fixed-dimensional multi-dimensional real vector that can represent the visual content of the sample. This multi-dimensional real vector is defined as the visual feature vector. The specific process of global average pooling is as follows: for the three-dimensional feature map output from the last convolutional layer, whose size is height multiplied by width multiplied by the number of channels, the arithmetic mean of all pixel values ​​of each channel is calculated in the two spatial dimensions of height and width, thereby compressing each channel into a scalar, and finally obtaining a one-dimensional vector with a dimension equal to the number of channels. This vector is the visual feature vector. This method can effectively aggregate spatial information and maintain feature translation invariance. Construct a heterogeneous knowledge graph containing sample nodes, target category nodes, interpretation marker nodes, and scene nodes, specifically including: Create a sample node for each sample in the sample library and assign a corresponding visual feature vector to the sample node; Create a target interpretation flag node for each target interpretation flag in the interpretation flag library; create target category nodes for all known target categories; create scene nodes for all known scene types; Establish directed edges representing the category affiliation relationship between sample nodes and target category nodes; Establish directed edges representing the relationship between target category nodes and target interpretation marker nodes; Establish directed edges representing the relationship between the source scene and the sample nodes; Establish an attribute edge with quantifiable weights between the sample node and the target interpretation marker node; The calculation process of attribute edge weights is as follows: First, a trainable attention mechanism network processes the visual feature vectors of sample nodes to focus on the feature components related to the description of the associated target interpretation marker nodes. The "trainable attention mechanism network" is specifically a multilayer perceptron structure. Its calculation process is as follows: The visual feature vectors of the sample nodes are taken as input, first undergoing a linear transformation through a fully connected layer followed by a non-linear activation function, and then outputting an attention weight vector with the same dimension as the input vector through a second fully connected layer. This attention weight vector is normalized using the Softmax function to obtain the attention score for each feature component. Finally, the original visual feature vector is multiplied element-wise with the normalized attention score vector to obtain a weighted feature vector, which is the "feature vector output by the attention mechanism network." This process enables the network to autonomously learn and enhance the visual features most relevant to the semantic meaning of the target interpretation marker, while suppressing irrelevant features. Simultaneously, a trainable projection mapping function is used to convert the text or feature descriptions of the associated target interpretation marker nodes into vector representations of the same dimension. The "trainable projection mapping function" is specifically implemented according to the different description forms of the target interpretation marker nodes: if the description is text, a pre-trained word vector model is used to convert the text into a sequence of word vectors, which is then encoded into a fixed-dimensional text feature vector through a recurrent neural network or a self-attention network, and finally projected onto the target dimension through a fully connected layer; if the description is a preset feature vector, it is directly projected through a fully connected layer. The purpose is to uniformly map different forms of description to the same vector space as the feature vectors output by the attention mechanism network, so as to perform subsequent matching calculations. Subsequently, the element-wise product of the feature vector output by the attention mechanism network and the vector output by the projection mapping function is calculated. This product is then passed through a learnable linear transformation layer and a non-linear activation function, ultimately mapping to a real value between zero and one. This real value represents the association strength. "Calculating the element-wise product" means multiplying the corresponding values ​​of two vectors of the same dimension to generate a new vector of the same dimension. The "learnable linear transformation layer" is a fully connected layer whose input is the vector obtained from the element-wise product, and whose output is a scalar. The "non-linear activation function" specifically refers to the Sigmoid function. The complete calculation process can be expressed mathematically as follows: Let A be the feature vector output by the attention mechanism network, and B be the vector output by the projection mapping function. The element-wise product is C, where each element of C is equal to the product of the corresponding elements of A and B. Next, the linear transformation is calculated: the vector C is input into the fully connected layer, and its dot product with the weight matrix is ​​calculated plus the bias to obtain an intermediate scalar Z. Finally, the Sigmoid function is applied to Z: the association strength is equal to one divided by one plus the negative Z power of the natural constant e. This design allows the association strength to dynamically and subtly reflect the degree of matching between the visual features of the sample and the semantic description of the target interpretation mark. The closer its value is to one, the higher the matching degree. The structured semantic network is composed of sample nodes, target category nodes, target interpretation marker nodes, scene nodes, and directed edges connecting these nodes, as well as attribute edges with association strength. The specific application of "association strength" in attribute edges is a filterable soft connection. During subsequent graph traversal, a strength threshold can be set, such as a value between 0.6 and 0.8, to retain only attribute edges with association strength greater than this threshold for calculation, thereby eliminating unreliable weak associations and improving the efficiency and accuracy of graph query and reasoning. This structured semantic network is stored and managed in the form of a graph database, supporting efficient multi-hop queries and neighbor node aggregation operations.

[0022] In this embodiment, it is particularly important to explain step S2, which involves parsing the target task description text and identifying the target entity and scene constraints. This specifically includes: A pre-trained natural language processing (NLP) model is used to perform sequence labeling and dependency parsing on the target task description text. Specifically, the pre-trained NLP model can be a neural network model trained on a general large-scale corpus with sequence labeling and syntactic analysis capabilities, such as a pre-trained language model based on the Transformer architecture. During parsing, the target task description text is first input into the model. The model's sequence labeling module predicts the entity type label for each lexical unit in the text, such as a "location target" label or a "scene" label, thereby identifying potential target entities and scene words. Next, the model's dependency parsing module parses the grammatical structure of the text, constructs a dependency relationship tree between lexical units, and then clearly identifies which scene words representing environment, time, or geographical state function as attributives or adverbs to modify target entity words that are the core object or subject. Sequence labeling is used to extract lexical units representing specific features or targets from the target task description text and identify them as target entities. The sequence labeling operation here is to extract all consecutive lexical units labeled as "feature target" based on the entity type labels predicted by the model, forming an independent target entity recognition result. Each identified target entity is a text segment. Dependency parsing is used to extract lexical units that modify the target entity and represent the state of environment, time, or location, and these units are identified as scene constraints. The extraction operation here involves searching for lexical nodes in the dependency tree that have a direct modification relationship with the identified target entity nodes, such as modifiers in "attributive-head relation" or "adverbial-head relation". These lexical units that conform to the modification relationship and semantically represent environment, time, or location are extracted to form an independent scene constraint recognition result. Finally, a structured list of target entities and a list of scene constraints are obtained. The structured list of target entities and the list of scene constraints refer to storing all the identified target entity text and scene constraint text in a list data structure, which provides clear input for subsequent graph node mapping. Mapping to corresponding nodes in heterogeneous knowledge graphs and forming dynamic query vectors, specifically including: Each target entity in the identified target entity list is semantically matched with predefined target category nodes in the heterogeneous knowledge graph. Successfully matched target category nodes are then added to the target entity set. The specific process of semantic similarity matching is as follows: the target entity text and the name or alias text of the target category node are each converted into a semantic vector using the same text embedding model; then, the cosine similarity between the two semantic vectors is calculated; if the cosine similarity value is greater than or equal to a preset matching threshold, such as 0.85, a successful match is determined. The matching threshold of 0.85 is an empirical value designed to balance matching accuracy and recall. The target entity set is a data structure that stores all successfully matched target category nodes. Each scene constraint in the identified scene constraint list is semantically similar to a predefined scene node in the heterogeneous knowledge graph. Successfully matched scene nodes are added to the scene constraint set. The matching process here is similar to that of target entity matching. The scene constraint text and the name text of the scene node are semantically vectorized and cosine similarity is calculated. The same matching threshold, such as 0.85, is used for judgment. The scene constraint set is a data structure that stores all successfully matched scene nodes. For each target category node in the target entity set, query all target interpretation marker nodes directly connected to it through relation edges in the heterogeneous knowledge graph, and select key target interpretation marker nodes according to preset rules to form a key interpretation marker set. One specific implementation of the preset rules is a marker relevance scoring screening rule. First, calculate a relevance score for each queried target interpretation marker node. This score is obtained by weighted summation of two parts: the first part is the weighted sum of the inverse document frequency values ​​of the target interpretation marker node in the entire sample database. The higher the inverse document frequency value, the rarer the marker is and the greater its distinguishability. The second part is the weighted sum of the maximum cosine similarity between the semantic vector of the description text of the target interpretation marker node and the overall semantic vector of the scene constraint set, which is used to measure the relevance of the marker to the current task scene. When calculating the overall semantic vector, the average of the semantic vectors of the description texts of all scene nodes in the scene constraint set is obtained. Set the weight coefficients of the two parts, for example, 0.6 and 0.4 respectively. Finally, select the top K target interpretation marker nodes with the highest relevance scores to form a key interpretation marker set, where K is a preset positive integer, for example, ten. Combining preset or user-specified task preference parameters, the target entity set, key interpretation marker set, scene constraint set, and task preference parameters are collectively encapsulated to form a dynamic query vector. The task preference parameters are a set of configurable numerical parameters used to guide subsequent sample retrieval preferences, typically encapsulated in vector form. They include at least four dimensions: a diversity weight parameter, controlling the degree of preference for the diversity of sample visual features (default value 0.3); a difficulty preference parameter, controlling the degree of preference for difficult samples (default value 0.2); a scene coverage intensity parameter, controlling the urgency of covering a specified scene (default value 0.4); and a redundancy penalty coefficient, controlling the degree of exclusion of similar samples (default value 0.1). The dynamic query vector is a composite data structure object containing the target entity set, key interpretation marker set, scene constraint set, and task preference parameter vector, serving as a unified input interface for subsequent steps.

[0023] In this embodiment, it is particularly important to explain step S3, which involves executing a walk strategy that combines transition probabilities to dynamically explore sample nodes and construct a subset of candidate samples. The specific process is as follows: Using the target interpretation marker nodes in the key interpretation marker set as initial anchors, all sample nodes directly connected to these initial anchors through attribute edges are located in the heterogeneous knowledge graph. These sample nodes are used as the initial set to be explored, and an empty subset of candidate samples is initialized. The construction of the initial set to be explored ensures that the retrieval starts from the samples most relevant to the core semantics of the task, laying a high-quality starting point for subsequent exploration. The transition probability in the walk strategy is determined by three factors: the first factor is the geometric mean of the association strength between the sample node to be evaluated and all target interpreter nodes in the key interpreter set; to calculate the geometric mean of association strength, first obtain the association strength value of the attribute edge between the sample node to be evaluated and each target interpreter node in the key interpreter set, then multiply these multiple association strength values, and then take the Nth root of the product, where N is the total number of target interpreter nodes in the key interpreter set; using the geometric mean instead of the arithmetic mean can more evenly require the sample node to maintain a high association with all key markers, avoiding the situation where a very strong association with a certain marker masks a very weak association with other markers; The second factor is the matching degree between the sample node to be evaluated and the scene constraint set in the dynamic query vector. If the sample node to be evaluated is connected to any scene node in the scene constraint set through a directed edge representing the relationship of the source scene, the factor is set to one; otherwise, it is set to a very small positive number close to zero. The very small positive number close to zero in the matching degree factor can usually be set to 0.01. Its function is to allow a very small number of samples from non-specified scenes to be explored with a very low probability, providing a theoretical possibility for the algorithm to escape local optima in extreme cases, while ensuring that the walking process is strongly constrained within the scene range specified by the task. The third factor is the cosine similarity between the visual feature vector of the sample node to be evaluated and the average visual feature vector of all sample nodes in the current candidate sample subset. This factor guides the flow towards the unexplored region in the feature space by actively reducing the transition probability of nodes that are similar to the average feature of the existing candidate set. It is the core driving mechanism for automatically injecting sample diversity. When calculating the average visual feature vector, the arithmetic mean of each dimension of the visual feature vector of all sample nodes in the candidate sample subset is calculated. The final transition probability is the result of normalizing the product of the first, second, and third factors. Normalization ensures that the sum of the probabilities of all possible transition directions starting from the current sample node is one. The specific process of normalization is as follows: For all possible neighbor sample nodes starting from the current sample node, calculate the product of the first, second, and third factors for each of them to obtain a series of unnormalized transition tendency values; then divide the tendency value corresponding to each neighbor node by the sum of the tendency values ​​of all neighbor nodes to obtain the normalized transition probability. Each time a new sample node is selected to be added to or replaced in the candidate sample subset based on the transition probability, the overall utility score of the current candidate sample subset is recalculated. This real-time re-evaluation mechanism ensures that every change in the candidate sample subset can be quantified, providing data for the immediate adjustment of the walk strategy. If the new overall utility score is higher than the highest score in the history, the adjustment is accepted and the highest score record is updated; this design preserves the globally optimal state found during the search process. If a predetermined number of consecutive walks and adjustments fail to produce a higher overall utility score, a walk strategy adjustment is triggered. This can be achieved by increasing the weight of the third factor in the transition probability calculation, or by randomly selecting a sample node from the subset of candidate samples with the highest historical score as the new walk starting point, thus changing the exploration direction. The predetermined number of consecutive walks is usually set to five to ten. The specific way to increase the weight of the third factor is to multiply it by an amplification factor greater than one, such as 1.5 or 2.0. Randomly restarting the walk from the historical best subset can effectively help the search process escape the current local optimum and explore new solution spaces. The comprehensive utility score of the candidate sample subset is calculated in real time using an online evaluation function, which is a weighted sum of four utility measures: Diversity measures, difficulty measures, coverage measures, and redundancy measures; Diversity measures are characterized by calculating the information entropy of the distribution of visual feature vectors of all sample nodes in the candidate sample subset within the vector space. The specific process for calculating the information entropy of the visual feature distribution is as follows: First, using the kernel density estimation algorithm, the probability density of the data distribution in the feature space is estimated based on the visual feature vectors of all sample nodes in the candidate sample subset. Then, the product of the estimated probability density function and its natural logarithm is integrated over the entire domain of the feature space, and the negative value is taken. The resulting information entropy value is the diversity measure. This method can more accurately measure the dispersion and uncertainty of the feature distribution, and is superior to simple variance calculation. The difficulty metric evaluates each sample node in the candidate sample subset using a pre-trained neural network model that takes visual feature vectors as input and outputs predicted difficulty values. The arithmetic mean of the predicted difficulty values ​​for all sample nodes is then calculated. The pre-trained difficulty prediction neural network model is a small, fully connected feedforward network that is trained using a large amount of training loss curve data (such as the average loss of each sample in iterations) as a supervision signal during the historical model training task. It learns the mapping relationship from the visual feature vector of a sample to its learning difficulty (represented by the loss value). During evaluation, the output value of this network is normalized to the interval between zero and one, serving as the predicted difficulty value for a single sample. The arithmetic mean is then taken to obtain the average difficulty of the entire subset. Coverage is calculated by counting how many target interpreter nodes in the key interpreter set are covered by sample nodes in the candidate sample subset. A target interpreter node is considered covered if and only if there is at least one sample node in the candidate sample subset with a correlation strength greater than a preset coverage threshold. The preset coverage threshold is usually set to 0.7, which is an empirical value used to determine whether a sample significantly possesses a certain interpreter. The value of the coverage metric is equal to the number of covered target interpreter nodes divided by the total number of nodes in the key interpreter set. Redundancy is measured by calculating the arithmetic mean of the cosine similarity of the visual feature vectors between all unordered sample node pairs in the candidate sample subset. Calculating all unordered sample node pairs means selecting two different nodes from N sample nodes in the candidate sample subset without repetition to form a node pair, resulting in a total of N multiplied by (N minus one) and divided by two pairs. The cosine similarity of the visual feature vectors of each pair of nodes is calculated, and then the arithmetic mean of the similarity of all node pairs is obtained to obtain the redundancy measure. The weighting coefficients for the four dimensions—diversity, difficulty, coverage, and redundancy—are provided by the task preference parameters in the dynamic query vector. The search process continues until a preset upper limit for the total number of steps is reached, or the improvement in the highest overall utility score across multiple iterations is less than a preset small positive threshold. At this point, the subset of candidate samples corresponding to the highest historical overall utility score is determined as the optimal subset of samples that meets the conditions. The preset upper limit for the total number of steps can be set according to the graph size and task complexity, for example, from one thousand to five thousand steps. The small positive threshold can be set to 0.001. The design of the termination condition strikes a balance between computational efficiency and solution quality.

[0024] In this embodiment, it is specifically necessary to explain step S4, which involves determining whether the coverage of the optimal sample subset to a specified combination or a specified scene combination in the key interpretation flag set is lower than a preset threshold. This specifically includes: For each combination of pairs of target interpretation marker nodes in the key interpretation marker set and each combination of pairs of scene nodes in the scene constraint set of the dynamic query vector, the coverage state of the optimal sample subset for each combination is calculated. A "combination of pairs of target interpretation marker nodes" refers to an unordered pair formed by arbitrarily selecting two different target interpretation marker nodes from the key interpretation marker set. A "combination of pairs of scene nodes" refers to an ordered pair formed by arbitrarily selecting two different scene nodes from the scene constraint set of the dynamic query vector, taking into account the hierarchical or sequential relationship between scenes. Evaluating all such combinations is to verify whether the optimal sample subset covers the important feature co-occurrence relationships and complex scene context in the task semantics. A specified combination is considered covered if and only if there exists at least one sample node in the optimal sample subset such that the association strength between that sample node and every node in the combination is greater than a preset coverage strength threshold; the preset coverage strength threshold is usually set to 0.7; for target interpretation marker nodes, the association strength comes directly from the association strength value stored on the attribute edge between the sample node and the target interpretation marker node in the heterogeneous knowledge graph; for scene nodes, it is necessary to determine whether the sample node is directly connected to the scene node through the "from" relation edge. If connected, the association strength is considered to be one; otherwise, it is considered to be zero; only when the comparison result of the association strength between the sample node and all nodes in the combination satisfies the condition of being greater than the threshold is the combination considered to be covered by the current sample node; Subsequently, the proportion of the number of covered specified combinations to the total number of all specified combinations is counted to obtain the combination coverage. The combination coverage is a value between zero and one, reflecting the overall coverage ratio of the optimal sample subset to the task-related combinations. When calculating the total number of all specified combinations, it is the number of pairwise unordered combinations of the key interpretation mark set plus the number of pairwise ordered combinations of the scene constraint set. The calculated combination coverage is compared with a preset combination coverage threshold. If it is lower than the threshold, it is determined that there is insufficient coverage, and all uncovered specified combinations are identified as targets for triggering programmatic sample synthesis. The preset combination coverage threshold is usually set to 0.8. If the combination coverage is lower than this threshold, it means that the currently retrieved optimal sample subset has a significant gap in semantic combination coverage, and samples need to be actively synthesized to make up for it. All the uncovered specified combinations constitute a clear and specific list of semantic targets for programmatic sample synthesis. The weight parameters of the online evaluation function for programmatic sample synthesis and subsequent updates, as well as the utility weights of sample nodes in heterogeneous knowledge graphs, specifically include: For each uncovered specified combination, based on the semantic and structural constraint descriptions of each node within the combination in the interpretation label library, high-quality sample image patches that satisfy the constraints of a single node are retrieved and segmented from the sample library. High-quality sample image patches refer to the image regions corresponding to the original samples that are visually clear, accurately labeled, and strongly associated with the target interpretation label nodes. Segmentation refers to using the labeling information of the samples to accurately crop the target instance from the original image to form independent, transparent image patches. Based on physical laws and scene context knowledge, a constraint satisfaction optimization model is used to calculate the optimal spatial layout of each high-quality sample image patch on a selected background image. The goal of the constraint satisfaction optimization model is to maximize the satisfaction of spatial relationship constraints derived from the interpretation marker library. Specifically, the constraint satisfaction optimization model can employ an optimizer based on simulated annealing or a genetic algorithm. Spatial relationship constraints are extracted from the textual descriptions in the interpretation marker library, such as "above," "nearby," and "inside," and are quantified as mathematical inequality constraints such as relative position, distance, and overlap between target instances. The objective function of the optimization model is defined as the weighted sum of all satisfied constraints, with weights representing the importance of the constraints. By iteratively adjusting the position, rotation, and scale of each sample image patch on the background image, the layout scheme that maximizes the objective function value is found, which is the optimal spatial layout. The selected background image is randomly chosen from images in the sample library that conform to the current task scene. A pre-trained generative adversarial network (GAN) is used to perform illumination, shadow, and edge fusion processing on high-quality sample image patches based on the calculated optimal spatial layout, generating a synthetic image. The pre-trained GAN consists of a generator and a discriminator, both trained on a large number of real remote sensing image patches that naturally blend the target and background. In the synthesis stage, the generator takes a background image stitched together with sample image patches (placed according to the optimal spatial layout) as input and outputs an image after illumination adjustment, shadow rendering, and edge feathering. The discriminator is used to assist in training the generator to ensure that its output is visually indistinguishable from real images. Finally, the image output by the generator is the synthetic image. A realism-based discriminant neural network is used to verify the quality of synthetic images. Only images whose realism confidence score output by the network exceeds a preset strict quality threshold are considered acceptable. This realism-based discriminant neural network is a separate, specially trained binary convolutional neural network used to determine whether an input image is a natural image rather than a synthetic product. The training data includes real remote sensing images and images synthesized using various methods. The network takes the synthetic image as input and outputs a realism confidence score between zero and one, representing the probability that the image is considered a real image. The preset strict quality threshold is typically set to 0.9 to ensure that only high-quality synthetic images proceed to subsequent processes. To automatically generate pseudo-labels for validated synthetic images, synthetic samples are added to the optimal sample subset to obtain an enhanced sample set. The automatic generation of pseudo-labels is achieved using a lightweight object detection model pre-trained on an existing sample database. This model is used to infer on the synthetic images, and its output target bounding boxes and class predictions, after post-processing such as non-maximum suppression, serve as the pseudo-labels for the synthetic images. Synthetic samples refer to paired data of synthetic images and their pseudo-labels; adding these to the optimal sample subset forms the enhanced sample set. The system collects performance metrics obtained from training the model using augmented sample sets. Based on these metrics, it adjusts the weight parameters of various utility metrics in the online evaluation function using optimization algorithms. Furthermore, it updates the utility weight attributes of corresponding sample nodes in the heterogeneous knowledge graph based on the gradient contributions of each sample node in the augmented sample set during model training. Performance metrics typically refer to the average precision of the model on the independent validation set. The optimization algorithm for adjusting the weight parameters of the online evaluation function can employ gradient-based meta-learning algorithms or performance-ranked heuristic search algorithms. Specifically, a surrogate model is established between the weight parameters and the final model performance, using gradient descent to find weight parameter directions that can improve performance. When updating the utility weights of sample nodes, the gradient contribution is approximated by recording the time average of the L2 norm of the gradient vector generated by the feature vector of each sample node during model training in the loss function calculation. The update formula for the utility weight attribute is: the new utility weight equals the old utility weight plus a learning rate multiplied by the average gradient norm of the sample node. This mechanism ensures that sample nodes that contribute significantly to model training receive higher attention in future searches. Example 2

[0025] This embodiment provides, for example Figure 2 The automated remote sensing image sample production system based on interpretation tag-driven architecture, as shown, specifically includes: The knowledge graph construction module extracts visual features of samples and obtains predefined interpretation symbols based on the sample library and interpretation symbol library. It then constructs a heterogeneous knowledge graph containing sample nodes, target category nodes, interpretation symbol nodes and scene nodes to form a structured semantic network. Query generation module: Based on task description text and heterogeneous knowledge graph, it parses and identifies target entities and scene constraints, completes the mapping with graph nodes, and generates dynamic query vectors by encapsulating target entity set, key interpretation mark set, constraint conditions and task preference parameters; Intelligent retrieval module: Taking dynamic query vectors as input, starting from the set of key interpretation markers in the heterogeneous knowledge graph, it performs a walk and online evaluation that combines transition probabilities, it iteratively adjusts the walk strategy and calculates the comprehensive utility score of the candidate sample subset, and searches for and determines the optimal sample subset; Enhancement and optimization module: When the coverage of the optimal sample subset for the specified interpretation tag combination or scene combination is lower than the preset threshold, the programmatic sample synthesis and quality verification are triggered. An enhanced sample set is obtained by generating and adding synthetic samples for model training, and training performance feedback is collected to update the weight parameters of the online evaluation function and the utility weights of sample nodes in the heterogeneous knowledge graph.

[0026] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0027] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0028] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0029] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0030] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0031] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0032] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for automated production of remote sensing image samples based on interpretation tag-driven methods, characterized in that, Specifically, the following steps are included: Step S1: Obtain sample metadata, annotation information and associated original images from the sample library, and extract sample visual features; Synchronously acquire predefined target interpretation tags from the interpretation tag library; based on this, construct a heterogeneous knowledge graph containing sample nodes, target category nodes, interpretation tag nodes, and scene nodes, and form a structured semantic network; Step S2: Parse the target task description text, identify target entities and scene constraints, and map them to corresponding nodes in the heterogeneous knowledge graph to form a dynamic query vector containing a set of target entities, a set of key interpretation markers, constraints, and task preference parameters. Step S3: Using the dynamic query vector as input, and the key interpretation marker set as the initial anchor point in the heterogeneous knowledge graph, a walking strategy combining transition probabilities is executed to dynamically explore sample nodes and construct candidate sample subsets; through an online evaluation function, the comprehensive utility score of the candidate sample subset is calculated in real time, and the walking strategy is iteratively adjusted based on the feedback of the comprehensive utility score until the optimal sample subset that meets the conditions is found. Step S4: If the coverage of the optimal sample subset to the specified combination or specified scene combination in the key interpretation label set is lower than the preset threshold, then the procedural sample synthesis is triggered to generate a synthetic image; pseudo-labels are generated for the synthetic image that passes the quality verification and added to the optimal sample subset as synthetic samples to obtain an enhanced sample set for model training. Collect training performance feedback and update the weight parameters of the online evaluation function and the utility weights of sample nodes in the heterogeneous knowledge graph.

2. The automated production method for remote sensing image samples based on interpretation flags as described in claim 1, characterized in that: In step S1, the visual features of the samples are extracted. The specific process is as follows: Using a deep convolutional neural network model pre-trained on a large-scale remote sensing image dataset, forward computation is performed on the corresponding region of the original image associated with each sample in the sample library. The feature map output by the last convolutional layer of the deep convolutional neural network model is extracted, and global average pooling is performed on the feature map in the spatial dimension to obtain a fixed-dimensional multi-dimensional real vector that can represent the visual content of the sample. This multi-dimensional real vector is defined as the visual feature vector.

3. The automated production method for remote sensing image samples based on interpretation flags as described in claim 2, characterized in that: Construct a heterogeneous knowledge graph containing sample nodes, target category nodes, interpretation marker nodes, and scene nodes, specifically including: Create a sample node for each sample in the sample library and assign a corresponding visual feature vector to the sample node; Create a target interpretation flag node for each target interpretation flag in the interpretation flag library; create target category nodes for all known target categories; create scene nodes for all known scene types; Establish directed edges representing the category affiliation relationship between sample nodes and target category nodes; Establish directed edges representing the relationship between target category nodes and target interpretation marker nodes; Establish directed edges representing the relationship between the source scene and the sample nodes; Establish an attribute edge with quantifiable weights between the sample node and the target interpretation marker node; The calculation process of attribute edge weights is as follows: First, a trainable attention mechanism network is used to process the visual feature vectors of sample nodes in order to focus on the feature components related to the description of the target interpreter node. Simultaneously, a trainable projection mapping function is used to convert the text or feature descriptions of the associated target interpretation marker nodes into vector representations of the same dimension; Then, the element-wise product of the feature vector output by the attention mechanism network and the vector output by the projection mapping function is calculated, and then mapped to a real value between zero and one through a learnable linear transformation layer and a non-linear activation function. This real value is the association strength. The structured semantic network is composed of sample nodes, target category nodes, target interpretation marker nodes, scene nodes, and directed edges connecting these nodes and attribute edges with association strength.

4. The automated production method for remote sensing image samples based on interpretation flags as described in claim 3, characterized in that: In step S2, the target task description text is parsed and the target entities and scene constraints are identified, specifically including: Pre-trained natural language processing models are used to perform sequence labeling and dependency parsing on target task description text; Lexical units representing specific geographical features in the target task description text are extracted through sequence labeling operations and identified as target entities. By using dependency parsing to extract lexical units that modify the target entity and represent its environment and time, these units are identified as scene constraints; ultimately, a structured list of target entities and a list of scene constraints are obtained.

5. The automated production method for remote sensing image samples based on interpretation flags as described in claim 4, characterized in that: Mapping to corresponding nodes in heterogeneous knowledge graphs and forming dynamic query vectors, specifically including: Each target entity in the identified target entity list is semantically matched with a predefined target category node in the heterogeneous knowledge graph, and the successfully matched target category node is added to the target entity set. Each scene constraint in the identified scene constraint list is semantically similar to a predefined scene node in the heterogeneous knowledge graph, and the successfully matched scene nodes are added to the scene constraint set. For each target category node in the target entity set, query all target interpretation marker nodes that are directly connected to it through relation edges in the heterogeneous knowledge graph, and filter out the key target interpretation marker nodes according to preset rules to form a key interpretation marker set; By combining preset task preference parameters, the target entity set, key interpretation flag set, scene constraint set, and task preference parameters are encapsulated together to form a dynamic query vector.

6. The automated production method for remote sensing image samples based on interpretation flags according to claim 5, characterized in that: In step S3, a walk strategy incorporating transition probabilities is executed to dynamically explore sample nodes and construct a subset of candidate samples. The specific process is as follows: Using the target interpretation marker nodes in the key interpretation marker set as initial anchors, locate all sample nodes in the heterogeneous knowledge graph that are directly connected to these initial anchors through attribute edges, use these sample nodes as the initial set to be explored, and initialize an empty subset of candidate samples. The transition probability in the walk strategy is determined by three factors: the first factor is the geometric mean of the association strength between the sample node to be evaluated and all target interpreter node in the key interpreter set; The second factor is the matching degree between the sample node to be evaluated and the set of scene constraints in the dynamic query vector. If the sample node to be evaluated is connected to any scene node in the set of scene constraints through a directed edge representing the relationship of the source scene, the factor takes a value of one; otherwise, it takes a very small positive number close to zero. The third factor is a subtraction of the cosine similarity between the visual feature vector of the sample node to be evaluated and the average visual feature vector of all sample nodes in the current candidate sample subset. The final transition probability is the product of the first factor, the second factor, and the third factor, after normalization. Normalization ensures that the sum of the probabilities of all possible transition directions starting from the current sample node is one. Each time a new sample node is selected to be added to or replaced in the candidate sample subset based on the transition probability, the overall utility score of the current candidate sample subset is recalculated. If the new overall utility score is higher than the highest score in history, then the adjustment will be accepted and the highest score record will be updated. If a predetermined number of consecutive walks and adjustments fail to produce a higher overall utility score, a walk strategy adjustment is triggered. This involves either increasing the weight of the third factor in the transition probability calculation or randomly selecting a sample node from the highest-scoring candidate sample subset in the historical records as a new walk starting point to change the exploration direction.

7. The automated production method for remote sensing image samples based on interpretation flags as described in claim 6, characterized in that: The comprehensive utility score of the candidate sample subset is calculated in real time using an online evaluation function, which is a weighted sum of four utility measures: Diversity measures, difficulty measures, coverage measures, and redundancy measures; Diversity measures are characterized by calculating the information entropy of the distribution of visual feature vectors of all sample nodes in a subset of candidate samples in the vector space. The difficulty metric evaluates each sample node in the candidate sample subset using a pre-trained neural network model that takes visual feature vectors as input and outputs predicted difficulty values, and takes the arithmetic mean of the predicted difficulty values ​​of all sample nodes. Coverage is measured by how many target interpreter nodes in the key interpreter set are covered by sample nodes in the candidate sample subset. A target interpreter node is considered covered if and only if there is at least one sample node in the candidate sample subset with a correlation strength greater than a preset coverage threshold. Redundancy is measured by calculating the arithmetic mean of the cosine similarity of the visual feature vectors between all unordered sample node pairs in the candidate sample subset. The weight coefficients of the four dimensions—diversity measure, difficulty measure, coverage measure, and redundancy measure—are provided by the task preference parameters in the dynamic query vector. The search process continues until the preset maximum number of total steps is reached, or the improvement of the highest comprehensive utility score in multiple consecutive iterations is less than a preset small positive threshold. At this point, the candidate sample subset corresponding to the highest historical comprehensive utility score is determined as the optimal sample subset that meets the conditions.

8. The automated production method for remote sensing image samples based on interpretation flags according to claim 7, characterized in that: In step S4, it is determined whether the coverage of the optimal sample subset for a specified combination or a specified scene combination in the key interpretation marker set is lower than a preset threshold, specifically including: For each combination of pairs of target interpretation marker nodes in the key interpretation marker set and each combination of pairs of scene nodes in the scene constraint set in the dynamic query vector, calculate the coverage state of the optimal sample subset for each combination. A specified combination is considered covered if and only if there exists at least one sample node in the optimal sample subset such that the association strength between that sample node and every node in the combination is greater than a preset coverage strength threshold. Subsequently, the proportion of the number of covered specified combinations to the total number of all specified combinations is calculated to obtain the combination coverage. The calculated combination coverage is compared with a preset combination coverage threshold. If it is lower than the combination coverage threshold, it is determined that there is insufficient coverage, and all the specified uncovered combinations are identified as targets to trigger programmatic sample synthesis.

9. The automated production method for remote sensing image samples based on interpretation flags as described in claim 8, characterized in that: The weight parameters of the online evaluation function for programmatic sample synthesis and subsequent updates, as well as the utility weights of sample nodes in heterogeneous knowledge graphs, specifically include: For each uncovered specified combination, based on the semantic and structural constraint descriptions of each node in the combination in the interpretation tag library, high-quality sample image patches that satisfy the constraints of a single node are retrieved and segmented from the sample library. Based on physical laws and scene context knowledge, a constraint satisfaction optimization model is used to calculate the optimal spatial layout of each high-quality sample image patch on a selected background image. The goal of the constraint satisfaction optimization model is to maximize the satisfaction of the spatial relationship constraints derived from the interpretation label library. Using a pre-trained generative adversarial network, high-quality sample image patches are processed for illumination, shadow and edge fusion based on the calculated optimal spatial layout to generate synthetic images. The quality of the synthetic image is verified by a realism discrimination neural network. The image is considered to pass only when the realism confidence score output by the realism discrimination neural network is higher than a preset strict quality threshold. To automatically generate pseudo-labels for the verified synthetic images, synthetic samples are formed and added to the optimal sample subset to obtain the enhanced sample set; The system collects performance metrics obtained from training the model using augmented sample sets. Based on these performance metrics, it adjusts the weight parameters of each utility metric in the online evaluation function using an optimization algorithm. It also updates the utility weight attributes of the corresponding sample nodes in the heterogeneous knowledge graph based on the gradient contribution of each sample node in the augmented sample set during the model training process.

10. An automated remote sensing image sample production system based on interpretation tag-driven method, applied to any one of claims 1-9, characterized in that: Specifically, it includes: The knowledge graph construction module extracts visual features of samples and obtains predefined interpretation symbols based on the sample library and interpretation symbol library. It then constructs a heterogeneous knowledge graph containing sample nodes, target category nodes, interpretation symbol nodes and scene nodes to form a structured semantic network. Query generation module: Based on task description text and heterogeneous knowledge graph, it parses and identifies target entities and scene constraints, completes the mapping with graph nodes, and generates dynamic query vectors by encapsulating target entity set, key interpretation mark set, constraint conditions and task preference parameters; Intelligent retrieval module: Taking dynamic query vectors as input, starting from the set of key interpretation markers in the heterogeneous knowledge graph, it performs a walk and online evaluation that combines transition probabilities, it iteratively adjusts the walk strategy and calculates the comprehensive utility score of the candidate sample subset, and searches for and determines the optimal sample subset; Enhancement and optimization module: When the coverage of the optimal sample subset for the specified interpretation tag combination or scene combination is lower than the preset threshold, the programmatic sample synthesis and quality verification are triggered. An enhanced sample set is obtained by generating and adding synthetic samples for model training, and training performance feedback is collected to update the weight parameters of the online evaluation function and the utility weights of sample nodes in the heterogeneous knowledge graph.

Images