An ontology logic constraint-based remote sensing image generative interpretation method and system

Abstract

The application provides a remote sensing image generative interpretation method and system based on ontology logic constraint, first constructs an entity-attribute-relation-rule four-layer geographic ontology knowledge system, quantizes feature attribute threshold, spatial topological relation and scene distribution logic into computable constraint rules, encodes and generates structured ontology knowledge features; secondly, the input remote sensing image is subjected to degradation repair and multi-scale layered feature extraction, and the obtained global image features are adaptively fused; then, two types of features are fused in a double mode, dynamic geographic logic constraints are embedded in the whole process of reverse denoising of the diffusion model, and an initial interpretation result is generated; and finally, the global error screening and local progressive fitting are iteratively optimized, and the interpretation quality is quantitatively evaluated from four dimensions of pixel, semantics, topology and scene. The application adopts a knowledge and data dual driving paradigm, replaces the traditional pure data driving and fragmented rule constraint mode, and effectively improves the geographic rationality, precision and robustness of the interpretation of complex degraded remote sensing images.

Description

Technical Field

[0001] This invention belongs to the field of intelligent interpretation of remote sensing images and generative artificial intelligence technology, specifically involving an intelligent interpretation method and system for remote sensing images that integrates geographic ontology logical rules and diffusion generation models. Background Technology

[0002] Generative interpretation of remote sensing images relies on generative artificial intelligence technology to complete semantic parsing of remote sensing scenes, identification of ground objects, and reconstruction of incomplete scenes. It can adapt to complex remote sensing images such as blurry, occluded, incomplete, and low-quality images, and is of great significance for improving the information mining and analysis capabilities of remote sensing big data. Currently, mainstream generative interpretation technologies for remote sensing are mainly divided into three categories: The first category is generative interpretation technology that integrates fragmented knowledge rules. This technology optimizes the basic generative model by embedding scattered artificial geographic rules and semantic label constraints, mitigating semantic confusion and improving the reasonableness of the results to some extent. However, this type of technology lacks a hierarchical, systematic, and structured geographic ontology knowledge system. The artificial rules have narrow coverage, weak correlation, and poor universality. Constraints are prone to failure in complex scenarios such as cloud cover and partial incompleteness, making it difficult to suppress semantic conflicts, scene distortion, and feature confusion. The second category is generative interpretation technology based on generative adversarial networks. This technology relies on adversarial training to complete image texture optimization, sample expansion, and feature detail generation, achieving high texture fidelity and good adaptability to common remote sensing scenes. However, this type of technology suffers from poor training stability, lacks prior geographic knowledge constraints, and the interpretation results are prone to spatial topological confusion and semantic contradictions, failing to conform to real geographic laws, resulting in low overall interpretation accuracy. The third category is generative interpretation technology based on diffusion models. This technology achieves high-quality image generation and semantic interpretation through multi-step progressive denoising iteration, exhibiting stable training, high accuracy, and strong generalization ability. However, this type of technology only mines shallow features such as pixel grayscale and texture, and cannot identify the inherent topological relationships, semantic matching rules and spatial distribution logic between ground features, which can easily generate pseudo ground features that violate common geographical sense.

[0003] In summary, existing generative remote sensing interpretation technologies generally suffer from the following technical shortcomings: First, existing knowledge-assisted technologies rely solely on fragmented manual rules, lacking a standardized ontology hierarchy and thus failing to adapt to complex remote sensing scenarios. Second, existing technologies all employ a purely data-driven learning model, lacking systematic geographic ontology logical constraints, making it easy for the generated interpretation results to violate objective geographic laws. Third, when faced with complex and degraded remote sensing images, the interpretation accuracy is low, the number of false targets generated is large, and the stability of the results is poor, severely restricting the engineering application of generative remote sensing interpretation technologies. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned defects of the prior art and to provide a method and system for generative interpretation of remote sensing images based on ontological logical constraints. This method constructs a standardized, hierarchical, and quantifiable geographic ontological knowledge system, integrates three types of logical rules—attribute constraints, semantic constraints, and topological constraints—into the entire process of generative interpretation of remote sensing images, and replaces the traditional pure data-driven paradigm and fragmented manual rule constraint mode. This effectively improves the geographic rationality, scene integrity, and interpretation accuracy of complex and degraded remote sensing images.

[0005] According to one objective of the present invention, the present invention provides a generative interpretation method for remote sensing images based on ontology logic constraints, comprising the following steps: Step 1: Construct a set of logical constraint rules and a geographic ontology knowledge system; Based on remote sensing industry standards and prior knowledge of samples, establish a hierarchical geographic ontology knowledge system that includes entity layer, attribute layer, relationship layer and logical rule layer, quantify and generate computable rules for feature attribute constraints, topological relationship constraints and scene rationality constraints, and encode and generate ontology knowledge structured features, while constructing a standard feature library; Step 2: Remote sensing image processing and multi-scale feature encoding; The input remote sensing image is preprocessed and degraded, and an improved hierarchical Vision Transformer encoder is used to extract low-level detail features, mid-level structural features and top-level global semantic features. The global fusion features are obtained by adaptive weighted fusion. Step 3: Diffusion-generative interpretation with ontology logical constraints; adaptively weightedly fuse the global fusion features with the ontology knowledge structured features to obtain dual-modal fusion features, perform forward noise addition using the dual-modal fusion features as the basis, embed global comprehensive constraint coefficients during the reverse iterative denoising process to dynamically correct the denoising results, and decode to generate the initial interpreted image after iterative convergence; Step 4: Dynamic correction and error optimization iteration based on ontology logic; Based on the set of geographic ontology logic constraint rules, the initial interpreted image is subjected to global error screening, abnormal feature areas are located, constraint weights are adaptively updated, and local feature progressive fitting correction is performed on abnormal areas in combination with the standard feature library. After iterative convergence, the optimized interpreted image is output. Step 5: Quantitative evaluation of interpretation results; Single-dimensional quantitative evaluation is carried out from four dimensions: pixel reconstruction accuracy, ground feature semantic recognition accuracy, spatial topology compliance rate and scene completeness, and a weighted comprehensive quality score is calculated to output the final interpretation results.

[0006] Furthermore, step 1 specifically includes: S101. Land Feature Attribute Threshold Quantification: For seven types of land features, namely water area, forest land, building, road, cultivated land, bare land and mountain, determine the upper and lower limits of attribute thresholds for gray scale, texture, size and elevation, construct a land feature attribute threshold library, and determine the compliance of single land feature attributes through the binary constraint formula of land feature attributes. S102. Construction of Topological Relationship Set: Define five types of topological relationships: isolation, adjacency, inclusion, connectivity, and mutual exclusion. Establish a standard topological relationship set for any two types of land features. Determine the spatial topological compliance of two land features through the land feature topological binary constraint formula. S103, Geographical rule quantification and transformation: Calculate the comprehensive rationality coefficient of a single feature and the comprehensive rationality of a scene respectively, and construct a two-level logical rule constraint system for individual features and global scenes; S104. Geographic ontology knowledge structured feature generation: Generate feature attribute compliance features, feature topology compliance features, and scene rationality features. After channel splicing, input them into the knowledge encoder for convolution and pooling encoding, and output ontology knowledge structured features. S105. Construction of Standard Land Feature Library: Calculate the mean feature of standard sample image blocks for each type of land feature, and construct a standard land feature library that maps land feature categories one-to-one with standard feature templates.

[0007] Furthermore, step 2 specifically includes: S201. Image preprocessing and degradation restoration: The original image is grayscale normalized and mapped to the [0,1] interval. The multi-scale Retinex algorithm is used to eliminate interference from clouds and uneven lighting. Then, bilateral filtering is used to remove noise and preserve edges to obtain the normalized image. S202, Multi-scale hierarchical feature extraction: A three-level cascaded improved Vision Transformer hierarchical encoder is adopted, which sequentially passes through convolutional downsampling and window multi-head self-attention module to output low-level detailed features, mid-level structural features and top-level global semantic features; S203, Multi-scale Feature Adaptive Fusion: The bottom and middle layer features are downsampled to the same size as the top layer features, and after being convolved with a 1×1 to unify the number of channels, adaptive weights are generated based on the global statistical features of the image, and weighted fusion is performed to obtain the global fusion features.

[0008] Furthermore, in step 3, the dual-modal feature fusion formula is:

[0009] In the formula, For dual-modal fusion features, F all For the feature of full-domain integration, F know As a structured feature of ontology knowledge, The dual-feature balanced weighting coefficient has a value range of [0,1]. For normal images, Use a value of 0.7-0.9; for degraded images, A value of 0.3-0.6 indicates more severe image degradation. The smaller; The diffusion model has a total of 1000 forward noise steps, and the noise intensity is... from The value is linearly increased to 0.02, and a continuous degradation feature sequence is constructed using the dual-modal fusion feature as the noise-adding basis.

[0010] Furthermore, in step 3, the reverse iterative denoising process specifically includes: The noisy features of the current step are used to generate a temporary land cover category label map through a lightweight classification head. After connected component labeling, the overall scene rationality is calculated as the overall constraint coefficient C for the entire domain. total ; The noisy features and time steps are input into the U-Net inverse denoising network to obtain the predicted denoised mean. The predicted mean is then dynamically weighted and corrected using a denoising iterative formula that embeds global comprehensive constraint coefficients. The formula is as follows: ; In the formula, For the first The previous clean feature output after each iteration of denoising is the optimized feature result after this iteration, which is corrected and conforms to the geographical logic. To predict the denoised mean, To adjust the strength coefficient, a value of 0.5 is used. The noise variance coefficient is adaptive and decays with the number of iterations. The noise is Gaussian random noise that follows a standard normal distribution. When the difference between the global comprehensive constraint coefficients of two consecutive steps is less than 10 -3 When the iteration converges, the initial decoded image is obtained.

[0011] Furthermore, in step 4, the error screening and weight update specifically involve: Connectivity components are marked on the initial interpreted image, and the logical error of each individual feature and the overall error of the entire region are calculated to locate anomalous feature areas. Combining global compliance with adaptive constraint weight updates, the formula is: ; In the formula, These are the updated adaptive constraint weights; The initial preset baseline weight for the system is set to 0.5; For adaptive adjustment index and .

[0012] Furthermore, in step 4, the local feature correction and iterative convergence specifically involve: For abnormal regions where the logical error of a single feature exceeds the constraint weight, the corresponding local feature blocks are cropped, and progressive fitting compensation is performed using a standard feature library. The correction formula is as follows: ; In the formula, This refers to the standardized local features after correction. Generate bias features for the original model; These are the standard features corresponding to the land cover categories; To correct the step size coefficient and control the correction magnitude in a single iteration; When the difference between the global comprehensive constraint coefficients of two adjacent iterations is less than 10 -5 When the optimization converges, the optimized interpretation image is output.

[0013] Furthermore, in step 5, the single-dimensional quantitative evaluation specifically includes: Pixel reconstruction accuracy is the pixel matching ratio between the interpreted image and the ground truth image; The accuracy of geographic feature semantic recognition is the precision rate, which is the proportion of correctly identified samples out of the total number of identified samples; The spatial topology compliance rate is the proportion of compliant topology association pairs to the total number of topology association pairs; Scene integrity is 1 minus the ratio of the area of ​​the incomplete area to the total effective area.

[0014] Furthermore, the comprehensive quality score is calculated using a weighted product combined with a fluctuation penalty, as shown in the formula: ； In the formula, The overall quality evaluation score for remote sensing interpretation results; In order: pixel precision, semantic precision, topology compliance rate, and scene completeness; The adaptive evaluation weights for the four indicators satisfy the normalization constraint: ; The standard deviation of the four indicators is calculated after normalization and is used to represent the degree of consistency fluctuation of multiple indicators. This is the fluctuation penalty coefficient, used to adaptively deduct points to correct interpretation results with unbalanced indicators and large evaluation fluctuations.

[0015] According to another objective of the present invention, the present invention provides a remote sensing image generative interpretation system based on ontology logic constraints for implementing the above-described method, comprising: The ontology knowledge construction module is used to construct a set of logical constraint rules and a geographic ontology knowledge system, quantify the constraints of land feature attributes, topology and scene rationality, and generate ontology knowledge structured features and a standard land feature feature library. The image coding module is used to preprocess and restore the input remote sensing images, extract multi-scale image features and adaptively fuse them to obtain global fusion features; The diffusion interpretation module is used to fuse global fusion features and ontology knowledge structured features, perform forward noise addition and reverse iterative noise reduction of the diffusion model with ontology logic constraints, and generate the initial interpreted image; The iterative correction module is used to perform global error screening based on ontology logic rules, adaptively update constraint weights, perform local feature progressive fitting correction on abnormal regions, and output optimized interpretation images. The quantitative evaluation module is used to evaluate accuracy from four dimensions: pixels, semantics, topology, and scene, and calculate a comprehensive quality score to output the final interpretation results.

[0016] The technical solution of this invention constructs a hierarchical and quantifiable geographic ontology knowledge system, transforming scattered geographic rules into systematic and computable constraints, thereby suppressing conflicts in geographic feature attributes, topological inconsistencies, and semantic mismatches at their source. It adopts a dual-driven paradigm of knowledge and data, embedding ontology logical constraints into the entire diffusion generation process, compensating for the lack of geographic priors in purely data-driven models and adapting to complex and degraded remote sensing scenarios. Combined with closed-loop iterative correction and multi-dimensional quantitative evaluation mechanisms, it effectively reduces the amount of pseudo-features generated, significantly improving interpretation accuracy, geographic rationality, and result stability, and can widely support various intelligent remote sensing analysis applications. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall process of the remote sensing image generative interpretation method based on ontology logic constraints according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the knowledge encoder according to an embodiment of the present invention; Figure 3 This is a system structure block diagram according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0021] Furthermore, 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 invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0022] Example 1 This invention is implemented through five steps: constructing a set of logical constraint rules and a geographic ontology knowledge system; processing remote sensing images and encoding multi-scale features; performing diffusion-generative interpretation that integrates ontology logical constraints; performing dynamic correction and error optimization iteration based on ontology logic; and quantitatively evaluating the interpretation results. Figure 1 As shown, the specific content is as follows: Step 1: Construct a set of logical constraint rules and a geographic ontology knowledge system S101, Threshold Quantization of Land Feature Attributes.

[0023] Based on national and Ministry of Natural Resources official remote sensing industry standards and specifications such as the "Land Use Dynamic Remote Sensing Monitoring Regulations," the "Technical Regulations for the Third National Land Survey," and the "Technical Specifications for Ecological Protection Red Line Supervision," and combined with prior knowledge of statistical analysis of massive high-definition, cloud-free standard remote sensing samples, this invention specifies the upper and lower limits of attribute thresholds for grayscale, texture, size, and elevation for seven types of land features: water areas, forest land, buildings, roads, cultivated land, bare land, and mountains. This forms a standardized, computable threshold library for ground features, with specific values ​​as follows:

[0024] To determine whether the attributes of a single land feature conform to objective geographical laws, a binary constraint formula for land feature attributes is adopted:

[0025] In the formula: It refers to a single ground object in a remote sensing image scene, including seven types of ground objects: water, woodland, buildings, roads, farmland, bare land, and mountains; Ground features The measured attributes (grayscale, texture, size, elevation, terrain relief); , These are the upper and lower limits of the threshold for land feature attributes. For mountains, there are no hard thresholds for grayscale and size; only topographic relief and elevation are verified. For the compliance coefficient of a single land feature attribute, This indicates that the land feature attributes are compliant. This indicates abnormal land feature attributes.

[0026] In this invention, ground features This represents a spatially connected set of ground feature pixels in the remote sensing image interpretation results. In the actual interpretation process (steps three and four), by labeling the ground feature classification map (each pixel is assigned a label for one of seven ground feature categories: water, woodland, building, road, farmland, bare land, and mountain) with eight connected components, spatially adjacent pixel regions belonging to the same category can be grouped into an independent entity. Each entity has attributes such as area, average grayscale, texture features, and minimum bounding rectangle size. All elements involved in step one... The formula ( , , All values ​​(etc.) are based on entity-level definitions, and the specific values ​​will be calculated in subsequent steps based on the current interpretation results.

[0027] S102, Construction of Topological Relationship Sets Based on the fundamental principles of geospatial topology, the relationships between any two types of land features are analyzed, classifying them into five categories: isolation, adjacency, inclusion, connectivity, and mutual exclusion. A standard set of topological relationships is then established. Clearly define the permissible spatial combinations of different land features to avoid unreasonable spatial arrangements; The five types of topological relationships are defined in this invention: Isolation: There is no physical contact between the objects in the two locations; Adjacent: Boundary contact, no overlap; Includes: a feature that is entirely located inside another feature; Connectivity: Spatially connected features of the same type; Mutually exclusive: Adjacent or overlapping objects are prohibited (e.g., water bodies and roads should generally not overlap, but adjacency is allowed; actual mutual exclusion relationships are set according to industry standards).

[0028] Establish a standard set of topological relationships Topological binary constraint formula for terrain features:

[0029] In the formula: It refers to a single ground entity in a remote sensing image scene, including seven types of ground features: water, woodland, buildings, roads, farmland, bare land, and mountains. To and Another spatially adjacent ground feature. For ground features Its neighboring entities The topology compliance coefficient is used to construct the topology constraint relationship between pairs of features; Ground features and neighboring features The actual spatial topological relationships; A set of standard topological relationships; The spatial topological relationships of the ground features are compliant and reasonable. This indicates that the topological relationships of geographical features are disordered and violate common geographical knowledge.

[0030] S103, Quantitative Transformation of Geographical Rules The fragmented geographical prior rules described by traditional natural language are uniformly mapped into quantitative functions that can be recognized and calculated by the model. This transforms fragmented geographical rules such as compliance of feature size, normal grayscale, elevation adaptation, topology matching, and scene rationality, and constructs a logical rule constraint system from the dimensions of individual features and the global scene.

[0031] The formula for the comprehensive rationality coefficient of a single feature is mainly used to evaluate the comprehensive rationality by integrating its own attributes with the topology of its neighborhood.

[0032]

[0033] In the formula: This refers to the comprehensive rationality coefficient of a single land feature; In relation to land features The set of 8 spatially connected features in the neighborhood. When there are no neighboring features. ,definition .

[0034] The scene comprehensive rationality formula is used to evaluate the overall geographic logical rationality of the entire remote sensing scene.

[0035]

[0036] In the formula: The number of independent ground features after the entire image is segmented by connected components is automatically counted by the connected component labeling algorithm. The comprehensive rationality coefficient of the scene is as follows: the closer the value is to 1, the higher the overall geographical logic rationality of the distribution of ground features, topological layout, and attribute characteristics of the remote sensing scene.

[0037] S104. Generation of Structured Features of Geographic Ontology Knowledge Geographic Ontology Definition: The "geographic ontology" as used in this invention refers to a standardized system for the formal and structured description of land cover types, attributes, spatial relationships, and logical rules in the field of remote sensing image interpretation. Specifically, it includes: Entity layer: Defines seven core land feature types: water area, woodland, buildings, roads, farmland, bare land, and mountains.

[0038] Attribute layer: Specifies the quantization threshold range for attributes such as grayscale, texture, size, and elevation for each type of land feature.

[0039] Relationship layer: Defines the spatial topological relationships (isolation, adjacency, containment, connectivity, mutual exclusion) between features.

[0040] Logical rule layer: Transforms prior geographical knowledge into computable quantifiable functions, including attribute compliance functions. Topology compliance functions Scene rationality function These functions take ground features as input and output compliance coefficients, which are used for subsequent interpretation of constraints and corrections.

[0041] The above four layers together constitute a hierarchical and quantifiable geographical ontology knowledge system.

[0042] Based on the land feature attribute threshold library constructed by S101, the standard topological relationship set constructed by S2, and the scene comprehensive and reasonable quantification rules generated by S3, the geographic ontology prior knowledge of the current remote sensing scene is uniformly structured and encoded, and the discrete land feature attribute rules, spatial topology rules, and scene compliance rules are transformed into ontology knowledge features.

[0043] The specific steps are as follows: First, the remote sensing image is segmented into connected components to obtain the land cover label for each pixel. Then, a single-channel map is created, and for each land cover entity, all pixels within it are assigned the value of that land cover. Obtain compliance characteristics of land cover attributes. , where H and W are the height and width of the feature map, respectively, and C represents the number of channels in the feature map.

[0044] For each pixel, obtain the feature it belongs to. Check all different features within the 8 neighborhoods. y ,calculate The maximum value is entered into the pixel. If there are no other features in the neighborhood, enter 1 (default compliance). This yields the feature topology compliance features. .

[0045] Assign values ​​to the pixel values ​​of the entire image To obtain the scene rationality features .

[0046] Secondly, the three features mentioned above are concatenated along the channel dimension to achieve feature aggregation, resulting in... :

[0047] Then, The input is fed into the KnowledgeEncoder module for encoding, such as... Figure 2 As shown, the knowledge encoder employs a three-level convolutional and two-level pooling cascaded encoding structure. Specifically, the structure includes three 3×3 convolutional layers (stride=1, padding=1) to extract constrained association features; two 2×2 average pooling layers (stride=2) to compress redundancy and unify dimensionality; a final fusion convolutional layer to achieve fusion; and ReLU activation after convolution to ensure non-negativity, alleviate gradient vanishing, and generate ontology knowledge structured features. .

[0048] formula:

[0049] S105, Construction of Standard Ground Feature Database From standard remote sensing samples, for each type of land cover Cut out the size as Image blocks ( The pixel length of the image block refers to the side length of the image block. This invention takes... Each image patch is then expanded into a one-dimensional feature vector to form a sample feature set. Calculate the mean vector of this feature set (equivalent to the centroid of K-means clustering with 1 class), and then reshape it as follows: The tensor serves as a standard feature library for this type of land cover. Establish a mapping library between land cover categories and standard feature templates. The formula is:

[0050] In the formula, For the first The number of samples of land cover types. This standard land cover feature is used in step four to perform deviation comparison and progressive fitting correction on the local features of abnormal areas.

[0051] Step 2: Remote Sensing Image Processing and Multi-Scale Feature Coding S201, Image Preprocessing and Degradation Restoration The original image is grayscale normalized by mapping the pixel grayscale values ​​to a standard range [0,1], thus generating a normalized preprocessed image. . formula:

[0052] In the formula: The grayscale value of a pixel at any location in the input image; These represent the maximum and minimum global pixel grayscale values ​​of a single input image, respectively, and are adaptively statistically analyzed pixel-by-pixel; Output results These are the normalized pixel values.

[0053] Standardized images Degradation repair is performed using the Retinex algorithm to eliminate interference from clouds, fog, and uneven lighting. Formula:

[0054] In the formula: These are the pixel values ​​of the original degraded image; The estimated values ​​for scene illumination components are obtained by analyzing... The result was obtained by performing Gaussian blurring at three scales (15, 45, and 90 pixels) and then averaging the results. The reflected component is (approximately cloudless and with uniform illumination).

[0055] Then, a bilateral filtering algorithm is used to remove noise while preserving feature edge details, resulting in a standardized image. Formula:

[0056]

[0057]

[0058] In the formula: The filtered pixel output value; These are the normalized weighting coefficients; The spatial distance weights are calculated from the pixel Euclidean distance Gaussian. The pixel grayscale difference weight is calculated from the Gaussian distribution of the pixel grayscale difference. A window with 8 neighboring pixels; The spatial standard deviation is taken as 0.2 in this invention; The grayscale standard deviation is 0.3 in this invention.

[0059] Note: If the input image is not degraded, you can skip this step and proceed directly. .

[0060] S202, Multi-scale Hierarchical Feature Extraction Standardized image restored in step S202 For input, an improved VisionTransformer is used to replace the traditional CNN / Transformer, which solves the problems of limited receptive field and weak global semantics of CNN, as well as the loss of details and computational redundancy of conventional Transformer.

[0061] The improved Vision Transformer hierarchical encoder consists of three cascaded stages, containing only 3 convolutional layers and 6 Transformer Blocks, achieving lightweight multi-scale feature extraction. The first stage uses a 7×7 convolution with a stride of 1 to map the input image to... The feature map is followed by two multi-head self-attention Transformer Blocks to output low-level detail features. ; The second stage uses a 3×3 convolution with a stride of 2 to... downsampling to Then, connect two more Transformer Blocks to output the mid-level structural features. ; In the third stage, a 3×3 convolution with a stride of 2 is used for downsampling to... Then, add two more TransformerBlocks to output the top-level global semantic features. .

[0062] Feature extraction formula:

[0063] In the formula: For the input standardized preprocessed restored image, the underlying fine-grained texture edge features Middle-layer geological features Top-level global semantic features The three-level features complement each other and are not redundant, providing multi-scale features for subsequent adaptive feature fusion.

[0064] The specific module operations and input / output size changes are detailed in the table below:

[0065] S203, Multi-scale Feature Adaptive Fusion Will Downsampling through two layers of 2×2 average pooling to , After one layer of pooling to Then, each channel is boosted to 256 using a 1×1 convolution, resulting in... , , The weights remain unchanged. Adaptive weights are generated using a lightweight fully connected network based on the global statistical features (information entropy, edge density) of the input image. (The sum is 1). Weighted summation yields the global fusion feature. . formula:

[0066] In the formula: It features full-domain integration; This represents the fine-grained texture edge features of the underlying layer after downsampling; The mid-level land cover structure features after downsampling; For adaptive weighting coefficients, satisfying .

[0067] Step 3: Integrate the diffusion generative interpretation of ontology logical constraints.

[0068] S301, Dual-modal feature fusion Input the global image fusion features obtained in step two With the ontology knowledge structured features constructed in step one An adaptive weighted fusion method is employed to achieve bimodal feature complementarity, generating features that combine pixel detail information with geographic logical information. The bimodal feature fusion formula is as follows:

[0069] In the formula: It is a dual-modal fusion feature; The dual-feature balanced weighting coefficient has a value range of [0,1]. For normal images, Use a value of 0.7-0.9; for degraded images, A value of 0.3-0.6 indicates more severe image degradation. The smaller the value, the better. Specifically, the weighting of the two feature types can be adaptively adjusted based on the complexity of the scenario.

[0070] S302, Forward noise addition to diffusion model Dual-modal fusion features output by S301 As input, a multi-level progressive forward noise addition operation of the diffusion model is performed to construct a continuously degraded feature sequence. This invention sets the total number of diffusion steps. Noise Intensity Plan Using linear increments, from arrive .definition Cumulative diffusion coefficient This invention utilizes a dual-modal fusion feature with geographic logical priors. This serves as a noisy base, enabling subsequent reverse denoising iterations to reconstruct the data based on geographic logic, thereby suppressing issues such as pseudo-features, topological distortion, and semantic misrepresentation from the perspective of generation mechanism.

[0071] Forward noise addition formula for diffusion model:

[0072] In the formula: For the first Features after adding noise; It is Gaussian random noise that follows a standard normal distribution.

[0073] S303, Constrained Backward Iterative Denoising In the reverse denoising iteration, in order to calculate the currently generated features Scene rationality First of all A lightweight classification head (1×1 convolution) is used to map a temporary land cover category probability map. The land cover category with the highest probability value in the category probability vector of each pixel output by the classification head is taken as the predicted label for that pixel, thus obtaining a map. The feature category label map is generated, and then the connected component labeling is performed on the label map to obtain the entity set. Then, the formula in step S103 is used to calculate... As a comprehensive constraint coefficient for the entire domain The denoised output features of the model are dynamically weighted and corrected. The parameters of the classification head are jointly trained with the diffusion model.

[0074] Then, the current noisy features and time step Inputting a U-Net-based inverse denoising network yields the predicted denoised mean. . It is a with Tensors of the same size represent the clean features estimated by the network from noisy features. This U-Net (containing downsampling, intermediate attention, upsampling, and skip connections, with a total of 16 convolutional layers) was trained.

[0075] To enable ontology logical constraints to guide the generation process, Embedded denoising iterative formula, for Perform dynamic weighted correction. To avoid directly applying... Multiplying by the predicted mean leads to excessive compression of the feature amplitude. A linear adjustment based on 0.5 is used: Formula:

[0076] In the formula: For the first The previous-order clean feature output after each round of denoising is the optimized feature result after this round of iteration, which is corrected and conforms to geographical logic; among which To adjust the strength coefficient (taken as 0.5), when When the multiplier varies within the range of [0,1], it varies within the range of [0.75,1.25], thus preventing the compliance coefficient from being over-compressed. For the first The adaptive noise variance coefficient of the round iteration, Set the lower limit to The amplitude of the iteration decreases linearly with the number of iterations, controlling the fine-tuning amplitude of the iterations to ensure more stable noise reduction and reduce random jitter in the later stages. It is Gaussian random noise that follows a standard normal distribution.

[0077] S304, Iterative Convergence The reverse denoising iteration of S303 is executed iteratively to gradually restore clear and compliant image features and ground structure. A convergence criterion is set: a steady state is generated by monitoring the global constraint stability of consecutive iterations. Once the model iteration process becomes stable and features no longer exhibit significant logical deviations or detail fluctuations, the iteration is considered converged, and the reverse denoising loop is terminated. After the reverse denoising iteration converges, the converged features are obtained. .

[0078] Diffusion generation iterative convergence criterion formula:

[0079] In the formula: , These are the diffusion generation stages. Step, First The global comprehensive constraint coefficients corresponding to each iteration; When the difference between two consecutive steps is less than the threshold, the generation iteration is considered converged, and the reverse denoising loop is terminated. This proves that the feature structure and ground feature logic have stabilized and there is no further room for correction, thus determining that the generation iteration has converged.

[0080] S305, Decode to generate initial interpreted image The converged features The decoder restores the original image space, resulting in the initial interpreted image. This includes land cover classification results, reconstructed images of incomplete areas, and complete scene completion results. The decoder structure is as follows:

[0081] Step 4: Dynamic correction and error optimization iteration based on ontology logic S401 Error Screening and Compliance Coefficient Calculation The initial interpreted image generated in step three As input, connected component labeling is performed to obtain the entity set. The pre-built ontology logical constraint rule set from step one is then invoked to traverse the entire image, calculating the error for each individual feature. The logical error of each individual feature is then calculated along with the overall global error of the remote sensing plugin, determining the distribution area of ​​the positioning error and the degree of error deviation. Formula:

[0082] In the formula: This is the logical error deviation value for a single feature, used to indicate the degree of anomaly of a single feature. The larger the value, the more severe the deviation in the attributes and topological logic of the single feature. This is the overall error value for the entire remote sensing scene, representing the degree of global geographic logical distortion of the current initial interpreted image; The global comprehensive constraint coefficients are calculated for the iterative process in step three.

[0083] By using compliance coefficients to calculate the difference in reverse, the magnitude of the error can be quantified, enabling the differentiation of error types, the quantification of error severity, and the location of error regions.

[0084] S402, Constraint Weights Adaptive Update For the error-prone areas identified during screening and location, the constraint weights are dynamically and adaptively updated based on the current overall compliance values ​​to achieve differentiated constraints and avoid the loss of details and insufficient correction caused by uniform constraints across the entire domain.

[0085] Constraint weight adaptive adjustment formula:

[0086] In the formula: The updated adaptive constraint weights, i.e. As a correction threshold; The initial preset baseline weight for the system is set to 0.5, which serves as the benchmark for balancing knowledge constraints and data features. This weight can be fine-tuned according to the specific scenario. For adaptive adjustment index and It is used to amplify the constraint strength in low compliance areas and enhance error correction capabilities.

[0087] S403, Local Feature Error Correction for Abnormal areas are cropped from the image. (size Based on the current classification category, retrieve the corresponding feature from the standard land cover feature database. (Same size) The abnormal generation features of the image error area are progressively fitted and compensated, and the defects of local distortion, contour deviation, attribute disorder and semantic inconsistency are gradually corrected, so that the abnormal features of the ground features are continuously fitted to the standard geographical features.

[0088] Local feature correction and compensation formula:

[0089] In the formula: This refers to the standardized local features after correction. Generate bias features for the original model; To correct the step size coefficient and control the correction magnitude in a single iteration.

[0090] S404, Iterative Convergence Determination The system monitors the difference in global comprehensive constraint coefficients after two adjacent iterations in real time, dynamically evaluates the optimization convergence status, and determines that the error optimization has stabilized and there is no further room for optimization when the coefficient difference is less than the preset minimum threshold. The system then adaptively terminates the iteration process to avoid unnecessary iteration time consumption.

[0091] Iterative convergence criterion formula:

[0092] In the formula: , These are the global compliance coefficients for two adjacent iterations; To optimize the convergence threshold during the optimization phase, a value is set. It is used to determine whether the iterative optimization has reached the optimal steady state, with higher accuracy than the generation stage, ensuring that the final result reaches sub-pixel level stability.

[0093] S405, Optimization Result Output After iterative convergence, the optimized remote sensing interpretation image is output. .

[0094] Step 5: Quantitative Evaluation of Interpretation Results S501, Single-Dimensional Quantitative Accuracy Assessment The interpretation results are quantitatively evaluated from four dimensions: pixel reconstruction matching degree, ground object semantic recognition accuracy, spatial topology compliance rate, and scene completeness. This comprehensive evaluation covers the interpretation quality performance at the pixel level, target level, and scene level, achieving a thorough accuracy assessment.

[0095] Pixel reconstruction accuracy formula:

[0096] In the formula: The height and width of the remote sensing image; The remote sensing interpretation image generated in step four; For industry standard true value imaging; This is an indicator function used to count the number of pixel matches.

[0097] Semantic recognition accuracy formula:

[0098] In the formula: The number of samples for correct identification of ground features; This represents the number of misidentified ground features; it is used to quantify the accuracy of the model's semantic classification and recognition of ground features.

[0099] Topology logic compliance rate formula:

[0100] In the formula: The number of compliant and valid topological associations of ground features within the scene; This represents the total number of topological associations of all features within the scene; it is used to quantify the geographic compliance of the overall scene spatial layout.

[0101] Scene completeness formula:

[0102] In the formula: This refers to the area of ​​distorted, incomplete, or missing ground features within the image. The total area effectively covered by the image; used to evaluate the completeness and fidelity of the interpreted scene.

[0103] S502, Calculation of Comprehensive Quality Score Based on single-dimensional quantitative evaluation indicators, a weighted fusion method is used to calculate the comprehensive quality score of the interpretation results. Combined with the weighting standards for remote sensing interpretation industry applications, the weights of each dimension are reasonably allocated to achieve a unified quantitative rating of the quality of the interpretation results and to intuitively represent the overall interpretation performance.

[0104] Comprehensive evaluation score formula:

[0105] In the formula, The overall quality evaluation score for remote sensing interpretation results; In order: pixel precision, semantic precision, topology compliance rate, and scene completeness; The adaptive evaluation weights for the four indicators satisfy the normalization constraint: . The standard deviation of the four indicators is calculated after normalization and is used to represent the degree of consistency fluctuation of multiple indicators. This is the fluctuation penalty coefficient, used to adaptively deduct points to correct interpretation results with unbalanced indicators and large evaluation fluctuations.

[0106] S503, Output of Results Standardized encapsulation and interpretation of images, classification maps, ground feature statistics, and compliance assessment scores are used to output the final results.

[0107] Example 2 like Figure 3 As shown, this embodiment provides a remote sensing image generative interpretation system based on ontology logic constraints, used to execute the method described in Embodiment 1, including: Ontology knowledge construction module: It has a built-in industry standard land feature attribute threshold library and topological relationship set, and supports geographic rule quantitative calculation, ontology knowledge structured coding and standard land feature library management; Image coding module: integrates image preprocessing, degradation restoration and multi-scale Vision Transformer feature extraction units, and outputs global fusion features; The diffusion interpretation module includes a dual-modal fusion unit, a diffusion forward noise addition unit, a constraint-type reverse noise reduction unit, and a decoding unit, which realizes diffusion generative interpretation with geographic logic constraints. Iterative correction module: includes error screening unit, adaptive weight unit, local feature correction unit and convergence determination unit, to complete closed-loop optimization of the interpretation results; Quantitative evaluation module: includes multi-dimensional accuracy calculation unit and comprehensive scoring unit, outputting standardized interpretation results and quality report.

[0108] Compared with the prior art, the present invention has the following positive effects: This invention constructs a hierarchical geographic ontology logic system, significantly improving geographic rationality. It establishes a four-layer ontology system of "entity-attribute-relationship-rule function," creating a hierarchical and quantifiable set of ontology logic constraint rules. This enables the systematic, structured, and algorithmic expression of geographic knowledge, suppressing issues such as conflicting land cover attributes, spatial topological distortions, and semantic mismatches from a generation mechanism perspective, thereby greatly improving the geographic logic compliance of interpretation results for complex remote sensing scenes.

[0109] This invention establishes a knowledge- and data-driven interpretation mechanism to enhance robustness in complex scenarios. It integrates ontology knowledge features with image data features, leveraging geographic ontology knowledge to compensate for feature loss in degraded images while preserving detail accuracy based on original image data. Simultaneously, it implements dynamic logical constraint correction throughout the diffusion iteration process, retaining the advantages of generative models in scene reconstruction and detail restoration, effectively improving model accuracy and generalization ability in complex geographic scenarios such as fog, incompleteness, and low quality.

[0110] This invention designs a closed-loop iterative optimization mechanism to improve the accuracy and stability of the interpretation results. The invention constructs a closed-loop optimization system of "error screening - adaptive weight adjustment - local feature correction - iterative convergence," which can adaptively strengthen the constraint of abnormal areas, progressively correct minor errors, and ensure optimal results through a high-precision convergence mechanism. This effectively reduces the probability of false ground features being generated, eliminates semantic and topological deviations, and significantly improves the consistency, stability, and scene adaptability of the interpretation results.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A generative interpretation method for remote sensing images based on ontology logic constraints, characterized in that, Includes the following steps: Step 1: Construct a set of logical constraint rules and a geographic ontology knowledge system; Based on remote sensing industry standards and prior knowledge of samples, establish a hierarchical geographic ontology knowledge system that includes entity layer, attribute layer, relationship layer and logical rule layer, quantify and generate computable rules for feature attribute constraints, topological relationship constraints and scene rationality constraints, and encode and generate ontology knowledge structured features, while constructing a standard feature library; Step 2: Remote sensing image processing and multi-scale feature encoding; The input remote sensing image is preprocessed and degraded, and an improved hierarchical Vision Transformer encoder is used to extract low-level detail features, mid-level structural features and top-level global semantic features. The global fusion features are obtained by adaptive weighted fusion. Step 3: Diffusion generative interpretation that integrates ontology logical constraints; The global fusion features and ontology knowledge structured features are adaptively weighted and fused to obtain dual-modal fusion features. The dual-modal fusion features are used as the basis to perform forward noise addition using a diffusion model. During the reverse iterative denoising process, global comprehensive constraint coefficients are embedded to dynamically correct the denoising results. After iterative convergence, the initial interpreted image is generated by decoding. Step 4: Dynamic correction and error optimization iteration based on ontology logic; Based on the set of logical constraint rules of geographic ontology, the initial interpreted image is screened for errors across the entire domain to locate abnormal land cover areas. The constraint weights are updated adaptively, and the local feature of abnormal areas is gradually fitted and corrected by combining the standard land cover feature library. After iterative convergence, the optimized interpreted image is output. Step 5: Quantitative evaluation of interpretation results; Single-dimensional quantitative evaluation is carried out from four dimensions: pixel reconstruction accuracy, ground feature semantic recognition accuracy, spatial topology compliance rate and scene completeness, and a weighted comprehensive quality score is calculated to output the final interpretation results.

2. The remote sensing image generative interpretation method based on ontology logic constraints according to claim 1, characterized in that, Step 1 specifically includes: S101. Land Feature Attribute Threshold Quantification: For seven types of land features, namely water area, forest land, building, road, cultivated land, bare land and mountain, determine the upper and lower limits of attribute thresholds for gray scale, texture, size and elevation, construct a land feature attribute threshold library, and determine the compliance of single land feature attributes through the binary constraint formula of land feature attributes. S102. Construction of Topological Relationship Set: Define five types of topological relationships: isolation, adjacency, inclusion, connectivity, and mutual exclusion. Establish a standard topological relationship set for any two types of land features. Determine the spatial topological compliance of two land features through the land feature topological binary constraint formula. S103, Geographical rule quantification and transformation: Calculate the comprehensive rationality coefficient of a single feature and the comprehensive rationality of a scene respectively, and construct a two-level logical rule constraint system for individual features and global scenes; S104. Geographic ontology knowledge structured feature generation: Generate feature attribute compliance features, feature topology compliance features, and scene rationality features. After channel splicing, input them into the knowledge encoder for convolution and pooling encoding, and output ontology knowledge structured features. S105. Construction of Standard Land Feature Library: Calculate the mean feature of standard sample image blocks for each type of land feature, and construct a standard land feature library that maps land feature categories one-to-one with standard feature templates.

3. The remote sensing image generative interpretation method based on ontology logic constraints according to claim 1, characterized in that, Step 2 specifically includes: S201. Image preprocessing and degradation restoration: The original image is grayscale normalized and mapped to the [0,1] interval. The multi-scale Retinex algorithm is used to eliminate interference from clouds and uneven lighting. Then, bilateral filtering is used to remove noise and preserve edges to obtain the normalized image. S202, Multi-scale hierarchical feature extraction: A three-level cascaded improved Vision Transformer hierarchical encoder is adopted, which sequentially passes through convolutional downsampling and window multi-head self-attention module to output low-level detailed features, mid-level structural features and top-level global semantic features; S203, Multi-scale Feature Adaptive Fusion: The bottom and middle layer features are downsampled to the same size as the top layer features, and after being convolved with a 1×1 to unify the number of channels, adaptive weights are generated based on the global statistical features of the image, and weighted fusion is performed to obtain the global fusion features.

4. The remote sensing image generative interpretation method based on ontology logic constraints according to claim 1, characterized in that, In step 3, the dual-modal feature fusion formula is as follows: In the formula, For dual-modal fusion features, F all For the feature of full-domain integration, F know As a structured feature of ontology knowledge, The dual-feature balanced weighting coefficient has a value range of [0,1]. For normal images, Use a value of 0.7-0.9; for degraded images, A value of 0.3-0.6 indicates more severe image degradation. The smaller; The diffusion model has a total of 1000 forward noise steps, and the noise intensity is... from The value is linearly increased to 0.02, and a continuous degradation feature sequence is constructed using the dual-modal fusion feature as the noise-adding basis.

5. The method for generative interpretation of remote sensing images based on ontology logic constraints according to claim 1, characterized in that, In step 3, the reverse iterative denoising process specifically includes: The noisy features of the current step are used to generate a temporary land cover category label map through a lightweight classification head. After connected component labeling, the overall scene rationality is calculated as the overall constraint coefficient C for the entire domain. total ; The noisy features and time steps are input into the U-Net inverse denoising network to obtain the predicted denoised mean. The predicted mean is then dynamically weighted and corrected using a denoising iterative formula that embeds global comprehensive constraint coefficients. The formula is as follows: ； In the formula, For the first The previous clean feature output after each iteration of denoising is the optimized feature result after this iteration, which is corrected and conforms to the geographical logic. To predict the denoised mean, To adjust the strength coefficient, a value of 0.5 is used. The noise variance coefficient is adaptive and decays with the number of iterations. The noise is Gaussian random noise that follows a standard normal distribution. When the difference between the global comprehensive constraint coefficients of two consecutive steps is less than 10 -3 When the iteration converges, the initial decoded image is obtained.

6. The method for generative interpretation of remote sensing images based on ontology logic constraints according to claim 1, characterized in that, In step 4, the error screening and weight update are specifically as follows: Connectivity components are marked on the initial interpreted image, and the logical error of each individual feature and the overall error of the entire region are calculated to locate anomalous feature areas. Combining global compliance with adaptive constraint weight updates, the formula is: ； In the formula, These are the updated adaptive constraint weights; The initial preset baseline weight for the system is set to 0.5; For adaptive adjustment index and .

7. The remote sensing image generative interpretation method based on ontology logic constraints according to claim 6, characterized in that, In step 4, the local feature correction and iterative convergence are specifically as follows: For abnormal regions where the logical error of a single feature exceeds the constraint weight, the corresponding local feature blocks are cropped, and progressive fitting compensation is performed using a standard feature library. The correction formula is as follows: ； In the formula, This refers to the standardized local features after correction. Generate bias features for the original model; These are the standard features corresponding to the land cover categories; To correct the step size coefficient and control the correction magnitude in a single iteration; When the difference between the global comprehensive constraint coefficients of two adjacent iterations is less than 10 -5 When the optimization converges, the optimized interpretation image is output.

8. The method for generative interpretation of remote sensing images based on ontology logic constraints according to claim 1, characterized in that, Step 5, the single-dimensional quantitative evaluation specifically includes: Pixel reconstruction accuracy is the pixel matching ratio between the interpreted image and the ground truth image; The accuracy of geographic feature semantic recognition is the precision rate, which is the proportion of correctly identified samples out of the total number of identified samples; The spatial topology compliance rate is the proportion of compliant topology association pairs to the total number of topology association pairs; Scene integrity is 1 minus the ratio of the area of ​​the incomplete area to the total effective area.

9. The remote sensing image generative interpretation method based on ontology logic constraints according to claim 8, characterized in that, The overall quality score is calculated using a weighted product combined with a fluctuation penalty, and the formula is as follows: ； In the formula, The overall quality evaluation score for remote sensing interpretation results; In order: pixel precision, semantic precision, topology compliance rate, and scene completeness; The adaptive evaluation weights for the four indicators satisfy the normalization constraint: ; The standard deviation of the four indicators is calculated after normalization and is used to represent the degree of consistency fluctuation of multiple indicators. This is the fluctuation penalty coefficient, used to adaptively deduct points to correct interpretation results with unbalanced indicators and large evaluation fluctuations.

10. A remote sensing image generative interpretation system based on ontology logic constraints for implementing the method of any one of claims 1-9, characterized in that, include: The ontology knowledge construction module is used to construct a set of logical constraint rules and a geographic ontology knowledge system, quantify the constraints of land feature attributes, topology and scene rationality, and generate ontology knowledge structured features and a standard land feature feature library. The image coding module is used to preprocess and restore the input remote sensing images, extract multi-scale image features and adaptively fuse them to obtain global fusion features; The diffusion interpretation module is used to fuse global fusion features and ontology knowledge structured features, perform forward noise addition and reverse iterative noise reduction of the diffusion model with ontology logic constraints, and generate the initial interpreted image; The iterative correction module is used to perform global error screening based on ontology logic rules, adaptively update constraint weights, perform local feature progressive fitting correction on abnormal regions, and output optimized interpretation images. The quantitative evaluation module is used to evaluate accuracy from four dimensions: pixels, semantics, topology, and scene, and calculate a comprehensive quality score to output the final interpretation results.

Images