A multi-source remote sensing data retrieval enhancement generation method, device, equipment and medium

Abstract

The present application relates to the technical field of data processing, and provides a multi-source remote sensing data retrieval enhancement generation method, device, equipment and medium, the method comprises: using a spatial adapter to analyze and process the multi-source remote sensing data obtained in advance to obtain a fusion data set; constructing a spatial observation hypergraph based on the fusion data set, and mapping the spatial observation hypergraph to a hyperbolic space according to a coupling relationship to obtain a hyperbolic hypergraph; in response to a query index given by a user, extracting relevant hyperedges in the hyperbolic hypergraph, and respectively performing cross-resolution aggregation on the corresponding relevant hyperedges based on the coupling relationship to obtain corresponding multi-source evidence; performing conflict triage on each multi-source evidence, calibrating the corresponding multi-source evidence according to the triage result to obtain calibrated enhanced evidence; and inputting the multi-source evidence and the enhanced evidence as retrieval enhancement context information into a large language model. The present application effectively realizes efficient fusion, accurate retrieval and credible generation of multi-source remote sensing data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method, apparatus, device, and medium for enhancing the generation of multi-source remote sensing data retrieval. Background Technology

[0002] With the development of space exploration technology, various observation platforms (such as orbital remote sensing payloads and in-situ exploration instruments) have accumulated a large amount of multi-source Mars remote sensing data. These data vary significantly in terms of resolution, spectral coverage, observation geometry, and acquisition time. How to effectively fuse and intelligently retrieve this data has become a core challenge in the field of space data science.

[0003] In recent years, Retrieval-Augmented Generation (RAG) technology has effectively mitigated the illusion problem of large language models by injecting external knowledge base retrieval results as contextual information into them. This information serves as prior knowledge for question answering within the large language model, allowing it to generate answers based on this contextual information during subsequent user question-answering processes. Multi-source RAG methods further address multi-source data sources by constructing multi-source knowledge graphs and employing multi-level confidence calculations, serving fields such as knowledge-based question answering and reasoning.

[0004] However, existing multi-source RAG methods have two shortcomings: First, existing methods construct knowledge graphs based on discrete text entities and binary triples, which is insufficient for representing the inherent continuous spatiotemporal topological relationships and heterogeneous reference frames of Mars remote sensing data, easily leading to edge explosion and distortion at the scale level; Second, the core assumption of existing methods is that inconsistencies between source data usually stem from misinformation or model illusions, directly treating all inter-source conflicts as noise and filtering them uniformly. However, this assumption does not hold true in deep space exploration scenarios, because different instruments produce different observation results on the same target due to differences in physical principles, which often contain scientific value. Blindly filtering will lose these valuable observational discrepancies. Summary of the Invention

[0005] The present invention aims to solve at least one of the above-mentioned technical problems.

[0006] To address the aforementioned problems, this invention provides a method, apparatus, device, and medium for enhancing the generation of multi-source remote sensing data retrieval.

[0007] In a first aspect, the present invention provides a method for enhancing and generating multi-source remote sensing data retrieval, comprising: A spatial adapter is used to analyze and process the pre-acquired multi-source remote sensing data to obtain a fused dataset; A spatial observation hypergraph is constructed based on the fused dataset, and the spatial observation hypergraph is mapped to hyperbolic space according to the coupling relationship to obtain a hyperbolic hypergraph. The coupling relationship includes coupling the radial depth of the hyperbolic space with the spatial resolution. In response to the query index given by the user, relevant hyperedges in the hyperbolic hypergraph are extracted, and cross-resolution aggregation is performed on the corresponding relevant hyperedges based on the coupling relationship to obtain the corresponding multi-source evidence; Each piece of multi-source evidence is subjected to conflict triage, and the corresponding multi-source evidence is calibrated according to the triage results to obtain calibrated enhanced evidence. The multi-source evidence and the enhanced evidence are used as retrieval enhancement context information input into the large language model.

[0008] Optionally, in response to a user-provided query index, relevant hyperedges in the hyperbolic hypergraph are extracted, and cross-resolution aggregation is performed on the corresponding relevant hyperedges based on the coupling relationship to obtain corresponding multi-source evidence, including: In response to the query index, spatial intent elements and topic entities are extracted from the query index; Based on the subject entity, retrieve the hyperedges that satisfy the spatial intent in the hyperbolic hypergraph to obtain the relevant hyperedges; Spatiotemporal encoding is performed on each of the related hyperedges to obtain a spatiotemporal encoded vector. A preset first classifier is used to score the relevance of each of the related hyperedges to obtain a corresponding score. Based on the spatiotemporal coding vector, the score, and the coupling relationship, cross-resolution aggregation is performed on the corresponding relevant hyperedges to obtain the corresponding multi-source evidence.

[0009] Optionally, the step of performing spatiotemporal encoding on each of the related hyperedges to obtain a spatiotemporal encoded vector includes: Based on each of the relevant hyperedges, a corresponding pseudo binary triplet is generated, consisting of the topic entity, the relevant hyperedge, and the associated entity, wherein the associated entity includes ground object entities or geological object entities that have spatial or observational association with the topic entity; Based on the pseudo-binary triples, semantic embedding, structural nearest neighbor encoding, and hyperbolic space encoding are fused to construct the spatiotemporal encoding vector for each relevant hyperedge.

[0010] Optionally, the step of performing cross-resolution aggregation based on the spatiotemporal coding vector, the score, and the coupling relationship as corresponding relevant hyperedges to obtain corresponding multi-source evidence includes: Resolution-aware radial weighting is determined based on the relationship between the radial depth and the spatial resolution, wherein the radial depth is proportional to the spatial resolution; For each of the relevant hyperedges, the product of the score and the resolution-perceived radial weighting is used as the comprehensive weight. Based on the spatiotemporal coding vector, the spatial outward Einstein midpoint projection vector projected onto the hyperbolic space is calculated to obtain the corresponding multi-source evidence.

[0011] Optionally, the conflict triage of each of the multi-source pieces of evidence includes: For any two of the multi-source evidences, input the two multi-source evidences and the query index into the large language model respectively to obtain the corresponding probability distribution entropy; Based on the probability distribution entropy, the cross-source interaction entropy of the two multi-source pieces of evidence is obtained, and the two multi-source pieces of evidence whose cross-source interaction entropy exceeds a preset threshold are determined to be in conflict.

[0012] Optionally, the step of calibrating the multi-source evidence according to the triage results to obtain calibrated enhanced evidence includes: In response to the triage result indicating a conflict, the fused feature vector of the conflicting multi-source evidence is input into a preset second classifier to obtain the conflict type; The corresponding score is used as the base confidence level, and the base confidence level is recalibrated according to the conflict type to obtain the calibrated enhanced evidence.

[0013] Optionally, the step of using a spatial adapter to parse and process the pre-acquired multi-source remote sensing data to obtain a fused dataset includes: Match the corresponding spatial adapter according to the type of the multi-source remote sensing data; The spatial adapter is used to perform metadata parsing, spatial coverage extraction, time window unification, and observation parameter normalization on the matched multi-source remote sensing data to obtain a fused dataset under a globally unified spatiotemporal coordinate system.

[0014] Secondly, the present invention provides a multi-source remote sensing data retrieval enhancement generation device, comprising: The parsing module is used to parse and process pre-acquired multi-source remote sensing data using a spatial adapter to obtain a fused dataset; The mapping module is used to construct a spatial observation hypergraph based on the fused dataset, and map the spatial observation hypergraph to hyperbolic space according to the coupling relationship to obtain a hyperbolic hypergraph, wherein the coupling relationship includes coupling the radial depth of the hyperbolic space with the spatial resolution. The aggregation module is used to extract relevant hyperedges in the hyperbolic hypergraph in response to a query index given by the user, and perform cross-resolution aggregation on the corresponding relevant hyperedges based on the coupling relationship to obtain the corresponding multi-source evidence; The calibration module is used to perform conflict triage on each of the multi-source evidences, calibrate the corresponding multi-source evidences according to the triage results, and obtain calibrated enhanced evidence. The input module is used to input the multi-source evidence and the enhanced evidence as retrieval enhanced context information into the large language model.

[0015] Thirdly, the present invention provides an electronic device, including a memory and a processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the multi-source remote sensing data retrieval enhancement generation method as described in the first aspect.

[0016] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multi-source remote sensing data retrieval enhancement generation method as described in the first aspect.

[0017] The beneficial effects of the multi-source remote sensing data retrieval enhancement generation method of this invention are as follows: It employs a spatial adapter to analyze and process multi-source remote sensing data. By performing standardized analysis on heterogeneous multi-source remote sensing data through the spatial adapter, it can unify the spatial coverage, time windows, and observation parameters of data from different observation platforms and payload types. This solves the problem of heterogeneous differences in spatiotemporal, spectral, and geometric dimensions of multi-source remote sensing data, achieving global normalization and efficient fusion of multi-source data, and providing a standardized and unified data foundation for subsequent hypermap construction. A spatial observation hypermap is constructed based on the fused dataset, and the radial depth and spatial resolution are coupled and mapped to hyperbolic space to obtain a hyperbolic hypermap. Using a spatial observation hypermap to replace the traditional binary map structure allows for the aggregation of multi-source observation entities in the same region with a single hyperedge, fundamentally avoiding the edge explosion problem caused by binary maps. Simultaneously, coupling and mapping the radial depth and spatial resolution of hyperbolic space can faithfully represent the inherent continuous spatiotemporal topology and scale hierarchy characteristics of remote sensing data, solving the technical problems of scale hierarchy distortion and inability to adapt to heterogeneous reference frames in existing methods, and achieving high-fidelity representation of multi-scale spatial structures. Multi-source evidence is obtained by extracting relevant hyperedges from the query index and performing cross-resolution aggregation. Based on the hyperbolic hypergraph and the radial depth-resolution coupling relationship, relevant hyperedges matching the query intent can be accurately retrieved, and multi-resolution evidence can be effectively aggregated. While preserving high-resolution detail information, multi-scale evidence fusion is completed, improving the completeness and spatial consistency of the retrieval results, making the retrieval output more consistent with the spatial distribution patterns of remote sensing data and the actual query needs of users. Conflict triage and calibration are performed on multi-source evidence to obtain enhanced evidence. Conflict triage identifies and classifies inconsistencies among multi-source evidence, abandoning the erroneous assumption of treating all conflicts as noise in existing technologies. It can distinguish between noise conflicts and scientifically valuable instrument-inherent conflicts, scale-dependent conflicts, and temporal evolution conflicts, and performs differentiated confidence calibration. While eliminating invalid noise, it retains scientifically significant observational discrepancies in deep space exploration, avoids the loss of valuable information, and significantly improves the scientific credibility and reliability of the evidence. By using multi-source data and augmented data as context input to the large language model to complete the retrieval enhancement generation, and injecting calibrated high-credibility evidence as retrieval enhancement context into the large language model, the model can be provided with real, accurate and traceable external knowledge support, which can significantly alleviate the illusion problem of large language models in the field of remote sensing, improve the accuracy, spatial consistency and scientific rigor of answer generation, and make the model output more suitable for deep space exploration and remote sensing scientific analysis scenarios.

[0018] This invention utilizes a space adapter to achieve standardized integration of multi-source heterogeneous remote sensing data. It accurately expresses the inherent spatiotemporal topological relationships of remote sensing data through a hyperbolic hypergraph structure, avoiding edge explosion and scale distortion. Simultaneously, cross-resolution aggregation ensures the completeness and spatial consistency of retrieval results. To address multi-source observation conflicts, a physically informed conflict triage mechanism is employed to distinguish between noise and scientifically significant observational differences and to perform confidence calibration, preserving the scientific value of the data while eliminating invalid interference. High-quality, calibrated evidence, used as enhanced context input to the large language model, significantly reduces model illusions and improves the accuracy and scientific credibility of the answers. This invention effectively solves the problems of insufficient representational ability, scale distortion, false conflict filtering, and unreliable generated results in existing multi-source RAG technology for remote sensing data processing. It achieves efficient fusion, accurate retrieval, and reliable generation of multi-source remote sensing data, better adapting to the intelligent question-answering and scientific analysis needs of deep space exploration remote sensing data such as Mars, and possesses strong practical value and scientific significance. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the multi-source remote sensing data retrieval enhancement generation method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the multi-source remote sensing data retrieval enhancement generation device according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0021] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0022] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0023] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0024] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0025] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for enhancing and generating multi-source remote sensing data retrieval, comprising: Step S1: Use a spatial adapter to analyze and process the pre-acquired multi-source remote sensing data to obtain a fused dataset.

[0026] Specifically, multi-source remote sensing data originates from different platforms such as orbital payloads, in-situ instruments, and derivative products, exhibiting heterogeneous differences in spatial coverage, temporal reference, and observation parameters, making it unsuitable for direct use in hypermap construction and retrieval. Therefore, standardization processing is necessary. Multi-source Mars remote sensing data includes spatial data from different observation platforms, specifically orbital remote sensing data, in-situ exploration data, and derivative data products. Orbital remote sensing data can include optical imagery, hyperspectral data, and topographic data, while in-situ exploration data can include radar data and rover data. By employing a space adapter to perform normalization processing on the multi-source remote sensing data, a fused dataset containing unified spatial, temporal, and parameter remote sensing observation information is obtained, providing a high-quality, unified data foundation for subsequent hypermap construction.

[0027] Step S2: Construct a spatial observation hypergraph based on the fused dataset, and map the spatial observation hypergraph to hyperbolic space according to the coupling relationship to obtain a hyperbolic hypergraph, wherein the coupling relationship includes coupling the radial depth of the hyperbolic space with the spatial resolution relationship.

[0028] Specifically, the hyperedge f is defined based on the fused dataset. nspa For example, multiple source observation entities in the same region can be aggregated into a single hyperedge, containing a tuple of parameters such as instrument entity, spatial coverage, temporal acquisition window, spectral band set, target geological feature set, and ground sampling distance, achieving a high degree of aggregation of multi-source observations in the same location. In traditional paired spatial maps, Each coexisting space entity requires O(k) 2 There are 10 near-neighbor edges. However, with a hyperedge structure, a single multi-factor fact can bind all k entities, reducing the edge complexity from O(k^k) to O(k^k). 2 The edge explosion problem is effectively solved by reducing the edge size to O(k). Based on the above hyperedges, a spatial observation hypergraph is constructed. Using scale-aware Lorentz embedding, the spatial observation hypergraph is represented in hyperbolic space, where the curvature constant K is a constant negative curvature, i.e., K < 0. The coupling relationship is expressed as: , Where r() represents the radial depth function, Φ(f n spa ) represents the Lorentz embedding vector of the spatial observation hyperedge. The scale normalization coefficient is represented by cosh(), which represents the hyperbolic cosine function, and g( res ) represents the mapping function from spatial resolution to geometric distance. res Indicates spatial resolution. As resolution increases... res Reduce (more refined scale), N ( () refers to the resolution level The number of observations at a given location exhibits a quadratic growth, demonstrating the exponential branching characteristics of a negative curvature space. Therefore, the spatial scale hierarchy is inherently hyperbolic, and a polynomial volume-growing Euclidean embedding cannot faithfully represent it. Through the aforementioned embedding, global coarse-resolution data is placed near the hyperbolic origin (small radial depth), while local high-resolution data is placed far from the origin (large radial depth).

[0029] Step S3: In response to the query index given by the user, extract the relevant hyperedges in the hyperbolic hypergraph, and perform cross-resolution aggregation on the corresponding relevant hyperedges based on the coupling relationship to obtain the corresponding multi-source evidence.

[0030] Specifically, the system receives the user-input query index and extracts its content using, for example, a semantic parsing model. It matches query elements, retrieves relevant hyperedges in the hyperbolic hypergraph, and encodes each hyperedge semantically, structurally, and spatially. Each relevant hyperedge is then scored, for example, by calculating the similarity between the query index and relevant hyperedges. Weighted aggregation is performed on the multi-resolution hyperedges, with weights set based on coupling relationships or scoring, prioritizing high-resolution observations to preserve fine-scale information and avoid coarse-resolution smoothing. After aggregation, spatiotemporally consistent, resolution-fused multi-source spatial evidence is obtained. This accurately acquires multi-source observational evidence matching the query, achieving cross-resolution fusion while preserving high-resolution details, making the evidence more aligned with query requirements and spatial distribution patterns.

[0031] Step S4: Perform conflict triage on each of the multi-source evidences, and calibrate the corresponding multi-source evidences according to the triage results to obtain calibrated enhanced evidence.

[0032] Specifically, the cross-source interaction entropy between any two data sources is calculated, and this entropy is used to determine whether a conflict exists. If the triage result indicates a conflict, the feature vectors corresponding to the multi-source evidence are adaptively calibrated based on the conflict. For example, if there is a temporal evolution conflict, i.e., the actual change in the target due to different observation times, a time decay weighting is applied accordingly, prioritizing the retention of recent observations and preserving temporal change signals. After calibration, highly reliable, conflict-marked, and scientifically interpretable enhanced evidence is obtained.

[0033] Step S5: Input the multi-source evidence and the enhanced evidence as retrieval enhanced context information into the large language model.

[0034] Specifically, the system integrates multi-source data and augmented data, combines them with user queries to retrieve enhanced contextual information, and inputs this information into a large language model. The model then generates answers based on real remote sensing evidence. When using the system, users can directly input this large language model to obtain scientifically credible, traceable, and conflict-inclusive enhanced search answers.

[0035] This embodiment employs a spatial adapter for parsing and processing multi-source remote sensing data. By performing standardized parsing on heterogeneous multi-source remote sensing data through the spatial adapter, it can unify the spatial coverage, time windows, and observation parameters of data from different observation platforms and payload types. This addresses the heterogeneous differences in spatiotemporal, spectral, and geometric dimensions of multi-source remote sensing data, achieving global normalization and efficient fusion of multi-source data. This provides a standardized and unified data foundation for subsequent hypermap construction. A spatial observation hypermap is constructed based on the fused dataset, and the radial depth and spatial resolution are coupled and mapped to hyperbolic space to obtain a hyperbolic hypermap. Replacing the traditional binary map structure with a spatial observation hypermap allows for the aggregation of multi-source observation entities in the same region with a single hyperedge, fundamentally avoiding the edge explosion problem caused by binary maps. Simultaneously, coupling and mapping the radial depth and spatial resolution of hyperbolic space can faithfully represent the inherent continuous spatiotemporal topology and scale hierarchy characteristics of remote sensing data, solving the technical problems of scale hierarchy distortion and inability to adapt to heterogeneous reference frames in existing methods, and achieving high-fidelity representation of multi-scale spatial structures. Multi-source evidence is obtained by extracting relevant hyperedges from the query index and performing cross-resolution aggregation. Based on the hyperbolic hypergraph and the radial depth-resolution coupling relationship, relevant hyperedges matching the query intent can be accurately retrieved, and multi-resolution evidence can be effectively aggregated. While preserving high-resolution detail information, multi-scale evidence fusion is completed, improving the completeness and spatial consistency of the retrieval results, making the retrieval output more consistent with the spatial distribution patterns of remote sensing data and the actual query needs of users. Conflict triage and calibration are performed on multi-source evidence to obtain enhanced evidence. Conflict triage identifies and classifies inconsistencies among multi-source evidence, abandoning the erroneous assumption of treating all conflicts as noise in existing technologies. It can distinguish between noise conflicts and scientifically valuable instrument-inherent conflicts, scale-dependent conflicts, and temporal evolution conflicts, and performs differentiated confidence calibration. While eliminating invalid noise, it retains scientifically significant observational discrepancies in deep space exploration, avoids the loss of valuable information, and significantly improves the scientific credibility and reliability of the evidence. By using multi-source data and augmented data as context input to the large language model to complete the retrieval enhancement generation, and injecting calibrated high-credibility evidence as retrieval enhancement context into the large language model, the model can be provided with real, accurate and traceable external knowledge support, which can significantly alleviate the illusion problem of large language models in the field of remote sensing, improve the accuracy, spatial consistency and scientific rigor of answer generation, and make the model output more suitable for deep space exploration and remote sensing scientific analysis scenarios.

[0036] This invention utilizes a space adapter to achieve standardized integration of multi-source heterogeneous remote sensing data. It accurately expresses the inherent spatiotemporal topological relationships of remote sensing data through a hyperbolic hypergraph structure, avoiding edge explosion and scale distortion. Simultaneously, cross-resolution aggregation ensures the completeness and spatial consistency of retrieval results. To address multi-source observation conflicts, a physically informed conflict triage mechanism is employed to distinguish between noise and scientifically significant observational differences and to perform confidence calibration, preserving the scientific value of the data while eliminating invalid interference. High-quality, calibrated evidence, used as enhanced context input to the large language model, significantly reduces model illusions and improves the accuracy and scientific credibility of the answers. This invention effectively solves the problems of insufficient representational ability, scale distortion, false conflict filtering, and unreliable generated results in existing multi-source RAG technology for remote sensing data processing. It achieves efficient fusion, accurate retrieval, and reliable generation of multi-source remote sensing data, better adapting to the intelligent question-answering and scientific analysis needs of deep space exploration remote sensing data such as Mars, and possesses strong practical value and scientific significance.

[0037] Optionally, in response to a user-provided query index, relevant hyperedges in the hyperbolic hypergraph are extracted, and cross-resolution aggregation is performed on the corresponding relevant hyperedges based on the coupling relationship to obtain corresponding multi-source evidence, including: In response to the query index, spatial intent elements and topic entities are extracted from the query index.

[0038] Specifically, after receiving the query index input by the user, the system uses a semantic parsing and spatial feature extraction model to separate spatial intent elements and subject entities from the query. Spatial intent elements include spatial coverage, observation time constraints, resolution preferences, target geological types, etc., while subject entities are core objects such as Martian features, geological units, and exploration areas that the query points to.

[0039] Based on the subject entity, retrieve the hyperedges that satisfy the spatial intent in the hyperbolic hypergraph to obtain the relevant hyperedges.

[0040] Specifically, taking the subject entity as the center, the spatial observation hyperedges associated with the entity are traversed in the hyperbolic hypergraph, and filtered according to the spatial intent elements. Hyperedges that match the spatial range, time window, and resolution are retained as relevant hyperedges related to the query.

[0041] Spatiotemporal encoding is performed on each of the aforementioned related hyperedges to obtain a spatiotemporal encoded vector, including: Based on each of the relevant hyperedges, a corresponding pseudo-binary triplet is generated, consisting of the topic entity, the relevant hyperedge, and the associated entity, wherein the associated entity includes ground object entities or geological object entities that have spatial or observational association with the topic entity.

[0042] Specifically, the relevant hyperedges themselves are high-dimensional structured data and cannot be directly input into a classifier for relevance scoring; furthermore, relying solely on semantic information cannot accurately characterize the spatial location, scale level, and topological relationship of the hyperedges in the hyperbolic hypergraph, leading to distorted relevance judgments. Therefore, taking the topic entity e... h With core and related super-edge f n spa For the relationship and related entity e t For the object, construct a standard pseudo-binary triple (e h f n spa e t The associated entity is a ground object or geological object that has a spatial proximity, inclusion, intersection, or joint observation relationship with the subject entity. This triple transforms the multi-dimensional structure of the hyperedge into a computable reasoning unit, which is adapted to the subsequent coding and classification process.

[0043] Based on the pseudo-binary triples, semantic embedding, structural nearest neighbor encoding, and hyperbolic space encoding are fused to construct the spatiotemporal encoding vector for each relevant hyperedge.

[0044] Specifically, a text embedding model is used to encode the query statement, topic entity, related hyperedge, and associated entity into semantic embedding vectors, which are used to represent the textual semantic information of each element in the triple and capture the semantic association between the query and the hyperedge. Based on the topological connectivity of the hyperbolic hypergraph, the proximity, connectivity strength, and hierarchical relationship between entities and hyperedges in the triple are calculated to obtain structural nearest neighbor encoding, which is used to represent the degree of structural association of hyperedges in the graph. The spatial features of corresponding elements in the triple are calculated in hyperbolic space to obtain hyperbolic space encoding, such as hyperbolic geodesic distance, which is used to represent the degree of spatial proximity; radial depth difference, which is used to represent the resolution scale difference between different hyperedges; and azimuth cosine, which is used to represent the spatial relative direction relationship. The semantic embedding vector, structural nearest neighbor encoding vector, and hyperbolic space encoding vector are concatenated and fused according to their dimensions to form a high-dimensional spatiotemporal encoding vector that simultaneously contains semantic information, structural information, and spatial information, which is expressed by the formula: X=[φ(q)||φ(e h )||φ(f n spa )||δ(e h f n spa e t )||ψ geo (e) h e t )], ψ geo (e) h e t )=[dK (φ(e) h ), φ (e t ), △r (e h e t ), cosθ bearing )], Where φ() represents the semantic embedding function, q represents the user query index, || represents the concatenation operation, δ() represents the structural nearest neighbor encoding, and ψ geo ( ) represents hyperbolic space encoding, d K ( , ) represents the hyperbolic geodesic distance function, Δr( , ) represents the radial depth difference, and θ bearing Cosθ represents the azimuth angle. bearing It represents the cosine of the azimuth angle.

[0045] A pre-defined first classifier is used to score the relevance of each related hyperedge, resulting in a score that reflects the semantic and spatial matching degree between the hyperedge and the user query. Specifically, the spatiotemporal encoded vector is input into the pre-defined first classifier to calculate the matching degree between each related hyperedge and the query index, outputting a relevance score between 0 and 1. The pre-defined first classifier is either pre-trained or a similarity algorithm carrier.

[0046] Based on the spatiotemporal coding vector, the score, and the coupling relationship, cross-resolution aggregation is performed on the corresponding relevant hyperedges to obtain the corresponding multi-source evidence, including: Resolution-aware radial weighting is determined based on the relationship between the radial depth and the spatial resolution, wherein the radial depth is proportional to the spatial resolution; For each of the relevant hyperedges, the product of the score and the resolution-perceived radial weighting is used as the comprehensive weight. Based on the spatiotemporal coding vector, the spatial outward Einstein midpoint projection vector projected onto the hyperbolic space is calculated to obtain the corresponding multi-source evidence.

[0047] Specifically, remote sensing data at different resolutions exhibit significant differences in information density and level of detail. Low-resolution data cannot provide sufficient detailed information, while high-resolution data better reflects the true geological features. Furthermore, user queries often implicitly indicate a preference for high-resolution information. Existing methods employ equal-weight aggregation of data at different resolutions, which can easily lead to the dilution of high-resolution details by low-resolution information, failing to reflect the advantages of high-resolution data. To aggregate this evidence into a unified representation without losing fine-scale information, cross-resolution aggregation of multi-source evidence is proposed.

[0048] Based on a predefined coupling relationship between radial depth and spatial resolution, where radial depth is directly proportional to spatial resolution (higher spatial resolution, smaller ground sampling distance), the corresponding hyperbolic radial depth is greater; conversely, lower resolution results in a smaller radial depth. For each relevant hyperedge, its radial depth in hyperbolic space is calculated using a predefined mapping function based on its spatial resolution parameters. Resolution-aware radial weighting is determined based on radial depth, employing a monotonically increasing function of radial depth (such as an exponential or linear scaling function) to ensure that hyperedges with greater radial depth (i.e., higher resolution) receive larger radial weighting values, guaranteeing a higher weighting proportion for high-resolution data in aggregation. The relevance score between the relevant hyperedges output by the first classifier and the query is taken, and multiplied by the resolution-aware radial weighting to obtain the comprehensive weight. This comprehensive weight considers both the hyperedge's relevance to the query and its resolution information; hyperedges with higher relevance and higher resolution have a larger comprehensive weight, while hyperedges with low relevance or low resolution have their comprehensive weight suppressed. The spatiotemporal encoding vector of each relevant hyperedge is taken. This vector contains multi-dimensional information such as semantics, structure, and hyperbolic space, and has been mapped to a hyperbolic space with constant negative curvature. Based on the hyperbolic space, the weighted sum of all relevant hyperedges is calculated, that is, the spatiotemporal encoding vector of each hyperedge is multiplied by its corresponding comprehensive weight, and then all weighted vectors are summed to obtain the weighted vector sum; at the same time, the sum of all weights is calculated as a normalization factor. In the hyperbolic space, the calculation of the Einstein midpoint needs to consider the Lorentz inner product constraint. By dividing the weighted vector sum by the weight sum, and then using the reprojection operator of the Lorentz model, the result is mapped back to a legal hyperbolic manifold, ensuring that the calculation result satisfies the geometric constraints of the hyperbolic space. During the reprojection process, the calculated vector needs to be normalized to satisfy the inner product condition of the Lorentz model, ensuring that the result lies on the manifold of the hyperbolic space, rather than an outlier outside the manifold. The specific process is expressed as follows: , in, This represents the vector reprojected from the spatial outward Einstein midpoint projection vector, i.e., multi-source evidence, where p represents the projection parameter, and Π represents the projection parameter. K ( ) denotes the hyperbolic reprojection operator, n denotes the number of related hyperedges, and w i Indicates the score, φ res (f) i ) represents resolution-perceptual radial weighting, λ i Denotes the Lorentz factor, Φ(f) i ) represents the hyperbolic embedding vector of the hyperedge.

[0049] The reprojected vector is the multi-source evidence vector obtained through cross-resolution aggregation. This vector integrates information from all relevant hyperedges, with high-resolution and highly relevant hyperedges dominating. This preserves the details of high-resolution data while maintaining spatial consistency across different resolutions. A resolution-aware radial weighting mechanism effectively ensures the dominant role of high-resolution data in aggregation, preventing the dilution of high-resolution details by low-resolution data. Furthermore, combining relevance scoring ensures a high degree of match between the aggregation results and user queries. High-fidelity fusion of multi-resolution data is achieved through Einstein midpoint calculation in hyperbolic space. The resulting multi-source evidence combines rich detail, spatial consistency, and query relevance, providing high-quality foundational data for subsequent conflict triage and generation.

[0050] Optionally, the conflict triage of each of the multi-source pieces of evidence includes: For any two pieces of multi-source evidence, the two pieces of multi-source evidence and the query index are respectively input into the large language model to obtain the corresponding probability distribution entropy.

[0051] Specifically, select any one piece of multi-source evidence to be evaluated, and input the two multi-source data and the user query index into the large language model. When generating the answer, the large language model will output the probability distribution of each word. Based on this distribution, the probability distribution entropy is calculated. The higher the entropy value, the greater the uncertainty of the model in generating the answer based on the evidence. The lower the entropy value, the clearer the evidence and the more stable the model generation.

[0052] Based on the probability distribution entropy, the cross-source interaction entropy of the two multi-source pieces of evidence is obtained, and the two multi-source pieces of evidence whose cross-source interaction entropy exceeds a preset threshold are determined to be in conflict.

[0053] Specifically, two multi-source datasets are used as context and input into the large language model along with the same query index. When the model generates an answer, it outputs a joint probability distribution, and the corresponding joint probability distribution entropy is calculated. This entropy value reflects the uncertainty of the model when generating an answer based on two pieces of evidence simultaneously. If the two pieces of multi-source evidence are completely consistent, the uncertainty of the model will not increase significantly when input simultaneously, the joint probability distribution entropy is close to the entropy of the two probability distributions, and the interaction entropy is close to 0. If the two pieces of evidence conflict, the uncertainty of the model will increase significantly when generating an answer based on contradictory information simultaneously, the joint probability distribution entropy will be significantly higher than the average entropy of the two individual inputs, and the interaction entropy will be positive. Conversely, a negative value indicates complementarity. The cross-source interaction entropy is calculated based on the probability distribution entropy and the joint probability distribution entropy, and is expressed as: , in, p represents the cross-source interaction entropy. i and p jdenoted as multi-source evidence to be evaluated, ans represents the response text generated by the model, P() represents the conditional probability distribution, H() represents the information entropy function, and ⊕ represents the concatenation operation.

[0054] A conflict determination threshold is preset, which can be determined through experimental calibration based on model characteristics and data scenarios. If the cross-source interaction entropy is greater than the preset threshold, a significant conflict is determined between the two pieces of multi-source evidence; if it does not exceed the preset threshold, the two pieces of multi-source evidence are determined to have no significant conflict and good consistency.

[0055] This invention quantifies the uncertainty impact of multi-source evidence on the model generation process by using probability distribution entropy, and accurately captures contradictory information between evidence by using cross-source interaction entropy, thus realizing unsupervised automatic detection of conflicts without relying on manual annotation or domain prior knowledge. This avoids misjudging normal observation differences as noise and can effectively identify real contradictions and conflicts, laying the foundation for subsequent conflict triage and confidence calibration.

[0056] Optionally, the step of calibrating the multi-source evidence according to the triage results to obtain calibrated enhanced evidence includes: In response to the triage result indicating a conflict, the fused feature vector of the conflicting multi-source evidence is input into a preset second classifier to obtain the conflict type.

[0057] Specifically, all multi-source evidence pairs deemed to be conflicting are selected. A fusion feature vector is constructed for each pair, containing cross-source interaction entropy, differences in observation geometric parameters, logarithm of resolution ratio, time interval, data source authority ratio, and hidden state features of the large language model. These features characterize the nature and causes of the conflict from multiple dimensions. For example, if the conflict originates from inherent instrument differences, the differences in observation geometric parameters will be significantly higher; if it originates from scale differences, the logarithm of resolution ratio will become the dominant feature. A second classifier is pre-defined as a conflict type classification model trained using supervised learning. The input is the aforementioned fusion feature vector, and the output is the conflict type label. The classifier classifies conflicts into four categories: noisy conflict, instrument-inherent conflict, scale-dependent conflict, and temporal evolution conflict. Noisy conflict refers to scientifically meaningless conflicts caused by low-authority data sources or observation errors; instrument-inherent conflict refers to interpretable conflicts caused by differences in the observation principles and geometric perspectives of different instruments; scale-dependent conflict refers to observational biases caused by differences in spatial resolution; and temporal evolution conflict refers to differences in target state changes caused by different observation times. These physical features and semantic hidden state features are independent of each other and complementary, making the four types of conflicts linearly separable in the augmented feature space.

[0058] The corresponding score is used as the base confidence level, and the base confidence level is recalibrated according to the conflict type to obtain the calibrated enhanced evidence.

[0059] Specifically, the relevance score obtained during the cross-resolution aggregation stage of the multi-source evidence is extracted as the base confidence level. This score reflects the semantic and spatial matching degree between the evidence and the user query; a higher score indicates that the evidence itself better matches the query requirements, and the initial confidence level is higher. Differentiated confidence recalibration strategies are adopted for different conflict types: For noise conflicts: apply a penalty coefficient to lower the base confidence level and suppress the impact of invalid noise evidence; for instrument-inherent conflicts or scale-dependent conflicts: apply a positive boost coefficient to preserve or even strengthen these scientifically valuable observational disagreements and enhance the integrity of the evidence; for temporal evolution conflicts: apply a time decay weighting, adjusting the confidence level based on the distance between the observation time and the current query time, with evidence from more recent observations having higher confidence, expressed as: , Among them, C triage (v) represents the calibrated confidence level, i.e., the confidence level of the enhanced evidence, C base (v) represents the base confidence level, v represents the multi-source evidence vector to be calibrated, and C detected This indicates that a set of conflicting evidence has been detected. This represents the conflict type label output by the preset second classifier; α represents the confidence retention coefficient for noise conflict, with a value range of (0,1), controlling the degree of suppression of the base confidence under noise conflict; η represents the lower confidence limit for noise conflict; noise represents noise conflict; and β represents the confidence enhancement coefficient for instrument / scale conflict. Let {inst, scale} represent the normalized cross-source interaction entropy, {inst, scale} represent the instrument-inherent conflict and scale-dependent conflict, and temp represent the temporal evolution conflict. This represents the exponential decay function. It should be noted that the above differential recalibration mechanism ensures that scientifically valuable conflict nodes are not filtered out by the confidence mechanism. For instrument-inherent conflicts and scale-dependent conflicts, since β > 0 and the interaction entropy... >Preset threshold > 0, the confidence level after recalibration is strictly greater than the base confidence level; for temporal evolution conflicts, the time decay function >1 Ensure that scientifically significant time-series comparisons are amplified.

[0060] This invention achieves refined processing of conflicts of different origins by classifying and identifying conflicting evidence and calibrating differentiated confidence levels. It avoids the shortcomings of "one-size-fits-all" conflict filtering, and while eliminating invalid noise, it fully preserves the scientific observation discrepancies caused by differences in instruments, scale, and time series in deep space exploration scenarios. This significantly improves the reliability, completeness, and scientific value of multi-source evidence, and provides high-quality contextual support for subsequent large language model generation.

[0061] Optionally, the step of using a spatial adapter to parse and process the pre-acquired multi-source remote sensing data to obtain a fused dataset includes: Match the corresponding spatial adapter according to the type of the multi-source remote sensing data; The spatial adapter is used to perform metadata parsing, spatial coverage extraction, time window unification, and observation parameter normalization on the matched multi-source remote sensing data to obtain a fused dataset under a globally unified spatiotemporal coordinate system.

[0062] Specifically, multiple spatial adapters are predefined based on the data source type, such as orbital remote sensing data adapters, in-situ sounding data adapters, and derived data product adapters. After acquiring multi-source remote sensing data, the data source and type are automatically identified and matched with the corresponding spatial adapter type to ensure that the parsing logic is consistent with the data source structure. Using the matched spatial adapters, structured metadata extraction is performed on the remote sensing data, including instrument number, payload type, data version, production unit, data quality identifier, calibration parameters, etc., converting unstructured and semi-structured metadata into structured information in a unified format. For example, for orbital remote sensing data, the image corner coordinates, projection method, ground sampling distance, imaging swath width, and footprint geometry information are extracted; for example, for in-situ sounding data, the sounding trajectory points, sounding range, effective sounding radius, and relative positioning coordinates are extracted; for example, for derived data products, the spatial range corresponding to the original data is extracted to form a unified spatial location description. The original observation time, imaging time, and acquisition timestamps from each data source are extracted. Different time bases and formats are then uniformly converted to a globally unified spatiotemporal coordinate system, completing time alignment and time window standardization to ensure that data from different sources can be compared and correlated on the same time axis. Observational parameters such as spectral bands, radiometric accuracy, incident angle, detection depth, penetration capability, signal-to-noise ratio, and observation geometry are dimensionless, standardized, and normalized to eliminate numerical deviations and magnitude differences caused by instrument variations, making the observational parameters from different sources comparable and fusionable. All multi-source data that have undergone metadata parsing, spatial range extraction, time unification, and parameter normalization are integrated to form a globally consistent fused dataset in terms of spatiotemporal base, data format, and parameter system, which can be directly used for constructing space observation hypermaps.

[0063] This invention enables accurate parsing and standardized processing of multi-source heterogeneous remote sensing data through a typed spatial adapter, achieving spatiotemporal benchmark unification and observation parameter normalization. This eliminates heterogeneous differences between data from different platforms at the source, improves data availability and consistency, and provides high-quality, directly usable data support for subsequent hypergraph construction, hyperbolic embedding, and retrieval aggregation.

[0064] like Figure 2As shown in the figure, an embodiment of the present invention provides a multi-source remote sensing data retrieval enhancement generation device, comprising: The parsing module is used to parse and process pre-acquired multi-source remote sensing data using a spatial adapter to obtain a fused dataset; The mapping module is used to construct a spatial observation hypergraph based on the fused dataset, and map the spatial observation hypergraph to hyperbolic space according to the coupling relationship to obtain a hyperbolic hypergraph, wherein the coupling relationship includes coupling the radial depth of the hyperbolic space with the spatial resolution. The aggregation module is used to extract relevant hyperedges in the hyperbolic hypergraph in response to a query index given by the user, and perform cross-resolution aggregation on the corresponding relevant hyperedges based on the coupling relationship to obtain the corresponding multi-source evidence; The calibration module is used to perform conflict triage on each of the multi-source evidences, calibrate the corresponding multi-source evidences according to the triage results, and obtain calibrated enhanced evidence. The input module is used to input the multi-source evidence and the enhanced evidence as retrieval enhanced context information into the large language model.

[0065] like Figure 3 As shown, an electronic device 300 provided in this embodiment of the invention includes a memory 310 and a processor 320; the memory 310 is used to store a computer program; the processor 320 is used to implement the multi-source remote sensing data retrieval enhancement generation method as described above when the computer program is executed.

[0066] Alternatively, an electronic device 300 includes a memory 310 and a processor 320 coupled to the memory 310; the memory 310 is configured to store a computer program; and the processor 320 is configured to perform the following operations when the computer program is executed: A spatial adapter is used to analyze and process the pre-acquired multi-source remote sensing data to obtain a fused dataset; A spatial observation hypergraph is constructed based on the fused dataset, and the spatial observation hypergraph is mapped to hyperbolic space according to the coupling relationship to obtain a hyperbolic hypergraph. The coupling relationship includes coupling the radial depth of the hyperbolic space with the spatial resolution. In response to the query index given by the user, relevant hyperedges in the hyperbolic hypergraph are extracted, and cross-resolution aggregation is performed on the corresponding relevant hyperedges based on the coupling relationship to obtain the corresponding multi-source evidence; Each piece of multi-source evidence is subjected to conflict triage, and the corresponding multi-source evidence is calibrated according to the triage results to obtain calibrated enhanced evidence. The multi-source evidence and the enhanced evidence are used as retrieval enhancement context information input into the large language model.

[0067] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the multi-source remote sensing data retrieval enhancement generation method described above.

[0068] Alternatively, a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the following operations: A spatial adapter is used to analyze and process the pre-acquired multi-source remote sensing data to obtain a fused dataset; A spatial observation hypergraph is constructed based on the fused dataset, and the spatial observation hypergraph is mapped to hyperbolic space according to the coupling relationship to obtain a hyperbolic hypergraph. The coupling relationship includes coupling the radial depth of the hyperbolic space with the spatial resolution. In response to the query index given by the user, relevant hyperedges in the hyperbolic hypergraph are extracted, and cross-resolution aggregation is performed on the corresponding relevant hyperedges based on the coupling relationship to obtain the corresponding multi-source evidence; Each piece of multi-source evidence is subjected to conflict triage, and the corresponding multi-source evidence is calibrated according to the triage results to obtain calibrated enhanced evidence. The multi-source evidence and the enhanced evidence are used as retrieval enhancement context information input into the large language model.

[0069] The present invention will now be described an electronic device 300 that can serve as a server or client of the present invention, which is an example of a hardware device that can be applied to various aspects of the present invention. Electronic device 300 is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 300 can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0070] Electronic device 300 includes a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or a computer program loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0071] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention according to actual needs. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units.

[0072] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for enhancing and generating multi-source remote sensing data retrieval, characterized in that, include: A spatial adapter is used to analyze and process the pre-acquired multi-source remote sensing data to obtain a fused dataset; A spatial observation hypergraph is constructed based on the fused dataset, and the spatial observation hypergraph is mapped to hyperbolic space according to the coupling relationship to obtain a hyperbolic hypergraph. The coupling relationship includes coupling the radial depth of the hyperbolic space with the spatial resolution. In response to the query index given by the user, relevant hyperedges in the hyperbolic hypergraph are extracted, and cross-resolution aggregation is performed on the corresponding relevant hyperedges based on the coupling relationship to obtain the corresponding multi-source evidence; Each piece of multi-source evidence is subjected to conflict triage, and the corresponding multi-source evidence is calibrated according to the triage results to obtain calibrated enhanced evidence. The multi-source evidence and the enhanced evidence are used as retrieval enhancement context information input into the large language model.

2. The multi-source remote sensing data retrieval enhancement generation method according to claim 1, characterized in that, In response to a user-provided query index, relevant hyperedges in the hyperbolic hypergraph are extracted, and cross-resolution aggregation is performed on the corresponding relevant hyperedges based on the coupling relationship to obtain corresponding multi-source evidence, including: In response to the query index, spatial intent elements and topic entities are extracted from the query index; Based on the subject entity, retrieve the hyperedges that satisfy the spatial intent in the hyperbolic hypergraph to obtain the relevant hyperedges; Spatiotemporal encoding is performed on each of the related hyperedges to obtain a spatiotemporal encoded vector. A preset first classifier is used to score the relevance of each of the related hyperedges to obtain a corresponding score. Based on the spatiotemporal coding vector, the score, and the coupling relationship, cross-resolution aggregation is performed on the corresponding relevant hyperedges to obtain the corresponding multi-source evidence.

3. The multi-source remote sensing data retrieval enhancement generation method according to claim 2, characterized in that, The step of performing spatiotemporal encoding on each of the relevant hyperedges to obtain a spatiotemporal encoded vector includes: Based on each of the relevant hyperedges, a corresponding pseudo binary triplet is generated, consisting of the topic entity, the relevant hyperedge, and the associated entity, wherein the associated entity includes ground object entities or geological object entities that have spatial or observational association with the topic entity; Based on the pseudo-binary triples, semantic embedding, structural nearest neighbor encoding, and hyperbolic space encoding are fused to construct the spatiotemporal encoding vector for each relevant hyperedge.

4. The multi-source remote sensing data retrieval enhancement generation method according to claim 2, characterized in that, The method of cross-resolution aggregation based on the spatiotemporal coding vector, the score, and the coupling relationship, respectively, to obtain the corresponding multi-source evidence, includes: Resolution-aware radial weighting is determined based on the relationship between the radial depth and the spatial resolution, wherein the radial depth is proportional to the spatial resolution; For each of the relevant hyperedges, the product of the score and the resolution-perceived radial weighting is used as the comprehensive weight. Based on the spatiotemporal coding vector, the spatial outward Einstein midpoint projection vector projected onto the hyperbolic space is calculated to obtain the corresponding multi-source evidence.

5. The multi-source remote sensing data retrieval enhancement generation method according to claim 1, characterized in that, The conflict triage of each of the multi-source pieces of evidence includes: For any two of the multi-source evidences, the two multi-source evidences and the query index are respectively input into the large language model to obtain the corresponding probability distribution entropy; Based on the probability distribution entropy, the cross-source interaction entropy of the two multi-source pieces of evidence is obtained, and the two multi-source pieces of evidence whose cross-source interaction entropy exceeds a preset threshold are determined to be in conflict.

6. The multi-source remote sensing data retrieval enhancement generation method according to claim 5, characterized in that, The step of calibrating the multi-source evidence according to the triage results to obtain calibrated enhanced evidence includes: In response to the triage result indicating a conflict, the fused feature vector of the conflicting multi-source evidence is input into a preset second classifier to obtain the conflict type; The corresponding score is used as the base confidence level, and the base confidence level is recalibrated according to the conflict type to obtain the calibrated enhanced evidence.

7. The multi-source remote sensing data retrieval enhancement generation method according to claim 1, characterized in that, The process involves using a spatial adapter to parse and process pre-acquired multi-source remote sensing data to obtain a fused dataset, including: Match the corresponding spatial adapter according to the type of the multi-source remote sensing data; The spatial adapter is used to perform metadata parsing, spatial coverage extraction, time window unification, and observation parameter normalization on the matched multi-source remote sensing data to obtain a fused dataset under a globally unified spatiotemporal coordinate system.

8. A multi-source remote sensing data retrieval enhancement generation device, characterized in that, include: The parsing module is used to parse and process pre-acquired multi-source remote sensing data using a spatial adapter to obtain a fused dataset; The mapping module is used to construct a spatial observation hypergraph based on the fused dataset, and map the spatial observation hypergraph to hyperbolic space according to the coupling relationship to obtain a hyperbolic hypergraph, wherein the coupling relationship includes coupling the radial depth of the hyperbolic space with the spatial resolution. The aggregation module is used to extract relevant hyperedges in the hyperbolic hypergraph in response to a query index given by the user, and perform cross-resolution aggregation on the corresponding relevant hyperedges based on the coupling relationship to obtain the corresponding multi-source evidence; The calibration module is used to perform conflict triage on each of the multi-source evidences, calibrate the corresponding multi-source evidences according to the triage results, and obtain calibrated enhanced evidence. The input module is used to input the multi-source evidence and the enhanced evidence as retrieval enhanced context information into the large language model.

9. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the multi-source remote sensing data retrieval enhancement generation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the multi-source remote sensing data retrieval enhancement generation method as described in any one of claims 1 to 7.

Images