A coastal vegetation index spatio-temporal reconstruction method and system based on geographic logic gating and multi-agent

Abstract

This invention discloses a spatiotemporal reconstruction method and system for coastal vegetation indices based on geographic logic gating and multi-agent systems, belonging to the fields of remote sensing ecological monitoring and deep learning. The method acquires and preprocesses multi-temporal high- and low-resolution remote sensing images to construct NDVI spatiotemporal feature sequences. Spatiotemporal features are obtained through deep encoding, and a multi-agent system incorporating information on vegetation, cloud cover, tides, and topography is established. Coastal geographic logic constraints are introduced to construct multi-scale spatiotemporal differences and generate geographic logic gating, which weights and modulates the spatiotemporal features to enhance reliable information and suppress interference noise. Multi-agent collaborative reconstruction is achieved under multi-objective constraints, outputting high-resolution vegetation indices and supporting edge deployment and online learning. This invention effectively overcomes problems such as cloud and fog obstruction, tidal inundation, and data gaps, significantly improving reconstruction accuracy, geographic rationality, and spatiotemporal continuity. It is more adaptable to the complex ecological environment of coastal zones, providing a stable and reliable technical means for long-term dynamic monitoring of vegetation.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing ecological monitoring and deep learning spatiotemporal data reconstruction, specifically involving a spatiotemporal reconstruction method and system for coastal vegetation index based on geographic logic gating and multi-agent systems. Background Technology

[0002] Coastal zones are characterized by intense land-sea interactions, complex ecological types, strong spatial heterogeneity, and significant spatiotemporal variations. Vegetation indices, as important remote sensing parameters reflecting vegetation growth and ecological changes, have significant application value in the monitoring and assessment of coastal ecosystems such as mangroves, salt marshes, and tidal flats.

[0003] However, in practical applications, the continuous acquisition of coastal vegetation indices in time and space faces numerous technical challenges. On the one hand, due to factors such as frequent cloud cover, thin cloud obstruction, and complex atmospheric conditions, medium- and high-resolution optical remote sensing images often suffer from large-scale and long-term data gaps in coastal areas, resulting in incomplete vegetation index time series and making it difficult to meet the needs of continuous monitoring. On the other hand, coastal areas commonly experience tidal fluctuations, alternation between land and water, and mixed water pixels. The same spatial location may exhibit different surface cover states such as water, bare land, or vegetation at different times, causing significant non-stationary changes in vegetation indices over time. If these differences are not properly distinguished, they can easily be misjudged as actual vegetation changes during reconstruction, thus introducing significant errors.

[0004] Existing spatiotemporal reconstruction methods for vegetation indices mainly include those based on statistical interpolation, time-series filtering, and deep learning models. Among these, statistical or filtering methods typically assume that vegetation index changes are smooth or periodic, making it difficult to effectively characterize the abrupt changes driven by tides in coastal areas. In recent years, spatiotemporal fusion methods based on deep learning have been widely used in remote sensing reconstruction; however, these methods mostly rely on feature learning from the image data itself, lacking explicit constraints on the unique geographical processes and environmental logic of coastal zones.

[0005] Specifically, existing deep learning spatiotemporal fusion models typically assign similar propagation mechanisms to features at different times and locations during feature transfer and information fusion. They fail to fully consider the impact of geographical factors such as tidal conditions, distance from the shoreline, and degree of water body influence on the reliability of vegetation index observations. This can easily introduce features that do not conform to geographical laws into the reconstruction results, leading to a decrease in reconstruction accuracy and stability. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for spatiotemporal reconstruction of coastal vegetation indices based on geographic logic gating and multi-agent systems. This method introduces a cyclically applicable geographic logic gating mechanism into the encoding-decoding structure, combined with multi-agent distributed spatiotemporal modeling and collaborative optimization, to achieve dynamic filtering, modulation, and constraint of spatiotemporal features. Even in complex coastal environments characterized by cloud cover, tidal inundation, and alternating water and land conditions, it can still obtain stable and reliable vegetation index reconstruction results with stronger spatiotemporal continuity, geographic consistency, and higher accuracy.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows:

[0008] A spatiotemporal reconstruction method for coastal vegetation indices based on geographic logic gating and multi-agent systems includes the following steps:

[0009] Step S1. Acquisition and preprocessing of multi-source remote sensing data: Acquire multi-temporal high and low resolution optical remote sensing images of the study area, including low-resolution high temporal resolution images LR(t0) and LR(t1) and high-resolution low temporal resolution images LR(t0); perform radiometric correction, atmospheric correction, geometric registration, and cloud and anomalous pixel labeling.

[0010] To achieve standardization of the radiometric gauge, the radiometric calibration results are first converted into top-atmosphere reflectance, specifically using the classical form:

[0011]

[0012] in The zenith angle of the sun. This is the Earth-Sun distance correction term, where The zenith angle of the sun. For the Earth-Sun distance correction term, L λ Converted to reflectivity ρ λ To control the adverse effects of abnormally high values ​​on subsequent indices and training.

[0013] Based on this, the Normalized Difference Vegetation Index (NDVI), which characterizes vegetation growth status, is calculated using surface reflectance. The NDVI is calculated based on the fused multispectral image, and the formula is as follows:

[0014]

[0015] in, and These represent the near-infrared and red bands of the Gaofen-6 multispectral image, respectively.

[0016] Step S2. Construction of Spatiotemporal Feature Sequence of Vegetation Indices: Based on the spatial distribution pattern and temporal change sequence of vegetation indices, multi-temporal vegetation indices are organized into a unified spatiotemporal structure to form a spatiotemporal feature sequence that can be used as model input, thereby fully representing the coordinated change pattern of vegetation indices in the temporal and spatial dimensions. The specific method for modeling the spatiotemporal features of vegetation indices is as follows:

[0017] NDVI images acquired at different times were constructed into a spatiotemporal sequence of vegetation indices in chronological order:

[0018]

[0019] in, Indicates the first NDVI distribution at each time phase, This represents the length of the time series.

[0020] Step S3. Spatiotemporal Feature Encoding and Multi-Agent System Construction: To capture the potential variation patterns of vegetation index in different time phases and spatial locations, deep encoding of the spatiotemporal feature sequences is performed to extract spatiotemporal correlation features. First, candidate information describing the changes in NDVI state between adjacent time phases is constructed:

[0021]

[0022] in, and These represent the spatiotemporal characteristics of adjacent time phases.

[0023] To reduce the impact of short-term abnormal fluctuations, a reference time phase is selected. Construct candidate information for changes relative to the reference state:

[0024]

[0025] The above change information is input into the encoder to obtain the spatiotemporal encoded features used for subsequent reconstruction and gating generation, thus obtaining the corresponding spatiotemporal encoded features:

[0026]

[0027] in, Indicates the first Spatiotemporal coding features of each phase.

[0028] At the same time, the reconstruction units at different spatiotemporal locations are defined as agents. i Construct a multi-agent system MAS={Agent1,Agent2,…,Agent2} n n};S_{n}, the state space of each agent is S_{n} i Represented as: Sᵢ = [NDVIi ,Δ t NDVI, Cloud i Tide i Water i Dist i DEM i Neighbor i ], of which NDVI i Δ represents the current pixel vegetation index. t NDVI is a time-varying variable, Cloud i For cloud cover, Tide i In a tidal state, Water i Dist represents the probability of water bodies. i DEM is the distance from the shoreline. i Neighbor is the terrain elevation. i This refers to the state information of neighboring intelligent agents.

[0029] Step S4. Construction of Geographic Logical Constraint Information: Considering the characteristics of coastal zones—tidal patterns, alternating water and land conditions, and frequent inundation—geographic logical information is constructed, including tidal state, inundation probability, distance from the shoreline, water mixing degree, topographic elevation, and land use type. To eliminate the influence of differences in the dimensions of different geographic elements, continuous geographic variables are normalized and mapped.

[0030]

[0031] in, For the original geographical variables, and The minimum and maximum values ​​are respectively used to combine the normalized geographic variables with land use category information to construct geographic logical constraint features, which are used to describe the geographic environmental conditions in which the vegetation index changes.

[0032] Step S5. Multi-scale spatiotemporal difference and geographic logic gating generation: In order to comprehensively characterize the changing characteristics of vegetation indices, multi-scale spatiotemporal difference features are constructed from three dimensions: temporal adjacency, reference steady state, and spatial neighborhood.

[0033] Adjacent phase differences (short-term variations):

[0034]

[0035] Reference phase differential (steady-state constraint): Selecting a reference phase , build

[0036]

[0037] Spatial difference (mixed pixel sensitivity):

[0038]

[0039] The above differential candidate information is fused to form a multi-scale differential representation:

[0040]

[0041] in, This represents the differential fusion mapping function, which can be a concatenated multilayer perceptron, attention fusion, or a combination thereof.

[0042] Simultaneously, a geographic logical constraint vector P is constructed by integrating information such as tides, shorelines, water bodies, and topography.

[0043]

[0044] For the geographic logical constraint vector By performing mapping and encoding, a geographic logical embedding representation is obtained:

[0045]

[0046] in, It can be a multilayer perceptron, an embedded map, or a combination thereof, used to map continuous and discrete geographic variables to a unified feature space.

[0047] Multiscale difference representation Geographical logic embedding Compared with the original spatiotemporal coding features By using a common input function to determine the reasonableness of changes, a score for the reasonableness of the changes is obtained.

[0048]

[0049] in, This is a fusion mapping function used to integrate change information with geographic logic information to determine whether the change conforms to coastal zone geographic processes.

[0050] Convert ratings to change confidence using the Sigmoid function or piecewise mapping:

[0051]

[0052] in, This indicates that the changes are highly credible in the current geographical context; This indicates that the changes are more likely caused by tidal inundation, water mixing, or non-vegetation disturbance.

[0053] To avoid the gating process relying entirely on data-driven results, an explicit suppression rule based on geographic logic is introduced to correct the change confidence level, resulting in the corrected change confidence level:

[0054]

[0055] in, This is a geographic logical suppression function used to express one or a combination of the following rules: when When the value exceeds a set threshold, the reliability of the corresponding change decreases; when Smaller and At lower levels (in low-lying nearshore areas), a penalty is imposed on the reliability of changes; when the land use type indicates a water body or tidal flat, high-frequency changes are suppressed.

[0056] Using the corrected change confidence level as the geographic logical gating weight, we obtain the gating information:

[0057]

[0058] The It is used to characterize the credibility of features at each spatiotemporal location in the reconstruction process, and can use continuous weights, segmented weights or binary gating to enable, enhance or suppress features.

[0059] Step S6. Gated Modulation and Trusted Feature Enhancement

[0060] By applying geographic logic gating information to spatiotemporal encoded features and performing element-wise weighted modulation on these features, enhanced spatiotemporal features constrained by geographic logic are obtained:

[0061]

[0062] Where F represents the original spatiotemporal coding features, and G represents the geographic logical gating weights. For element-wise multiplication, F' represents the enhanced spatiotemporal features after gating;

[0063] By using gating modulation, unreliable features caused by tidal inundation, water-land mixing, and cloud and fog obstruction are suppressed, while the geographically logically reliable spatiotemporal features of vegetation are enhanced, providing pure feature input for subsequent reconstruction.

[0064] Step S7. Multi-agent cooperative reconstruction and multi-objective constraint optimization

[0065] Design a multi-objective constraint function to achieve joint optimization of reconstruction accuracy, geographic compliance, temporal smoothness, and spatial consistency. The total loss function is expressed as:

[0066]

[0067] Where: L1 is the reconstruction accuracy constraint, using mean square error loss:

[0068]

[0069] L2 is a geographic compliance constraint used to penalize anomalous mutations that do not conform to geographic logic.

[0070]

[0071] L3 is a timing smoothing constraint that suppresses anomalous timing transitions.

[0072]

[0073] L4 is a neighborhood cooperation constraint that ensures spatial continuity and consistency.

[0074]

[0075] This is the balance coefficient;

[0076] Each agent performs local reconstruction reasoning under geographic logic gating constraints and engages in state interaction and information collaboration with neighboring agents.

[0077]

[0078] Where S i Given the current state of the agent, {S} j Let G be the state of the neighboring agent. i It uses geographic logic gating to achieve spatiotemporal consistency reconstruction through global collaborative optimization.

[0079] Step S8. Deployment of Spatiotemporal Reconstruction and Online Inference of Vegetation Index

[0080] Decoding and reconstruction are performed based on the gated enhanced spatiotemporal features, outputting high-resolution vegetation index results at the target time:

[0081]

[0082] in For spatiotemporal decoding networks, This is the final high-resolution NDVI reconstruction result.

[0083] The trained model is deployed to the remote sensing edge computing unit to build an online fine-tuning learning mechanism:

[0084]

[0085] in For model parameters, For learning rate, For online losses based on newly added observation data;

[0086] The system supports real-time remote sensing data access, rapid edge inference, and dynamic adaptive updates. It can stably adapt to changes in cloud and fog, tidal fluctuations, and seasonal changes over a long period of time, enabling continuous monitoring of coastal vegetation indices over long time series.

[0087] This invention also provides a spatiotemporal reconstruction system for coastal vegetation indices based on geographic logic gating and multi-agent systems, comprising a data preprocessing module, a spatiotemporal feature construction module, a multi-agent modeling module, a geographic logic constraint module, a geographic logic gating generation module, a feature modulation module, a multi-agent collaborative reconstruction module, and an edge reasoning and online learning module; the system is used to execute the methods described.

[0088] Beneficial effects

[0089] (1) This invention can effectively filter out abnormal observation information caused by cloud cover, abnormal pixels and spatiotemporal change credibility by dynamically judging and gating suppression. It can maintain the stability and continuity of reconstruction results even under large-scale and long-term data loss conditions, and solve the problem that the comparison method is easily affected by noise and outliers and the reconstruction results are distorted.

[0090] (2) In response to the non-stationary abrupt changes in vegetation index caused by tidal fluctuations, alternation between land and water, and mixed water pixels in the coastal zone, this invention introduces explicit geographic logic constraints such as tidal state, inundation probability, distance from the shoreline, and topographic elevation. It can automatically identify and suppress false changes caused by tidal inundation and water cover, avoid misjudging non-vegetation disturbances as vegetation growth changes, and significantly improve the geographic rationality and authenticity of the reconstruction results.

[0091] (3) The geographic logic gating proposed in this invention can dynamically weight and modulate spatiotemporal features, automatically enhance reliable vegetation features that conform to the ecological laws of the coastal zone, and suppress unreliable disturbance features. It breaks through the limitations of traditional deep learning models that adopt a unified feature propagation mechanism for all spatiotemporal locations and lack geographic prior constraints, making the reconstruction results more consistent with the real surface processes.

[0092] (4) By modeling each spatiotemporal location as an independent intelligent agent and conducting neighborhood state interaction and collaborative decision-making, this invention can ensure global spatial smoothness and spatiotemporal consistency while performing local fine reconstruction. It effectively alleviates the reconstruction jump and patch distortion problems caused by the high spatial heterogeneity of the coastal zone, making the reconstruction results more continuous in space, more natural in texture, and more accurate. Attached Figure Description

[0093] Figure 1 Flowchart of the method described in this invention;

[0094] Figure 2 A schematic diagram of the geographic logic gating unit structure constructed in this invention;

[0095] Figure 3 This invention provides a network structure diagram for geographic logic gating and multi-agent reconstruction.

[0096] Figure 4 Comparison of reconstruction accuracy under different cloud and fog coverage rates;

[0097] Figure 5 Comparison of errors under different tidal inundation intensities;

[0098] Figure 6 Comparison of reconstruction results between the present invention and the comparative method. Detailed Implementation

[0099] To make the above-mentioned objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0100] This embodiment takes a typical 2000m×2000m tidal flat-salt marsh complex ecological area in the southeastern coastal zone of my country as the study area. Using remote sensing data from the Gaofen-6 GF-6 satellite, including 2m panchromatic image and 16m multispectral image, low-resolution high-temporal-series images LR(t0) and LR(t1) and high-resolution low-temporal-resolution image LR(t0) are constructed to carry out a spatiotemporal reconstruction experiment of vegetation index.

[0101] Step A100: Multi-source remote sensing data acquisition and preprocessing: Refer to Appendix Figure 1 The overall process shown first involves acquiring multi-temporal high and low resolution optical remote sensing images of the study area, completing radiometric correction, atmospheric correction, geometric registration, cloud and anomalous pixel labeling, converting radiance values ​​into atmospheric top reflectance, and calculating the Normalized Difference Vegetation Index (NDVI) based on the near-infrared and red bands.

[0102] Step A200: Construction of spatiotemporal feature sequence of vegetation index: According to the temporal and spatial patterns, multi-temporal NDVI images are organized into a unified spatiotemporal structure to form a spatiotemporal feature sequence that can be input into a deep learning network, which fully represents the coordinated change law of vegetation index in the temporal and spatial dimensions.

[0103] Step A300: Spatiotemporal Feature Encoding and Multi-Agent System Construction: Deeply encode the NDVI spatiotemporal feature sequence to extract spatiotemporal correlation features; define each pixel's spatiotemporal location as an independent agent to construct a multi-agent system (MAS); the state space of each agent is: Sᵢ = [NDVI i ,Δ t NDVI, Cloud i Tide i Water iDist i DEM i Neighbor i This enables a unified expression of vegetation status, temporal changes, cloud cover, tides, water bodies, topography, and neighborhood information.

[0104] Step A400: Construction of Geographic Logical Constraint Information: Based on the characteristics of tidal inundation and alternation between water and land in the coastal zone, construct geographic logical information such as tidal state, inundation probability, distance from the shoreline, water mixing degree, topographic elevation (DEM), and land use type. Normalize the continuous geographic variables to eliminate dimensional differences and form unified geographic prior constraint features.

[0105] Step A500: Multi-scale Spatiotemporal Difference and Geographic Logical Gating Generation: Refer to Figure 2 The geographic logic gated unit structure shown is as follows:

[0106] (1) Construct multi-scale spatiotemporal difference features from three dimensions: adjacent temporal phases, reference steady state, and spatial neighborhood;

[0107] (2) Generate a geographic logical constraint vector by integrating information on tides, shorelines, topography, and water bodies;

[0108] (3) The correction is completed by embedding encoding, judging the rationality of changes, and mapping credibility, combined with geographical suppression rules;

[0109] (4) Output the geographic logic gating weights that can dynamically adjust the credibility of features.

[0110] Step A600 Gated Modulation and Trusted Feature Enhancement: Geographic logic gating and spatiotemporal coding features are weighted and modulated element by element to enhance trustworthy vegetation features that conform to the geographical laws of the coastal zone, suppress unreliable features caused by cloud, tide, and water mixing, and obtain pure and stable enhanced spatiotemporal features.

[0111] Step A700: Multi-agent cooperative reconstruction and multi-objective constraint optimization: Refer to... Figure 3 The network structure shown allows each agent to complete local reconstruction reasoning and interact with neighboring agents in terms of state and information, achieving global spatiotemporal consistency optimization under the constraints of four multi-objectives: reconstruction accuracy, geographical compliance, temporal smoothness, and neighborhood collaboration.

[0112] Step A800 High-Resolution Reconstruction and Edge Inference Deployment: Decode the enhanced spatiotemporal features and output the high-resolution vegetation index result HR(t1) at the target time;

[0113] The trained model is deployed to the remote sensing edge computing unit, supporting real-time data access, rapid inference and online fine-tuning learning, and realizing automated continuous monitoring of long-term series.

[0114] A traditional deep learning-based spatiotemporal fusion reconstruction method is used as a comparative example.

[0115] Figure 4 It can be seen that as the cloud and fog coverage ratio increases, the reconstruction effect of different methods all decreases to a certain extent, but the comparative methods are more significantly affected and the accuracy decreases significantly. The method of the present invention maintains a relatively stable output under different cloud and fog coverage conditions, the overall change is more gradual, and it shows a stronger adaptability to clouds, noise and missing data.

[0116] Figure 5 This reflects the trend of reconstruction error variation under different tidal inundation intensities. Under no or weak inundation conditions, all methods can maintain good reconstruction results; as the tidal inundation intensity increases, the error of the comparative methods increases rapidly, while the error of the method of this invention is lower overall and increases more slowly, and can still maintain stable and reliable output in areas with frequent water-land alternation and strong tidal influence.

[0117] Figure 6 This provides a direct comparison of the reconstruction results of the method of this invention and the comparative method. It can be seen that the reconstruction results of the comparative method have more obvious patchiness, abrupt changes and local anomalies, and weak spatial continuity; the reconstruction results of the method of this invention are smoother in spatial distribution, the temporal changes are more in line with the actual evolution of surface vegetation, the overall texture is natural, the boundary transition is soft, and the consistency with the real surface spatial pattern is higher.

[0118] From the perspective of the overall reconstruction structure and effect, combined with Figure 3 Network architecture and Figure 6 The visualization results show that this invention, through multi-agent distributed modeling and neighborhood information interaction, enables each spatial unit to maintain global consistency while reconstructing locally, effectively alleviating the jump and distortion problems that are prone to occur in highly heterogeneous coastal areas, and improving the continuity and reliability of long-term vegetation indices.

Claims

1. A spatiotemporal reconstruction method for coastal vegetation indices based on geographic logic gating and multi-agent systems, characterized in that, Includes the following steps: S1 acquires multi-temporal high- and low-resolution optical remote sensing images of the study area, performs radiometric correction, atmospheric correction, geometric registration, and anomaly labeling preprocessing, converts the radiometric calibration results into top atmospheric reflectance, and calculates the normalized vegetation index based on the near-infrared and red bands. S2 Constructs a spatiotemporal feature sequence of vegetation index based on the spatial distribution and temporal variation order of NDVI. ; S3 performs deep encoding on the spatiotemporal feature sequence of vegetation index, constructs candidate information for changes in adjacent time phases and candidate information for changes in reference time phases, and inputs it into the encoder to obtain spatiotemporal encoded features F; The reconstruction units at different spatiotemporal locations are defined as agents. i Construct a multi-agent system MAS={Agent1,Agent2,…,Agent2} n n};S_{n}, the state space of each agent is S_{n} i Represented as: Sᵢ = [NDVI i ,Δ t NDVI, Cloud i Tide i Water i Dist i DEM i Neighbor i ]; S4 constructs geographic logical constraints on tidal states, inundation probabilities, distance from shoreline, water mixing degree, topographic elevation, and land use types, and performs normalization mapping on continuous geographic variables. ； S5 constructs multi-scale spatiotemporal difference features from three dimensions: adjacent time phase, reference time phase, and spatial neighborhood. It integrates the geographic logical constraint vector P and generates geographic logical gate G after the change rationality judgment and credibility correction. S6 performs element-wise weighted modulation of the geographic logic gate G and the spatiotemporal coding feature F to obtain the enhanced spatiotemporal features constrained by geographic logic: To suppress unreliable features caused by clouds, tidal inundation, and mixed water and land, and to enhance the spatiotemporal features of vegetation with high credibility; S7, under multi-objective constraints, uses the total loss function. The optimization is performed, where L1 is the reconstruction accuracy loss, L2 is the geographic compliance loss, L3 is the temporal smoothing loss, and L4 is the neighborhood cooperation loss. Each agent performs local reconstruction inference under geographic logic gating constraints and interacts with the state of neighboring agents to achieve global spatiotemporal consistency optimization. S8 decodes and reconstructs the spatiotemporal features F' based on gating enhancement, outputs a high-resolution vegetation index at the target time, and deploys the model to the remote sensing edge computing unit to achieve real-time data access, rapid inference, and online fine-tuning learning.

2. The method according to claim 1, characterized in that, Step S1 calculates NDVI based on the near-infrared and red bands of the Gaofen-6 multispectral image.

3. The method according to claim 1, characterized in that, In step S3, the spatiotemporal coding features are obtained by jointly encoding the candidate information of adjacent temporal phase changes and the candidate information of reference temporal phase changes.

4. The method according to claim 1, characterized in that, In step S4, the normalized geographic variables and land use categories are combined to form a unified geographic logical constraint feature.

5. The method according to claim 1, characterized in that, In step S5, the geographic logic gating adopts continuous weight, segmented weight, or binary gating.

6. The method according to claim 1, characterized in that, In step S7, spatiotemporal consistency reconstruction is achieved through multi-agent local reasoning and neighborhood collaborative decision-making.

7. A spatiotemporal reconstruction system for coastal vegetation indices based on geographic logic gating and multi-agent systems, characterized in that, The system includes a data preprocessing module, a spatiotemporal feature construction module, a multi-agent modeling module, a geographic logic constraint module, a geographic logic gating generation module, a feature modulation module, a multi-agent collaborative reconstruction module, and an edge reasoning and online learning module; the system is used to execute the method described in any one of claims 1-6.

Images