A remote sensing time series data analysis method and system considering spatiotemporal correlation of geographic objects

Abstract

The application discloses a kind of remote sensing time series data analysis method and system considering the spatio-temporal correlation of geographical object, belong to data analysis technical field.The method includes: obtaining multi-source multi-temporal remote sensing image and auxiliary data and pre-processing, generate consistent time series data stack;In three-dimensional space-time domain, image is divided into voxel and is adjacent tracking, identify geographical process object with evolution behavior and record its attribute;Determine the spatio-temporal topological relationship between geographical process object, generate unified spatio-temporal relationship table by composite reasoning;With geographical process object as node, spatio-temporal relationship as edge constructs geographical process object spatio-temporal graph model, extracts and standardizes the attribute characteristics of node and edge;Importance index of each node is calculated by using graph convolution network and topological dynamic mechanism joint modeling.The application realizes closed-loop analysis from data processing, relationship modeling, importance evaluation to decision support, improves the interpretability and decision effectiveness of remote sensing time series change detection.

Description

Technical Field

[0001] This invention relates to the field of data analysis methods, specifically to a remote sensing time-series data analysis method and system that takes into account the spatiotemporal correlation of geographic objects. Background Technology

[0002] The analysis of remote sensing time series data is of great significance in fields such as global change monitoring and land use / cover change (LUCC) research. Existing pixel-level time-series remote sensing image change detection methods can grasp regional change trends as a whole, such as detecting urban expansion or forest cover changes through multi-temporal images. However, these methods mainly focus on absolute changes in geospatial space (i.e., changes in geographic coordinates), neglecting the relationships and evolutionary processes between different geographic objects (such as land cover patches).

[0003] In recent years, to delve deeper into the relationship between local changes and the overall landscape pattern, the academic community has begun to introduce the concept of "object-oriented" spatiotemporal analysis: dividing geographical regions into semantically meaningful patches or objects, and using these geographical process objects as basic units to track and analyze their evolution over time. This approach can better describe the changes within local patches themselves and the transformation processes between patches, thereby linking local changes with changes in the overall landscape pattern.

[0004] Meanwhile, the development of artificial intelligence has driven the application of deep learning in remote sensing time-series analysis. Among these, graph neural networks (GNNs) have gained increasing attention due to their ability to model relationships between entities. In particular, graph convolutional networks (GCNs) are used for scene understanding and change detection in remote sensing images to capture the connections between spatially non-adjacent regions, overcoming the limitation of traditional convolutional neural networks that can only extract local features. Furthermore, in the field of complex network analysis, recent advancements have yielded metrics and algorithms for measuring the importance of network nodes, such as DomiRank centrality. This metric combines local neighborhood dominance with global topological information by introducing adjustable parameters to quantitatively assess the dominance of nodes in the network. In summary, this field is moving towards object-oriented, relational, and intelligent approaches—that is, constructing object-level spatiotemporal relationship models and combining them with advanced network analysis algorithms to more comprehensively understand the changing processes in remote sensing time-series data.

[0005] The applicant has discovered that the prior art has at least the following technical problems:

[0006] (1) Lack of unified analysis of the relationship between local and global changes: Traditional pixel-based time series analysis only focuses on the magnitude of local changes and cannot characterize the transmission and relationship between changes in different regions, resulting in an incomplete understanding of the overall pattern changes. Although object-oriented models can record the evolution process of patches, they are usually used for database storage and simple retrieval, and lack a mechanism for automatically assessing the relative importance of each patch.

[0007] (2) Insufficient integration of spatiotemporal relationships: Some existing methods only consider spatial adjacency relationships (such as simple graph theory connectivity), but do not make full use of information in the temporal dimension and cannot represent the influence relationship between objects over time. For example, existing methods fail to establish temporal inheritance chains of patches and cannot identify how changes in one patch affect neighboring patches in subsequent times. The lack of spatiotemporal integration means that it is difficult to discover the propagation patterns or key nodes of changes.

[0008] (3) Difficulty in identifying key change regions: Most current change detection technologies (including deep learning methods) ultimately output the location or category of the change, without distinguishing which change regions play a core role in the network. In other words, detecting a change does not equate to understanding its importance. The lack of quantification of "node importance" makes it impossible to identify the regions with the greatest impact on the overall structure from numerous changes, potentially missing the timely discovery of ecologically vulnerable areas or key turning points. Summary of the Invention

[0009] The purpose of this invention is to provide a remote sensing time series data analysis method and system that takes into account the spatiotemporal correlation of geographic objects. Addressing the core pain point of the invisible impact of changes in remote sensing time series analysis, this invention proposes a node importance analysis method that integrates DomiRank and Graph Convolutional Network (GCN) on a unified spatiotemporal map of geographic process objects.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] In a first aspect, the present invention provides a remote sensing time-series data analysis method that takes into account the spatiotemporal correlation of geographic objects, including:

[0012] Step 1: Acquire multi-source, multi-temporal remote sensing images and auxiliary data, and preprocess them to generate a consistent time-series data stack;

[0013] Step 2: Divide the image into voxels in the three-dimensional spatiotemporal domain and perform adjacency tracking to identify geographic process objects with evolutionary behavior and record their attributes;

[0014] Step 3: Determine the spatial and temporal topological relationships between geographic process objects, generate a unified spatiotemporal relationship table through composite reasoning, and quantify their directionality, correlation strength, and temporal distance;

[0015] Step 4: Construct a spatiotemporal graph model of geographic process objects with geographic process objects as nodes and unified spatiotemporal relationships as edges. Extract and standardize the attribute features of nodes and edges to form structured data input for graph learning.

[0016] Step 5: Utilize graph convolutional networks and topology dynamics mechanisms to jointly model, calculate the importance index of each node, and output hierarchical layers, a list of key nodes, and interface services that can be connected.

[0017] Further, in step 1, the preprocessing includes:

[0018] Differences in imaging conditions are eliminated through radiometric and atmospheric corrections;

[0019] Geometric correction and registration ensure accurate spatial alignment of data from multiple periods;

[0020] By unifying projection and clipping the range, a time-series data stack with consistent spatial resolution and coordinate system is generated.

[0021] Input quality is ensured through masking.

[0022] Furthermore, step 2 specifically includes:

[0023] Multiple remote sensing images are constructed into spatiotemporal cubes according to spatial and temporal dimensions, with each cube unit representing the observation value of a specific spatial location within a specific time period;

[0024] The spatiotemporal cube is segmented based on spectral features, texture features, and temporal change patterns. The cube is divided into multiple object units with consistent features in both space and time. Each object unit is a continuous region in three-dimensional space and is accompanied by spatial boundaries and category labels.

[0025] In the cube segmentation results, the continuous state of the same object unit in the time dimension is identified and tracked to form geographic process objects. The temporal continuity of the objects is directly determined by the adjacency relationship of voxels within the cube.

[0026] Objects that appear for the first time within a certain time period are marked as newly created; objects that disappear within a certain time period are marked as terminated. Each geographic process object carries complete temporal information, including the time of appearance, the time of termination, and attribute snapshots of each time slice.

[0027] Furthermore, in step 3, a spatiotemporal topological relationship model is used to uniformly calculate the relationships between geographic process objects. Eight types of spatial topological relationships are determined within the same time slice, and 13 types of temporal topological relationships are determined between consecutive time slices. The two are then combined for inference, the relationship strength is calculated, and a unified spatiotemporal relationship table is output.

[0028] Furthermore, step 4 specifically includes:

[0029] A spatiotemporal graph of geographic process objects is constructed using geographic process objects as nodes and unified spatiotemporal relationships as weighted edges. This graph includes an adjacency matrix, a node feature matrix, and an edge feature matrix. Missing data is then handled and standardized. Training, validation, and inference subsets are then segmented according to the application scenario. Serialized graph data objects are output for graph learning and inference. Node features cover location and environment, morphology, evolution, and semantic categories, while edge features include relation type encoding, strength, direction, and time interval.

[0030] Furthermore, step 5 specifically includes:

[0031] The node embedding vector is obtained by convolutional aggregation of node features and neighborhood information through multi-layer GCN.

[0032] DomiRank dynamic competitive iteration is performed based on embedding vectors and topological structure to calculate dominance centrality score;

[0033] The GCN output is weighted and fused with the DomiRank score to generate a node importance index, and then sorted and ranked.

[0034] Secondly, the present invention provides a remote sensing time-series data analysis system that takes into account the spatiotemporal correlation of geographic objects, for performing the aforementioned remote sensing time-series data analysis method that takes into account the spatiotemporal correlation of geographic objects, including:

[0035] The data acquisition and preprocessing module is used to acquire raw data from multi-source and multi-temporal remote sensing data and perform quality control.

[0036] The geographic process object construction module is used to construct spatiotemporal cubes from multiple remote sensing images according to the spatial-temporal dimensions and perform three-dimensional segmentation to obtain continuous and consistent object units in three-dimensional space.

[0037] The spatiotemporal relationship calculation module is used to determine the type and intensity of relationships between objects based on the spatiotemporal topological relationship model.

[0038] The GPO-STG modeling module is used to construct a unified spatiotemporal relationship network with GPOs as nodes and spatiotemporal relationships as edges.

[0039] The feature extraction and attribute enhancement module is used to calculate and enhance the geometric, temporal, semantic, neighborhood, and environmental features of nodes and edges;

[0040] The graph convolutional network embedding and dynamic competition mechanism cascade analysis module is used to perform node importance analysis on a unified spatiotemporal relationship graph;

[0041] The results output and visualization module is used to highlight key nodes in a map overlay format and display their evolution trajectory and important neighborhood relationships; the module outputs a node list, node importance index level and evaluation indicator report, and provides interface integration capabilities with business systems.

[0042] Thirdly, the present invention provides an electronic device, including a processor and a memory; the processor executes the steps of the remote sensing time-series data analysis method that takes into account the spatiotemporal correlation of geographic objects by calling programs or instructions stored in the memory.

[0043] Fourthly, the present invention provides a computer-readable storage medium storing a program or instructions that cause a computer to perform the steps of the remote sensing time-series data analysis method that takes into account the spatiotemporal correlation of geographic objects.

[0044] Based on the above technical solution, the embodiments of the present invention can produce at least the following technical effects:

[0045] This invention revolves around the Geographic Process Object Spatiotemporal Graph (GPO-STG), which uses geographic process objects as nodes and unified spatiotemporal relationships as edges. It integrates spatiotemporal cube segmentation, spatiotemporal topology computation, relationship-aware graph modeling and feature engineering, and a cascaded mechanism of "GCN embedding × DomiRank dynamic competition" within a single graph structure. Ultimately, it provides interpretable, quantifiable, and reproducible key object identification and ranking using the Node Importance Index (NII). This forms a closed-loop capability from data organization, relationship characterization, representation learning to decision output, achieving a simultaneous improvement in structural representation and decision effectiveness compared to solutions that only perform pixel-level change detection or only model object databases.

[0046] (1) Unifying the expression of spatiotemporal relationships eliminates structural fragmentation and reduces information loss. This invention uses "unified spatiotemporal relationship edges" to carry spatial adjacency and temporal evolution information. It no longer structurally splits edges into different types, but only marks the relationship source, relationship type, and strength in the edge attributes, thereby avoiding cross-graph alignment, duplicate storage, and information loss. The unified graph structure enables spatiotemporal constraints to work together on the same topology, reducing modeling and maintenance costs, and providing a stable and consistent input basis for subsequent graph convolution and dynamic competition.

[0047] (2) Object construction based on spatiotemporal cube segmentation ensures temporal consistency and traceability. By performing cube segmentation in the XYT three-dimensional space and forming GPOs by tracking objects based on voxel adjacency, this invention naturally obtains continuous identity and attribute snapshots of objects in the time dimension, avoiding mismatches and error accumulation caused by two-dimensional cross-period matching. Stable object temporal trajectories make relationship calculation more reliable and feature statistics more consistent, directly improving the credibility of subsequent topology determination and learning inference.

[0048] (3) Complete spatiotemporal topology calculation accurately reveals propagation links and synchronous interactions. By employing 8 types of spatial topology and 13 types of temporal topology and performing composite reasoning, and combining the relationship strength composed of contact length, overlap ratio, temporal overlap rate and interval, it is possible to clearly determine "who interacts with whom at what location, in what order, and whether there is synchronous overlap". This mechanism structurates and quantifies changing propagation paths (such as "adjacent + prior") and co-evolution (such as "intersection + overlap"), providing a direct and interpretable basis for the identification of key nodes and key links.

[0049] (4) Relationship-aware graph modeling and feature engineering enhance representation and analytical capabilities. Through a ternary representation of “node feature matrix X + sparse adjacency matrix A + edge attribute matrix A_attr,” this invention encodes the source, type, strength, and time delay of relationships into learnable signals. Simultaneously, it introduces multimodal features such as geometric, temporal, semantic, neighborhood, and environmental features and performs standardization / one-hot processing. Graph convolution can then be weighted and aggregated according to relationship differences, and dynamic competition can rationally allocate dominance based on embedding similarity and topological position, significantly improving the ability to analyze complex spatiotemporal relationships and fine-grained anomalies.

[0050] (5) The cascaded fusion of "learnable × interpretable" directly produces crucial evidence. On the unified graph, GCN is first used for nonlinear representation of multi-hop neighborhoods, followed by dynamic competitive iteration using the DomiRank concept to obtain interpretable dominance centrality. The two are then fused through normalized weighting to form NII, achieving a balance between performance and interpretability. Unlike traditional methods that only output "whether / what kind of change," NII clearly identifies "which objects are more critical and why," providing managers with traceable ranking results and causal explanations.

[0051] (6) Robust and scalable engineering implementation, adaptable to a wide range of long-term applications. By adopting COO / CSR sparse storage, adjacency sampling and batch training, mixed precision and GPU-accelerated graph convolutional inference, and combining parallelized dynamic competitive iteration and incremental updates with sliding time windows, this invention achieves high throughput and low latency while ensuring accuracy. The design of neighborhood consistency, time-delay coding and three-dimensional continuous objects effectively suppresses spurious changes and noise interference, ensuring stable and reliable output even under conditions of varying data quality from multiple sources and ultra-large graph scales.

[0052] (7) Standardized outputs and strong transferability for decision-making, forming reusable business value. This invention directly connects to the business side with standardized products such as the NII's hierarchical and sorting lists, highlighted maps, and link diagrams, supporting early warning triggering and priority governance. The method has only a weak dependency on the scenario and can be transferred to various tasks such as forest degradation, wetland shrinkage, farmland abandonment, and urban expansion by adjusting category priors and a few parameters. This achieves a closed loop of "discovering changes → identifying key issues → determining priorities → guiding intervention", significantly improving resource allocation efficiency and governance effectiveness. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0054] Figure 1 This is a flowchart of the steps of the method according to an embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of the process of geographic object extraction and process object construction in an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of the GPO-STG mapping process according to an embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the cascade fusion process according to an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram of the system framework of an embodiment of the present invention. Detailed Implementation

[0059] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0060] The objective of this invention is achieved through the following technical solution:

[0061] Example 1

[0062] like Figure 1 As shown, this embodiment provides a remote sensing time-series data analysis method that takes into account the spatiotemporal correlation of geographic objects, including the following steps:

[0063] Step 101: Acquire multi-source, multi-temporal remote sensing images and auxiliary data, perform preprocessing, and generate a consistent time-series data stack.

[0064] It should be noted that in step 101, multi-source, multi-temporal remote sensing images and auxiliary data (such as DEM, land cover products, and vector boundaries) are collected for the study area and analysis period. Radiometric and atmospheric corrections, geometric / orthorectification, and fine registration are uniformly completed. The images are resampled to a consistent spatial resolution and coordinate projection is unified. The images are then cropped to the study area. Low-quality pixels are removed by combining masks such as clouds / shadows / snow / water bodies and quality markers. Necessary temporal gaps are filled using adjacent temporal replacement or removal strategies to maintain statistical consistency. At the same time, timestamps and metadata are standardized, and quality control indicators (such as registration error and cloud cover threshold) are recorded. Finally, a multi-temporal data stack with consistent spatial reference, meeting quality standards, and being traceable is output, along with its quality report, providing reliable input for subsequent three-dimensional spatiotemporal processing.

[0065] Step 102: Divide the image into voxels in the three-dimensional spatiotemporal domain and perform adjacency tracking to identify geographic process objects with evolutionary behavior and record their attributes;

[0066] It should be noted that in step 102, objects are first extracted by constructing a spatiotemporal cube from multiple remote sensing images according to spatial and temporal dimensions. Each cube unit represents the observation value of a specific spatial location within a specific time period. The spatiotemporal cube is segmented based on spectral features, texture features, and temporal change patterns, dividing the cube into multiple object units with consistent characteristics in both space and time. Each object unit is a continuous region in three-dimensional (X, Y, T) space and is accompanied by spatial boundaries and category labels. Then, process objects are constructed. In the cube segmentation results, the continuous state of the same object unit in the temporal dimension is identified and tracked to form geographic process objects (GPOs). The temporal continuity of objects can be directly determined by the voxel adjacency relationship within the cube, avoiding the error accumulation of simple two-dimensional matching. Object units that appear for the first time within a certain time period are marked as "new"; object units that disappear within a certain time period are marked as "terminated". Each GPO carries complete temporal information, including the time of appearance, the time of termination, and attribute snapshots of each time slice (area, category, spectral features, shape parameters, etc.).

[0067] like Figure 2As shown, a spatiotemporal cube segmentation method is demonstrated, which stacks multiple remote sensing images into a three-dimensional spatiotemporal data volume, and uses spectral, texture and temporal features to segment and generate three-dimensional object units. Then, the same object units are tracked along the time axis to form geographic process objects (GPOs).

[0068] During the aggregation process, the currently processed cubes can be merged and expanded in both the spatial and temporal domains. Therefore, as shown in the formula, the heterogeneity criterion H used for segmentation... h It also includes spatial heterogeneity S h With time heterogeneity T h It should be noted that for continuous data (Mode 1) and discrete data (Mode 2), S h With T h The specific calculation formulas differ: H h =(S h ,T h ).

[0069] The first stage (spatial clustering), such as Figure 2 As shown in (a), based on the spatial heterogeneity criterion S h The process involves evaluating and merging spatial neighbors' voxels pixel by pixel to form spatially continuous sub-regions with highly homogeneous internal attributes. For continuous data, spatial heterogeneity is a concern. Defined as a formula:

[0070]

[0071] Where k represents the number of bands, ω k Let n be the weight of the k-th band. cb ΔT represents the number of pixels at the bottom of the merged cube. cb This indicates the time span covered by the cube; This represents the mean and variance of the combined cube in the k-th band, while and These represent two adjacent spacetime cubes (denoted as cb) before the merger. i With cb i+1 The number of pixels, and For their respective time spans, and Let H represent their mean and variance in the k-th band. k Defined as a formula:

[0072]

[0073] In the formula, T represents the number of time nodes contained in the spacetime cube; The standard deviation of the k-th band at the i-th time point is calculated as follows:

[0074]

[0075] Where n is the number of pixels in the spatial domain at time point i, and p j Let j be the observation value of the j-th pixel in this band. This is the average value of these pixels.

[0076] For discrete data, spatial heterogeneity Based on the definition of pixel category similarity, the calculation is defined by the formula:

[0077]

[0078] Among them, H kd This is an indicator of heterogeneity for spatially discrete data, defined as follows:

[0079]

[0080] In the above formula, c p With c q δ(c) represents the category values ​​for pixels p and q, respectively; p ,c q ) is an indicator function, when c p ≠c q It takes the value 1 when the condition is met, and 0 otherwise; SN represents the set of all cell pairs in the spatial neighborhood, and |SN| is the total number of cell pairs.

[0081] The second stage (temporal clustering), such as Figure 2 (b) After completing spatial clustering, the temporal heterogeneity criterion T is adopted. h Clustered sub-regions within a spatiotemporal neighborhood are merged on a time-by-time basis to ultimately form a temporally continuous and self-consistent geographic process object. For continuous data, temporal heterogeneity is a concern. The calculation formula is defined as follows:

[0082]

[0083] Wherein, as in the formula, H t This represents the average variance of continuous data values ​​over the entire time domain of the spacetime cube. The standard deviation of the i-th pixel in the k-th band of the time pixel column is:

[0084]

[0085] For discrete data, temporal heterogeneity The calculation formula is defined as follows:

[0086]

[0087]

[0088] c i and c i+1 Let δ(c) represent the discrete category values ​​for time periods i and i+1, respectively; i ,c j ) is an indicator function, when c i ≠c j The value is 1 if it is a time interval and 0 otherwise; |TN| represents the number of time pixel pairs in the spatiotemporal neighborhood.

[0089] During the spatiotemporal clustering process, the spatial and temporal heterogeneity indicators of each candidate neighbor voxel are examined. Clustering stops when all indicators exceed preset thresholds. At this point, the resulting clustered units are the finally identified geographic process objects, meeting the requirements of spatiotemporal continuity and consistency of evolutionary trends.

[0090] Step 103: Determine the spatial and temporal topological relationships between geographic process objects, generate a unified spatiotemporal relationship table through composite reasoning, and quantify their directionality, correlation strength, and temporal distance.

[0091] It should be noted that in step 103, a spatiotemporal topological relationship model is used to uniformly calculate the relationship between GPOs. By jointly reasoning the topological constraints of the spatial domain and the temporal domain, the spatial relationship, temporal relationship and spatiotemporal composite relationship between objects are obtained.

[0092] 1) Spatial topological relationship calculation

[0093] Within the same time slice, topological determination is performed using the object boundary polygons. The following eight spatial relationship types are determined according to the classic spatial topology model: Equal: The boundaries and interiors of the two objects are completely identical; Contains: Object A contains object B, and B is completely inside A; Within: Object A is completely inside object B; Intersect: Object A and B have partially overlapping areas, but no containment relationship; Touches: Object A and B have contact boundaries but do not overlap interiorly; Disjoint: Object A and B have no common points; Covers: Object A covers object B (boundaries may partially overlap); CoveredBy: Object A is covered by object B (boundaries may partially overlap). The relationship strength can be calculated based on contact length, overlap area ratio, etc., and normalized to [0,1].

[0094] 2) Calculation of temporal topological relationships

[0095] Between consecutive time slices, the temporal relationship is determined based on object ID matching or spatial overlap ratio. Referring to Allen's temporal topological algebra, the following 13 types of temporal relationships are defined: Equal: The two objects are completely identical in time; Before: A is completely earlier than B; After: A is completely later than B; Meets: The end time of A is contiguous with the start time of B; MetBy: The end time of B is contiguous with the start time of A; Overlaps: A and B partially overlap in time, and A starts first; Overlapped (O) (VerlappedBy): B and A partially overlap in time, and B starts first; During: A is entirely within B's time frame; Contains: B is entirely within A's time frame; Starts: A and B start simultaneously, but A ends earlier; StartedBy: B and A start simultaneously, but B ends earlier; Finishes: A and B end simultaneously, but A starts later; FinishedBy: B and A end simultaneously, but B starts later. Relationship strength can be calculated using the time overlap ratio, duration difference, etc.

[0096] 3) Reasoning about spatiotemporal composite relationships

[0097] The spatial and temporal topological determination results are combined to form a spatiotemporal relation triple (Node_i, Relation, Node_j).

[0098] Example of relation type:

[0099] Spatially adjacent + temporally prior: This indicates that a change may propagate from one object to its adjacent objects;

[0100] Spatial intersection + temporal overlap: This indicates that two objects change simultaneously and have spatial intersection.

[0101] The overall weight of the spatiotemporal composite relationship is calculated based on the weighted or multiplicative form of spatial intensity and temporal intensity.

[0102] Step 104: Construct a spatiotemporal graph model of geographic process objects with geographic process objects as nodes and unified spatiotemporal relationships as edges. Extract and standardize the attribute features of nodes and edges to form structured data input for graph learning.

[0103] Specifically, a spatiotemporal graph of geographic process objects is constructed using GPOs as nodes and unified spatiotemporal relationships as weighted (directed / undirected) edges, producing an adjacency matrix A, a node feature matrix X, and an edge feature matrix E. Node features cover location and environment (latitude and longitude / raster index, terrain factors and proximity), morphology (area, perimeter, compactness, texture statistics), evolution (duration, trend and volatility), and semantic category. Edge features include relation type encoding, strength, direction, and time interval. Subsequently, missing data processing and standardization (dimension removal / standardization, one-hot encoding of categories, outlier handling) are performed, and training / validation / inference subsets are segmented according to the application scenario. Serialized graph data objects (including A, X, E and an index dictionary) are output for graph learning and inference.

[0104] like Figure 3 As shown, the general construction process of a geographic process object spatiotemporal map (GPO-STG) consists of four parts: data input, geographic process object identification and feature extraction, spatiotemporal topological relationship calculation, and geographic process object spatiotemporal map generation.

[0105] The spatiotemporal map of geographic process objects is designed with the universality of data sources in mind, capable of accepting and processing inputs from various remote sensing time series data, such as physical attributes (e.g., surface temperature), chemical attributes (e.g., soil pollutant content), biological attributes (e.g., vegetation index), and classification attributes (e.g., land cover type). For a given type of input time series data, it is segmented and recombined using a spatiotemporal cube segmentation method to identify geographic process objects with spatiotemporal continuity and consistent evolutionary characteristics, and extracts multidimensional attributes such as location, morphology, and evolution as the basic features of nodes. Based on the identification of geographic process objects, qualitative relationship categories between geographic process objects are calculated according to the defined temporal and spatial topological types; and the relationships are quantitatively described using indicators such as distance, direction, overlap, and connectivity strength.

[0106] In the graph structure generation stage, each identified geographic process object is mapped to a graph node, and its location, morphology, evolution, and other attributes are recorded to comprehensively describe the characteristics of the geographic process object and provide rich information for subsequent analysis. The interactions between geographic process objects are determined based on spatiotemporal topological relationships. Undirected edges are established for pairs of geographic process objects that satisfy spatiotemporal topological relationships, thereby constructing spatiotemporal associations between nodes. Edge creation is based on the spatiotemporal topological relationships between geographic process objects, and the spatiotemporal topological relationship type and related quantitative indicators are attached to ensure that edge attributes accurately express the interactions between geographic process objects. All node and edge information is organized and stored in an adjacency list or adjacency matrix, forming a complete spatiotemporal graph of geographic process objects.

[0107] Step 105: Utilize graph convolutional networks and topology dynamics mechanisms to jointly model, calculate the importance index of each node, and output hierarchical layers, a list of key nodes, and interface services that can be connected.

[0108] It should be noted that in step 105, graph representation learning and topology mechanism modeling are jointly employed on the graph data object: First, node embeddings are obtained through graph convolution / attention methods to aggregate neighborhood information; second, a topological dynamic process based on relation strength and propagation / competition assumptions is introduced to calculate dominance or influence scores, and the two are weighted and fused to form a node importance index; stability is evaluated using cross-validation and multiple random resampling, and hierarchical classification is achieved according to natural breakpoints or quantiles, generating a list of key nodes (including ID, spatial location, category, adjacency evidence, and detailed indicators) and visualization products (overlay layers, hierarchical graphs, and atlas views). Simultaneously, report export and API / map service interfaces are provided, and version and audit logs are recorded to support business integration for batch processing and incremental updates. Figure 4 As shown, its overall structure can be divided into three stages, forming a cascaded computing link:

[0109] Step 1051: Graph Convolutional Network Embedding

[0110] (1) Input data structure

[0111] The node feature matrix X includes static geometric attributes (area, perimeter, shape index), temporal dynamic features (duration, frequency of change, rate of change of area), semantic category features, neighborhood statistical features (number of adjacent objects, category diversity), and optional external environmental features (such as terrain, climate).

[0112] Adjacency matrix A: Constructed from spatiotemporal relationship edges, reflecting the topological structure of all spatial adjacency, temporal inheritance, and composite relationships.

[0113] The edge attribute matrix A_attr contains information such as relation source (spatial / temporal / composite), relation weight (e.g., contact length normalization value, overlap ratio, temporal overlap rate), and time interval.

[0114] (2) Embedded computation process

[0115] First layer graph convolution:

[0116]

[0117] in (Introducing a self-loop), W is the degree matrix, σ is the nonlinear activation function (such as ReLU), and W is the degree matrix. (0) The first layer weight matrix is ​​used, and the edge weights and attributes can participate in neighbor aggregation as weighting factors during the convolution process.

[0118] Second layer graph convolution:

[0119]

[0120] This layer further aggregates second-order neighborhood features and dynamically adjusts the aggregation weights using edge attributes.

[0121] Output embedding vector: Each node obtains a high-dimensional representation vector z. i It integrates its own attributes with multi-hop neighborhood information, providing structured input for the dynamic competition mechanism.

[0122] Step 1052: Dynamic Competition Mechanism

[0123] Based on graph embedding, a dynamic competition mechanism based on the DomiRank centrality concept is introduced, so that the importance of a node is not only determined by its static structure and features, but also reflects its dominance and influence in the network, thus providing interpretability.

[0124] Calculation steps:

[0125] Initialize the dominance vector: Set the initial dominance value of all nodes to a constant p0;

[0126] Neighborhood competition propagation: In each iteration, node i acquires dominance from its neighbor node j, and the proportion acquired is determined by the following factors:

[0127] Boundary weight (strength of spatial / temporal / composite relationship);

[0128] Node embedding similarity (cosine similarity or inverse Euclidean distance);

[0129] Node degree normalization factor (to prevent highly connected nodes from excessively absorbing dominance);

[0130] Dynamic parameter adjustment: During the propagation process, balance coefficients α (local neighborhood dominance weight) and γ (global topological influence weight) are introduced and optimized based on the training or validation set;

[0131] Iterative convergence: Repeat propagation until the change in the overall dominance force Δp < ∈ or the maximum number of iterations T is reached. max ;

[0132] Output Dominance Score_DomiRank(i): Reflects the degree of dominance and propagation ability of a node in the network.

[0133] Step 3: Cascaded Fusion and Node Importance Index Calculation

[0134] (1) Fusion Formula

[0135] Integrate the interpretable dominance score with the GCN prediction score:

[0136]

[0137] in: and The score is normalized; β∈[0,1] is determined by cross-validation or robustness analysis to balance interpretability and feature learning ability.

[0138] (2) Sorting and Grading

[0139] Nodes are sorted in descending order of NII; they are classified into Level I / II / III key nodes based on percentile thresholds (e.g., Top 5%, 5-20%, 20%+); a list of key nodes and visual annotations (including node location, category, change trajectory, and scope of influence) are generated.

[0140] (3) Output

[0141] Highlight nodes with high NII on the map, overlay their evolution trajectory and important neighborhood relationships; list the influencing factors (DomiRank contribution, GCN feature importance) of each key node, and their potential influence range.

[0142] Example 2

[0143] like Figure 5 As shown, this embodiment provides a remote sensing time-series data analysis system that takes into account the spatiotemporal correlation of geographic objects, including the following functional modules:

[0144] The data acquisition and preprocessing module is used to acquire raw data from multi-source, multi-temporal remote sensing data (including high-resolution optical imagery and fused land cover products) and perform quality control. Specifically, this module performs radiometric and atmospheric correction, geometric correction and fine registration, projection unification and range clipping, and performs masking processing on invalid pixels such as clouds / shadows to output a time-series dataset with consistent spatial resolution and coordinate system.

[0145] The Geographic Process Object (GPO) construction module is used to construct spatiotemporal cubes from multiple remote sensing images according to the spatial-temporal dimensions and perform three-dimensional segmentation to obtain continuous and consistent object units in the X–Y–T three dimensions. Specifically, the object units are tracked along the time axis through voxel adjacency relationships to form Geographic Process Objects (GPOs) and to record the generation, termination, and attribute snapshots of each time slice.

[0146] The spatiotemporal relationship calculation module is used to determine the type and strength of the relationship between objects based on the spatiotemporal topological relationship model. Specifically, the module determines 8 types of spatial topological relationships within the same time slice and 13 types of temporal topological relationships between consecutive time slices. It then performs composite reasoning on the two, calculates the relationship strength, and outputs a unified spatiotemporal relationship table.

[0147] The GPO-STG modeling module is used to construct a unified spatiotemporal relationship network (GPO-STG) with GPOs as nodes and spatiotemporal relationships as edges. Specifically, it generates an adjacency matrix A, a node feature matrix X, and an edge attribute matrix A_attr, which uses sparse storage and supports batch processing and incremental updates.

[0148] The feature extraction and attribute enhancement module is used to calculate and enhance the geometric, temporal, semantic, neighborhood, and environmental features of nodes and edges. Specifically, it normalizes continuous features and performs one-hot encoding and uniformization on categorical features to form a numerical feature matrix that can be directly used by graph neural networks.

[0149] The cascaded analysis module of graph convolutional network embedding and dynamic competition mechanism is used to perform node importance analysis on a unified spatiotemporal relationship graph. Specifically, firstly, node features and neighborhood information are convolved and aggregated through multi-layer GCN to obtain node embedding vectors. Then, based on the embedding vectors and topology, the DomiRank dynamic competition iteration is performed to calculate the dominance centrality score. Finally, the GCN output and the DomiRank score are weighted and fused to generate the node importance index (NII) and complete the ranking and classification.

[0150] The results output and visualization module is used to highlight key nodes in a map overlay format and display their evolution trajectory and important neighborhood relationships; the module outputs a node list, NII level and evaluation index report, and provides interface integration capabilities with business systems.

[0151] The experimental environment for this invention uses an NVIDIA-A800 graphics card, the software environment is based on Python 3.8, the deep learning framework uses PyTorch 1.10.0, the graph neural network library uses PyTorch Geometric 2.0.0, and other dependent libraries include NumPy, Pandas, Matplotlib, NetworkX, Scikit-learn, and SciPy.

[0152] The hyperparameters of the model are set as follows: the input feature dimension of the GCN model is set to 4, which represents the initial number of features of the nodes, corresponding to the dimensions of node attributes (volume, surface area, centroid, duration). The hidden layer dimension is set to 64 to improve the network's ability to abstractly represent graph structure and attribute features. The output dimension is set to 32, which serves as the dimension for the final node embedding and is used in the subsequent DomiRank dynamic competition mechanism. The number of training epochs is set to 200. The optimizer is Adam, and the learning rate is set to 0.01 to balance optimization efficiency and stability. The parameter σ of the DomiRank algorithm is set to 0.5 to control the strength of the competition mechanism. The parameter θ is set to 1.0 to calculate the competition state between nodes. The convergence threshold is set to 1e-6 to determine whether the algorithm has converged.

[0153] The selected node attributes include the volume, surface area, spatiotemporal centroid coordinates, and duration of the land cover change process object. Edge attributes include the contact area between change processes, their temporal sequence or coexistence duration, and their geometric distance in three-dimensional spatiotemporal space.

[0154] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A remote sensing time-series data analysis method that takes into account the spatiotemporal correlation of geographic objects, characterized in that, include: Step 1: Acquire multi-source, multi-temporal remote sensing images and auxiliary data, and preprocess them to generate a consistent time-series data stack; Step 2: Divide the image into voxels in the three-dimensional spatiotemporal domain and perform adjacency tracking to identify geographic process objects with evolutionary behavior and record their attributes; Step 2 specifically includes: Multiple remote sensing images are used to construct a spatiotemporal cube according to spatial and temporal dimensions, with each cube unit representing the observation value of a specific spatial location in a specific time period; The spatiotemporal cube is segmented based on spectral features, texture features, and temporal change patterns. The cube is divided into multiple object units with consistent features in both space and time. Each object unit is a continuous region in three-dimensional space and is accompanied by spatial boundaries and category labels. In the cube segmentation results, the continuous state of the same object unit in the time dimension is identified and tracked to form geographic process objects. The temporal continuity of the objects is directly determined by the adjacency relationship of voxels within the cube. For an object unit that appears for the first time within a certain time period, it is marked as newly created; for an object unit that disappears within a certain time period, it is marked as terminated. Each geographic process object carries complete temporal information, including the time of appearance, the time of termination, and attribute snapshots of each time slice. Step 3: Determine the spatial and temporal topological relationships between geographic process objects, generate a unified spatiotemporal relationship table through composite reasoning, and quantify their directionality, correlation strength, and temporal distance; Step 4: Construct a spatiotemporal graph model of geographic process objects with geographic process objects as nodes and unified spatiotemporal relationships as edges. Extract and standardize the attribute features of nodes and edges to form structured data input for graph learning. Step 5: Utilize graph convolutional networks and topology dynamics mechanisms to jointly model, calculate the importance index of each node, and output a hierarchical layer, a list of key nodes, and interfacing interface services. Step 5 specifically includes: The node embedding vector is obtained by convolutional aggregation of node features and neighborhood information through multi-layer GCN. DomiRank dynamic competitive iteration is performed based on embedding vectors and topological structure to calculate dominance centrality score; The GCN output is weighted and fused with the DomiRank score to generate a node importance index, and then sorted and ranked.

2. The remote sensing time-series data analysis method considering the spatiotemporal correlation of geographic objects according to claim 1, characterized in that, In step 1, the preprocessing includes: Differences in imaging conditions are eliminated through radiometric and atmospheric corrections; Geometric correction and registration ensure accurate spatial alignment of data from multiple periods; By unifying projection and clipping the range, a time-series data stack with consistent spatial resolution and coordinate system is generated. Input quality is ensured through masking.

3. The remote sensing time-series data analysis method considering the spatiotemporal correlation of geographic objects according to claim 1, characterized in that, In step 3, a spatiotemporal topological relationship model is used to uniformly calculate the relationship between geographic process objects. Eight types of spatial topological relationships are determined within the same time slice, and 13 types of temporal topological relationships are determined between consecutive time slices. The two are then combined for inference, the relationship strength is calculated, and a unified spatiotemporal relationship table is output.

4. The remote sensing time-series data analysis method considering the spatiotemporal correlation of geographic objects according to claim 1, characterized in that, Step 4 specifically includes: A spatiotemporal graph of geographic process objects is constructed using geographic process objects as nodes and unified spatiotemporal relationships as weighted edges. This graph includes an adjacency matrix, a node feature matrix, and an edge feature matrix. Missing data is then handled and standardized. Training, validation, and inference subsets are then segmented according to the application scenario. Serialized graph data objects are output for graph learning and inference. Node features cover location and environment, morphology, evolution, and semantic categories, while edge features include relation type encoding, strength, direction, and time interval.

5. A remote sensing time-series data analysis system that considers the spatiotemporal correlation of geographic objects, used to execute the remote sensing time-series data analysis method considering the spatiotemporal correlation of geographic objects as described in any one of claims 1-4, characterized in that, include: The data acquisition and preprocessing module is used to acquire raw data from multi-source and multi-temporal remote sensing data and perform quality control. The geographic process object construction module is used to construct spatiotemporal cubes from multiple remote sensing images according to the spatial-temporal dimensions and perform three-dimensional segmentation to obtain continuous and consistent object units in three-dimensional space. The spatiotemporal relationship calculation module is used to determine the type and intensity of relationships between objects based on the spatiotemporal topological relationship model. The geographic process object spatiotemporal graph modeling module is used to construct a unified spatiotemporal relationship network with geographic process objects as nodes and spatiotemporal relationships as edges. The feature extraction and attribute enhancement module is used to calculate and enhance the geometric, temporal, semantic, neighborhood, and environmental features of nodes and edges; The graph convolutional network embedding and dynamic competition mechanism cascade analysis module is used to perform node importance analysis on a unified spatiotemporal relationship graph; The results output and visualization module is used to highlight key nodes in a map overlay format and display their evolution trajectory and important neighborhood relationships; the module outputs a node list, node importance index level and evaluation indicator report, and provides interface integration capabilities with business systems.

6. An electronic device, characterized in that, It includes a processor and a memory; the processor executes the steps of the remote sensing time-series data analysis method that takes into account the spatiotemporal correlation of geographic objects as described in any one of claims 1-4 by calling programs or instructions stored in the memory.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program or instructions that cause a computer to perform the steps of the remote sensing time-series data analysis method that takes into account the spatiotemporal correlation of geographic objects as described in any one of claims 1-4.

Images