Multi-source heterogeneous data fusion remote sensing map dynamic database construction method and system

Abstract

The invention belongs to the technical field of data processing, and discloses a multi-source heterogeneous data fused remote sensing map dynamic database construction method and system, and the method comprises the steps: firstly obtaining multi-source heterogeneous remote sensing data of a target area, extracting remote sensing time sequence data of a land parcel after preprocessing, and obtaining a remote sensing time sequence; constructing a time sequence semantic track containing a key semantic feature time sequence for each land parcel; identifying a semantic evolution trend by using a semantic drift detection model, and generating an abandoned suspicious score; the static factors, the dynamic factors and the prior knowledge of the land parcels are fused, a structure causal graph is constructed by means of a causal discovery algorithm, and abandoned causal credibility is reasoned based on a graph neural network model; and automatically generating abandoned land labels for the land parcels of which the double scores exceed the threshold value, and finally writing the land parcel information into a dynamic database. And the problems of construction lag and information incompleteness of the dynamic database are solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method, system, electronic device, medium, and program product for constructing a dynamic database of remote sensing maps by fusing multi-source heterogeneous data. Background Technology

[0002] Due to the fragmented landforms, complex vegetation dynamics, and high semantic similarity to grasslands and fallow land, the interpretation of abandoned land in remote sensing images has long faced challenges such as insufficient recognition accuracy, spatiotemporal discontinuity, and unstable boundaries. Furthermore, due to the lack of labeled data, traditional supervised models are difficult to train effectively, resulting in a lag in the construction of dynamic databases and incomplete information. Summary of the Invention

[0003] The purpose of this invention is to address the above-mentioned problems by providing a method, system, electronic device, medium, and program product for constructing a dynamic database of remote sensing maps based on the fusion of multi-source heterogeneous data, thereby at least partially solving the aforementioned problems.

[0004] According to a first aspect of the present invention, a method for constructing a dynamic database of remote sensing maps based on multi-source heterogeneous data fusion is provided, comprising: acquiring multi-source heterogeneous remote sensing data of a target area; preprocessing the remote sensing data to extract land parcel remote sensing time series data; constructing a temporal semantic trajectory for each land parcel, the temporal semantic trajectory including a sequential representation of key semantic features of land parcel attributes in the time dimension; based on the temporal semantic trajectory, identifying semantic evolution trends using a semantic drift detection model and generating a land abandonment suspicion score; and based on land parcel static factors, dynamic factors, and prior knowledge, using a causal discovery algorithm... A structural causal graph is constructed, and based on the structural causal graph, a graph neural network model is used to infer the causal credibility of land parcel abandonment. The static factors reflect the inherent attributes of the land parcel and the long-term stable geographical conditions, while the dynamic factors are feature change indicators extracted based on temporal semantic trajectories. The prior knowledge is auxiliary information from external data sources. For land parcels whose abandonment suspicion score and abandonment causal credibility both exceed a set threshold, an abandonment label is automatically generated. The land parcel information is written into a dynamic database, and the land parcel information includes land parcel identifier, abandonment label, semantic trajectory, discrimination source, scoring result, or timestamp.

[0005] According to a second aspect of the present invention, a system for constructing a dynamic database of remote sensing maps by fusing multi-source heterogeneous data is provided, comprising: a first acquisition module, configured to acquire multi-source heterogeneous remote sensing data of a target area, and preprocess the remote sensing data to extract remote sensing time series data of land parcels;

[0006] The first construction module is used to construct a temporal semantic trajectory for each of the land parcels, wherein the temporal semantic trajectory includes a sequential representation of the key semantic features of the land parcel attributes in the time dimension;

[0007] The first generation module is used to identify semantic evolution trends based on the temporal semantic trajectory using a semantic drift detection model, and generate a suspected abandonment score.

[0008] The first reasoning module is used to construct a structural causal graph based on static factors, dynamic factors and prior knowledge of the land parcel through a causal discovery algorithm, and to infer the causal credibility of the land parcel's abandonment through a graph neural network model based on the structural causal graph. The static factors reflect the inherent attributes of the land parcel and its long-term stable geographical conditions, the dynamic factors are feature change indicators extracted based on time-series semantic trajectories, and the prior knowledge is auxiliary information from external data sources.

[0009] The tag generation module is used to automatically generate abandonment tags for plots of land whose abandonment suspicion score and abandonment causal credibility both exceed the set thresholds.

[0010] The writing module is used to write land parcel information into a dynamic database. The land parcel information includes land parcel identifier, abandoned land tag, semantic trajectory, identification source, scoring result or timestamp.

[0011] According to a second aspect of the invention, a computing device is provided, the computing device comprising: at least one processing unit; at least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions, when executed by the at least one processing unit, enabling the computing device to perform steps of the method according to the first aspect.

[0012] According to a third aspect of the present invention, a computer-readable storage medium is also provided. The computer-readable storage medium stores a computer program that, when executed by a machine, performs the method of the first aspect of the present invention.

[0013] According to a fourth aspect of the present invention, a computer program product is also provided, comprising a computer program that, when executed by a machine, performs the method of the first aspect of the present invention.

[0014] Therefore, this invention proposes to effectively identify the abandonment trend and causal relationship of land parcels by introducing a semantic drift detection model and a graph neural network inference mechanism, breaking through the limitations of traditional models that rely heavily on supervised data and lack dynamism. By combining the static attributes of land parcels, temporal change characteristics, and external prior knowledge to construct a structural causal graph, it realizes in-depth mining of complex abandonment causes and automatic generation of credible labels. The final dynamic database has the advantages of strong timeliness, comprehensive information, and high label interpretability, which significantly improves the monitoring and management of abandoned land in remote sensing maps.

[0015] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for constructing a dynamic database for remote sensing maps by fusing multi-source heterogeneous data according to an embodiment of the present invention is shown.

[0017] Figure 2 A schematic diagram of the semantic drift detection model structure according to an embodiment of the present invention is shown.

[0018] Figure 3 A schematic diagram of a remote sensing map dynamic database construction system for multi-source heterogeneous data fusion according to an embodiment of the present invention is shown;

[0019] Figure 4 A schematic diagram of an electronic device structure according to an embodiment of the present invention is shown. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0021] The scope of the embodiments described herein includes the entire scope of the claims and all available equivalents thereof. Throughout this document, the terms “first,” “second,” etc., are used only to distinguish one element from another, without requiring or implying any actual relationship or order between the elements. The terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion and do not exclude the presence of additional identical elements in the structure, apparatus, or device that includes the stated element. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0022] Figure 1 A flowchart illustrating a method for constructing a dynamic database of remote sensing maps by fusing multi-source heterogeneous data according to an embodiment of the present invention is shown. It should be understood that the method may also include additional steps not shown and / or the steps shown may be omitted, and the scope of the invention is not limited in this respect.

[0023] In step 101, multi-source heterogeneous remote sensing data of the target area is acquired, and the remote sensing data is preprocessed to extract the remote sensing time series data of the plots.

[0024] Regarding multi-source remote sensing data, specifically multi-temporal remote sensing image sequences, examples include: optical data, such as Sentinel-2 satellite images, which provide high-resolution surface reflectance information for extracting vegetation indices (such as NDVI and EVI) and other features closely related to vegetation growth status; SAR (Synthetic Aperture Radar) data, such as Sentinel-1 satellite images, which have the characteristics of penetrating clouds and being unaffected by illumination, and can provide features such as VV / VH polarization ratio and backscattering values, reflecting changes in surface roughness and structure; and thermal infrared data, such as MODISLST (surface temperature data), which can capture surface temperature information and assist in analyzing changes in the thermal characteristics of land parcels.

[0025] Data preprocessing includes, for example, radiometric calibration: converting raw digital quantization (DN) values ​​acquired by sensors into physically meaningful radiance or reflectance values ​​to eliminate radiometric biases from different sensors or from the same sensor observed at different times; atmospheric correction: removing the influence of atmospheric scattering, absorption, and other factors on optical images to obtain the true surface reflectance and improve the comparability of images from different time periods; geometric registration: using a reference image to adjust the spatial positions of other images so that all images are aligned in the same geographic coordinate system, ensuring the accuracy of plot locations; and resampling: unifying images of different resolutions to the same spatial resolution to ensure the consistency of multimodal data at the spatial scale, while using interpolation and other methods to handle missing or inconsistent data in time series, achieving temporal consistency.

[0026] Regarding the extraction of land parcel-level remote sensing time series data, for example, based on existing land parcel outline vector data (such as land use patches or administrative boundary lines), the preprocessed remote sensing image is divided into independent spatial units (i.e., land parcels). Then, for each spatial unit, the corresponding pixel information is extracted from the multi-temporal remote sensing image sequence to form a land parcel-level remote sensing image sequence.

[0027] In step 102, a temporal semantic trajectory is constructed for each land parcel, the temporal semantic trajectory including the sequential representation of multimodal remote sensing features in the time dimension.

[0028] Regarding the construction of temporal semantic trajectories, for example, for each independent spatial unit (plot) divided in step 101, key semantic features that can characterize the plot attributes are extracted one by one from its corresponding multi-temporal remote sensing image sequence, including but not limited to four major categories of core features:

[0029] Vegetation indices: Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) are calculated from optical remote sensing images (such as Sentinel-2) to quantify vegetation cover and growth vigor. Their numerical changes can directly reflect crop growth status or vegetation degradation caused by abandonment.

[0030] Texture features: Based on the gray-level distribution of optical or SAR images, extract entropy (characterizing the complexity of image information), variance (reflecting the degree of pixel value dispersion), and gray-level co-occurrence matrix (GLCM) parameters (such as contrast, correlation, etc., describing the spatial distribution law of pixels) to capture the structural changes of the land surface. For example, when cultivated land is transformed into abandoned land, the texture will evolve from a regular farmland structure to a rough texture covered with weeds.

[0031] Thermal infrared characteristics: Surface temperature is extracted from thermal infrared remote sensing data (such as MODISLST) as an indicator reflecting the energy exchange and surface condition of the land. Abandoned land may exhibit different temperature change patterns than cultivated land due to reduced vegetation cover.

[0032] SAR intensity characteristics: Calculate the VV / VH polarization ratio (the ratio of backscatter intensity under different polarization modes) and backscatter values ​​from SAR images (such as Sentinel-1). Utilize the sensitivity of SAR to surface roughness and humidity to capture changes in surface structure caused by the cessation of farming activities (such as increased backscatter due to increased soil exposure).

[0033] In some embodiments, the extracted single-temporal features are organized in chronological order to form a unique temporal semantic trajectory (TST) for each plot. Specifically, the TST for each plot is a set of high-dimensional feature vectors arranged along a time axis, where each time node corresponds to a set of synthesized multimodal semantic features (including vegetation index, texture features, thermal infrared features, SAR intensity features, etc.).

[0034] In step 103, based on the temporal semantic trajectory, the semantic drift detection model is used to identify the semantic evolution trend and generate a suspected abandonment score.

[0035] In some embodiments, the temporal semantic trajectory (TST) of each land parcel constructed in step 102 is input into the semantic drift detection model. This model adopts a structure based on a temporal attention mechanism, which can focus on key time nodes and feature changes related to abandonment in the TST, and analyze the semantic feature evolution of the land parcel.

[0036] Regarding semantic drift detection models, such as Figure 2 As shown, it includes, for example, an input layer, a modality adaptation embedding layer, a basic temporal feature extraction layer, a temporal attention layer, a multimodal consistency verification layer, a temporal trend and periodic analysis layer, and a score fusion output layer.

[0037] Specifically, the input layer is the TST feature access layer, which receives the temporal semantic trajectory (TST) data generated in step 102. TST is a high-dimensional temporal vector, and each time step contains multimodal features (such as NDVI, EVI, SAR backscatter value, texture entropy value, surface temperature, etc.), and the feature dimension of each time step is d (composed of multimodal features).

[0038] Regarding modality adaptation embedding layers, they can be used, for example, to unify the scale and feature space of multimodal features, eliminating the heterogeneity of different types of features (such as vegetation indices of [-1,1] and SAR backscattering as absolute values). As an example, a small fully connected network (FC layer) is designed separately for each modality feature to map different modal features such as optical, SAR, thermal infrared, and texture to an embedding space of the same dimension, outputting the temporal embedding features of each modality.

[0039] Regarding the basic temporal feature extraction layer (LSTM layer), it captures, for example, the long-term temporal dependencies implicit in TST (such as the continuous downward trend of NDVI and the periodic basic pattern of crop growth), providing temporal context for the attention mechanism. As an example, a bidirectional LSTM network (Bi-LSTM) is used, taking the modality-adapted fused temporal features as input (concatenating the embedded features of each modality) and outputting the hidden state H(t) at each time step (containing forward and backward temporal information), focusing on extracting unbiased basic temporal patterns, such as the seasonal NDVI fluctuation pattern of normally cultivated plots.

[0040] Regarding the temporal attention layer, it focuses, for example, on key time points related to the semantic drift of abandonment (such as the time when NDVI first significantly decreases). k t m This assigns higher weights to these time steps and weakens irrelevant time steps. As an example, based on prior knowledge of abandoned features (such as "NDVI decrease" and "SAR reflection enhancement"), a learnable query vector Q is constructed to match the hidden state H(t) output by the LSTM. Attention weights are then calculated through... Obtain the weight α for each time step t (where, ∑α) t =1), the higher the weight, the stronger the correlation between the time step and the abandonment drift pattern.

[0041] Regarding the multimodal consistency verification layer, it can detect the logical consistency of different modal features within the same time step and identify anomalous contradictory signals (such as high NDVI in optical displays but strong backscattering in SAR). As an example, for the key time step (t corresponding to high weights) output by the attention layer, its modal embedding features are extracted. The cosine similarity between modes is calculated, and the intensity of the contradictory signal is obtained if the similarity is related to a preset threshold (such as 0.3).

[0042] Regarding the time series trend and cycle analysis layer, for example, trend analysis could be the attention-weighted time series features ∑α t H(t) extracts slope features (such as the linear fitting slope of the NDVI sequence) through a 1D convolutional layer (kernel size set to 3-5 to capture local continuous trends) to determine whether there are trends such as continuous decline and enhanced reflection.

[0043] Specifically, a 1D convolutional layer with a kernel size of 3-5 is used to perform sliding window operations on the weighted features. For example, with a kernel size of 3, each sliding operation covers features from three consecutive time phases, extracting local continuous change patterns (such as the NDVI value change rate of "t1→t2→t3"). The convolution process strengthens the correlation between adjacent time phases, highlighting local trend signals such as continuous decline and sustained enhancement, and suppressing isolated short-term fluctuations (such as accidental changes caused by a single extreme weather event). The local trend features output by the convolutional layer are linearly fitted, and the slope value is calculated (e.g., a negative and large absolute slope for the NDVI sequence indicates a continuous decline; a positive and significant slope for the SAR backscatter sequence indicates enhanced reflection). By using a preset threshold (e.g., a slope <-0.05 is considered a significant decline), the strength and direction of the trend are quantified, and the "trend degradation degree" (range 0-1, the higher the value, the more obvious the abandonment-related trend) is output.

[0044] For example, periodic analysis can be performed by using a Fourier transform layer to model the periodic characteristics of TST (such as the annual cycle of crop growth), compare the differences between the key time intervals where attention is focused and historical cycles, and output the "period disruption degree" (such as the percentage decrease in signal strength during normal cycles).

[0045] In some embodiments, the periodic signal strength is used to describe the “significance” or “completeness” of a periodic pattern (such as the annual “sow-grow-harvest” cycle) in the TST. For normally cultivated plots, their multimodal characteristics (such as NDVI and SAR backscatter) exhibit stable periodic fluctuations with agricultural activities (e.g., NDVI peaks annually from June to August due to crop growth), at which time the periodic signal strength is high; however, when plots show a tendency to be abandoned, cultivation activities cease, the original periodic pattern is disrupted, and the periodic signal strength decreases significantly (e.g., NDVI remains at a low level throughout the year without peak fluctuations).

[0046] As an example, taking the annual cycle as an example, the calculation of the periodic signal strength includes transforming the time-domain characteristics of the original TST (such as the NDVI time series) to the frequency domain through a Fourier transform layer, and identifying the dominant periodic component (such as the frequency component corresponding to the 12-month annual cycle). The amplitude of the dominant periodic component in the frequency domain is normalized (range 0-1). The higher the amplitude, the more significant the periodic pattern is in the TST, that is, the higher the periodic signal strength.

[0047] Extract the high-weight time intervals focused by the temporal attention layer (such as three consecutive quarters of abnormal NDVI decline), and compare the periodic characteristics of this interval with historical periods (such as the periodic patterns of the same period in the previous three years): if the periodic signal strength of the current interval decreases by more than 50% compared with the historical period (such as from 0.8 to 0.3), and / or the periodic phase shifts significantly (such as the peak crop growing season being delayed from June-August to September-November without a reasonable agricultural explanation), it is judged as "abnormal periodic pattern"; based on the combined indicators of the signal strength decline ratio and the phase shift amplitude, the "period disruption degree" (range 0-1) is output.

[0048] Specifically, based on historical TST data of the land parcel (such as records from 3-5 years prior to the emergence of the abandonment trend), periodic characteristics under normal cultivation conditions are extracted as a baseline period, including:

[0049] Signal strength benchmark: The average signal strength of the dominant cycle (e.g., the annual crop growth cycle) (e.g., the annual NDVI cycle signal strength benchmark value is 0.8);

[0050] Phase reference: the time point of the cycle peak / trough (e.g., July is the peak period of NDVI each year, corresponding to the peak crop growth season);

[0051] Calculate the actual signal strength of the dominant period within the critical interval (e.g., 0.2) and compare it with the signal strength of the reference period (e.g., 0.8) to obtain the decrease ratio: (reference strength - actual strength) / reference strength = (0.8 - 0.2) / 0.8 = 0.75, that is, the signal strength decreases by 75%.

[0052] Compare the actual time points of the peak / trough values ​​within the key interval with the reference phase (e.g., the reference peak is in July, and the actual peak is in October), calculate the offset duration (e.g., 3 months), and normalize it to the offset magnitude: offset duration / total cycle length (e.g., 3 months / 12 months = 0.25).

[0053] By weighting and integrating the two indicators mentioned above, the final periodicity is generated (range 0-1). The signal strength decrease ratio has a higher weight (e.g., 0.8) because it directly reflects the overall attenuation of the periodic signal; the phase shift amplitude has a lower weight (e.g., 0.2).

[0054] In some embodiments, the input features of the scoring fusion output layer include, for example, the degree of degradation of trend analysis, the degree of disruption of periodic analysis, and the intensity of contradictory signals in multimodal consistency verification. The above input features are standardized (ensuring that they are all in the range of [0,1]) to eliminate dimensional differences, and then weighted and fused to finally output the abandonment suspicion score (ASS) value, which is in the range of [0,1]. The higher the value, the more significant the abandonment trend of the land parcel.

[0055] Step 104: Based on the static factors, dynamic factors, and prior knowledge of the land parcel, construct a structural causal graph using a causal discovery algorithm. Based on the structural causal graph, infer the causal credibility score of the land parcel's abandonment using a graph neural network model. The static factors reflect the inherent attributes and long-term stable geographical conditions of the land parcel, the dynamic factors are feature change indicators extracted based on temporal semantic trajectories, and the prior knowledge is auxiliary information from external data sources.

[0056] In some embodiments, multidimensional variables that may be related to land abandonment include static factors: reflecting the inherent attributes of the land and long-term stable geographical conditions, including topographic slope (land with steep slopes is difficult to cultivate and is prone to abandonment), transportation accessibility (the farther away from the main transportation line, the lower the convenience of cultivation), arable land grade (land with poor inherent conditions such as soil fertility is prone to abandonment), distance from water source (insufficient irrigation water will reduce the feasibility of cultivation), and irrigation conditions (the degree of perfection of water conservancy facilities directly affects the sustainability of cultivation).

[0057] Dynamic factors: Based on the feature change indicators extracted from TST constructed in step 102, including the vegetation and other feature decline trends in TST (such as the continuous decline rate of NDVI), texture change rate (the speed of transformation from regular farmland texture to weed texture), and thermal anomaly frequency (the number of times the surface temperature deviates from the normal cultivation state). These dynamic changes directly reflect the evolution of the plot from cultivation to abandonment.

[0058] External priors: Auxiliary information from external data sources, including policy support scope (plots not included in agricultural subsidies or land consolidation policies are prone to abandonment), historical agricultural plots (past cultivation records can be used as a reference to determine whether the current state deviates from cultivation status), and agricultural machinery operation tracks (lack of agricultural machinery operation traces suggests that cultivation activities have stopped).

[0059] In some embodiments, the Structural Causal Graph (SCG) uses "abandonment" as the target node and static factors, dynamic factors, and external priors as input nodes. Through statistical association analysis and geographical causal logic reasoning (such as the chain relationship of "steep slope → difficult irrigation → high farming cost → abandonment"), it identifies the direct or indirect causal paths between each variable and the abandonment result.

[0060] Specifically, it includes the following steps:

[0061] I. Construction of Structural Cause-Effect Graph (SCG)

[0062] 1. Node definition and variable standardization

[0063] Define the node set of SCG: including the target node "abandoned" (a binary variable, 1 represents abandonment, 0 represents cultivation), and the input nodes (static factors, dynamic factors, external priors).

[0064] Static factor nodes: terrain slope, accessibility, farmland grade, distance from water source, irrigation conditions (all quantified as continuous or ordered discrete values, such as slope in degrees and accessibility in kilometers from the main road);

[0065] Dynamic factor nodes: TST decay trend (e.g., annual NDVI decline rate), texture change rate, thermal anomaly frequency (average number of anomalies per year);

[0066] External prior nodes: policy support scope (binary variable, 1 indicates the policy coverage area), historical agricultural plots (number of times cultivated in the past 5 years), and agricultural machinery operation trajectory (average number of days of operation per year).

[0067] Standardize all node variables (e.g., normalize to the [0,1] interval) to eliminate dimensional differences and ensure consistency in causal association analysis.

[0068] 2. Causal Relationship Direction and Connection Construction

[0069] Based on geographical knowledge, the initial causal connection is presupposed: According to the logic of agricultural production, static factors usually serve as "basic constraints" to influence dynamic factors (e.g., "poor irrigation conditions → significant TST decline trend"), while dynamic factors directly reflect changes in the state of the plot (e.g., "TST decline trend → abandonment"). External a priori factors moderate the influence intensity of static / dynamic factors (e.g., "policy support scope → weakening the impact of terrain slope on abandonment").

[0070] Optimize connections by combining causal discovery algorithms (such as PC algorithm): By analyzing the conditional independence between variables (such as "after controlling irrigation conditions, does the correlation between distance from water source and TST decline trend disappear?"), eliminate spurious associations (such as only statistically related but not causal variable pairs), and correct the causal direction (such as confirming "poor accessibility → reduced farming activities → faster texture change rate" rather than the reverse).

[0071] In the final SCG, the directed edges (→) between nodes represent the cause-effect relationship, and the weight of the edge is tentatively set as the initial association strength (such as the absolute value of the correlation coefficient).

[0072] II. Causal Path Identification

[0073] Causal path identification involves filtering out chain-like causal relationships (such as "static factor → dynamic factor → abandonment", "external prior → static factor → dynamic factor → abandonment", etc.) that have a significant impact on the "abandonment" target node in SCG. These include:

[0074] 1. Path traversal and candidate path generation

[0075] Using "abandoned" as the endpoint, traverse all directed paths in the SCG pointing to the endpoint to generate candidate causal paths. For example:

[0076] Direct path: Dynamic factors (such as the declining trend of TST) → abandonment;

[0077] Indirect Path 1: Static Factor (steep terrain slope) → Dynamic Factor (rapid texture change rate) → Abandonment;

[0078] Indirect Path 2: External Priorities (outside the scope of policy support) → Static Factors (poor irrigation conditions) → Dynamic Factors (high frequency of thermal anomalies) → Abandonment.

[0079] 2. Path strength quantification and critical path selection

[0080] Calculate the "causal strength" of each candidate path: the path strength is the product of the weights of all edges on the path (reflecting the cumulative effect of chain influence), and is corrected by conditional probability (such as "the probability that the texture change rate is fast when the terrain slope is >15°"), and finally obtain the comprehensive strength value of the path.

[0081] Key paths are selected by setting intensity thresholds: Paths with intensity values ​​higher than the threshold are retained, as these paths are considered to have a strong causal relationship with the formation of land abandonment. For example, the key paths for plot number "GZ-101-23" are: "12° terrain slope (static factor) → poor irrigation conditions (static factor) → TST decline trend (dynamic factor) → land abandonment", and "outside the scope of policy support (external prior) → low transportation accessibility (static factor) → lack of agricultural machinery tillage trajectory (external prior) → land abandonment".

[0082] 3. Calculate path contribution

[0083] Path contribution is calculated based on a graph neural network (GNN) model. As an example, GNNs include:

[0084] Input layer: Receives standardized data from three types of causal variable nodes;

[0085] Graph convolutional layers: Based on the edge weights of the SCG, convolutional operations are performed on node features to aggregate the influence of neighboring nodes (causally related variables) (e.g., the "irrigation conditions" node aggregates features of "distance from water source" and "terrain slope"). Specifically, the graph convolutional layer performs convolutional operations on the original features of each node (e.g., "irrigation conditions", "terrain slope", "TST decline trend") based on the edge weights of the SCG. During the operation, the features of the node's neighboring nodes (i.e., variables in the SCG with direct causal relationships, such as "distance from water source" and "terrain slope" for "irrigation conditions") are weighted and aggregated according to the edge weights (reflecting the strength of causal relationships), and finally output an enhanced feature vector for each node.

[0086] Causal attention layer: assigns weights to different paths, focusing on key paths with high intensity (e.g., allocating more attention to the path "static factor → dynamic factor → abandonment");

[0087] Output layer: Outputs the contribution of each path to the "abandonment".

[0088] 4. Credibility Rating of Abandonment Causes (CAC)

[0089] In some embodiments, CAC is an overall score that is a weighted fusion of the contributions of all critical paths, used to quantify "land abandonment". For the selected critical paths (paths with intensity higher than the threshold), different weights are assigned according to the path type (e.g., the weight of direct path of dynamic factors > the weight of indirect path of static factors). The contribution of each path is multiplied by the corresponding weight and then summed to obtain a preliminary causal credibility score.

[0090] As an example, suppose there are two critical paths for the "GZ-101-23" plot:

[0091] Path 1: TST decline trend → abandoned, contribution 0.35, weight 0.6;

[0092] Path 2: 12° terrain slope → poor irrigation conditions → TST decline → abandonment, contribution 0.42, weight 0.4;

[0093] CAC value = 0.35 × 0.6 + 0.42 × 0.4 = 0.21 + 0.168 = 0.378.

[0094] In this embodiment, CAC integrates the contribution of each causal path, with a scoring range typically between 0 and 1. A higher value indicates a more complete causal logic for the land abandonment.

[0095] Step 105: For plots of land where both the abandonment suspicion score and the abandonment causal credibility score exceed the set threshold, an abandonment label is automatically generated.

[0096] In some embodiments, the ASS is compared with a preset "semantic drift threshold" (e.g., ASS ≥ 0.7), and the CAC is compared with a preset "causal credibility threshold" (e.g., CAC ≥ 0.6). If both exceed the corresponding threshold, the land parcel is determined to be abandoned, and the automatic label generation process is initiated, automatically generating a "abandoned land (high credibility)" label. If either exceeds the corresponding threshold, an "abandoned land (manual verification)" label is automatically generated. If both are less than the corresponding threshold, a "non-abandoned land" label is automatically generated. It is understood that other types of labels can also be classified based on the comparison between the two and the corresponding threshold, but this embodiment does not impose any limitations.

[0097] In some embodiments, the tags are bound to key information of the land parcel, including: basic attributes of the land parcel (geographic code, location boundary); specific values ​​and calculation basis of ASS and CAC (e.g., the core feature of ASS is the downward trend of NDVI, and the key path of CAC is "steep slope → poor irrigation → TST decline"); key feature segments of TST (e.g., annual NDVI decline curve, texture change rate map); causal path explanation (e.g., "due to the terrain slope of 12° and the lack of irrigation conditions, farming activities ceased after 2022, and vegetation continued to degrade").

[0098] Step 106: Write the land parcel information into the dynamic database. The land parcel information includes land parcel identifier, abandoned land tag, semantic trajectory, identification source, scoring result or timestamp.

[0099] In some embodiments, for each plot of land, six categories of core information are extracted and integrated to ensure data integrity and relevance, including:

[0100] Land parcel identifier: Includes a unique geographic code (such as "GZ-101-23"), spatial location coordinates (latitude and longitude boundaries), and land parcel outline vector information, used to accurately locate the spatial range of the land parcel.

[0101] Abandoned Land Labels: Based on the judgment results, the label type generated, such as "Abandoned Land (Highly Reliable)", "Abandoned Land (Manually Verified)" or "Non-Abandoned Land", clearly defines the current abandonment status of the land parcel.

[0102] Semantic Trajectory: The key feature sequence and trend summary of the time-series semantic trajectory (TST), which includes, for example, multimodal feature time-series data (such as the annual NDVI variation curve, SAR backscatter value time-series sequence, and surface temperature fluctuation record); trend description (such as "NDVI continuously decreased from 0.7 to 0.2 from 2020 to 2024").

[0103] Source identification: Record the source of the tag generation, which may include, for example, the analysis results of semantic drift detection (e.g., “NDVI continues to decline + seasonal pattern disruption”); causal path (e.g., “12° terrain slope → no irrigation → TST decline”); manual verification records (if based on field research, the verification personnel and date should be noted).

[0104] Scoring results: A set of quantitative indicators, including, for example, ASS and calculation dimensions (e.g., "ASS = 0.82, mainly due to NDVI trend and texture changes"), CAC and critical path contribution (e.g., "CAC = 0.86, core path contribution accounts for 75%)).

[0105] Timestamps: record key time points in the generation and updating of information, such as the start time of the first extraction of TST; the completion time of semantic drift detection; the time of causal inference and label generation; and the latest update time of the database.

[0106] When the status of a plot changes (e.g., a plot that was previously "pending verification" is manually confirmed to be "highly reliable abandoned"), or when new time-series data is added (e.g., new TST features are added), the system automatically updates the corresponding fields (e.g. abandoned tag, semantic trajectory, timestamp) and retains historical records to trace the evolution process.

[0107] like Figure 3 As shown, this embodiment of the invention provides a remote sensing map dynamic database construction system 300 for multi-source heterogeneous data fusion, comprising:

[0108] The first acquisition module 301 is used to acquire multi-source heterogeneous remote sensing data of the target area and preprocess the remote sensing data to extract plot remote sensing time series data.

[0109] The first construction module 302 is used to construct a temporal semantic trajectory for each of the land parcels, wherein the temporal semantic trajectory includes a sequential representation of the key semantic features of the land parcel attributes in the time dimension;

[0110] The first generation module 303 is used to identify semantic evolution trends based on the temporal semantic trajectory using a semantic drift detection model, and generate a suspected abandonment score.

[0111] The first reasoning module 304 is used to construct a structural causal graph based on static factors, dynamic factors and prior knowledge of the land parcel through a causal discovery algorithm, and to reason about the causal credibility score of the land parcel's abandonment based on the structural causal graph through a graph neural network model. The static factors reflect the inherent attributes of the land parcel and its long-term stable geographical conditions, the dynamic factors are feature change indicators extracted based on time-series semantic trajectories, and the prior knowledge is auxiliary information from external data sources.

[0112] The tag generation module 305 is used to automatically generate abandoned tags for plots of land whose abandonment suspicion score and abandonment causal credibility both exceed the set thresholds.

[0113] The writing module 306 is used to write land parcel information into a dynamic database. The land parcel information includes land parcel identifier, abandoned land tag, semantic trajectory, identification source, scoring result or timestamp.

[0114] like Figure 4 As shown, an embodiment of the present invention provides an electronic device, including a processor and a memory. The memory stores program instructions, and the processor executes the program instructions to implement the method for constructing a dynamic remote sensing map database by fusing multi-source heterogeneous data as described in any of the preceding claims.

[0115] Another embodiment of this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0116] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a processor, it performs the functions defined in the methods of this application.

[0117] It should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0118] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a dynamic database for remote sensing maps by fusing multi-source heterogeneous data, characterized in that, include: Acquire multi-source heterogeneous remote sensing data of the target area, and preprocess the remote sensing data to extract plot remote sensing time series data; For each of the aforementioned land parcels, a temporal semantic trajectory is constructed, wherein the temporal semantic trajectory includes a sequential representation of the key semantic features of the land parcel attributes in the time dimension; Based on the aforementioned temporal semantic trajectory, a semantic drift detection model is used to identify semantic evolution trends and generate a score indicating the suspicion of abandoned land. Based on static factors, dynamic factors, and prior knowledge of land parcels, a structural causal graph is constructed using a causal discovery algorithm. Based on the structural causal graph, a graph neural network model is used to infer the causal credibility score of land parcels for abandonment. The static factors reflect the inherent attributes and long-term stable geographical conditions of the land parcels, the dynamic factors are feature change indicators extracted based on temporal semantic trajectories, and the prior knowledge is auxiliary information from external data sources. For plots of land where both the doubt score for abandonment and the credibility of the causal relationship for abandonment exceed the set threshold, an abandonment label is automatically generated; The land parcel information is written into a dynamic database. The land parcel information includes land parcel identifier, abandoned land tag, semantic trajectory, identification source, scoring result or timestamp.

2. The method for constructing a dynamic remote sensing map database by fusing multi-source heterogeneous data according to claim 1, characterized in that, include: The construction of a temporal semantic trajectory for each land parcel includes extracting key semantic features that characterize the attributes of the land parcel. These key semantic features include vegetation index, texture features, thermal infrared features, and SAR intensity features. The key semantic features are then organized in chronological order, and the key semantic features corresponding to each time node are combined at each time node.

3. The method for constructing a dynamic database of remote sensing maps by fusing multi-source heterogeneous data according to claim 2, characterized in that, include: The semantic drift detection model includes an input layer, a modality adaptation embedding layer, a basic temporal feature extraction layer, a temporal attention layer, a multimodal consistency verification layer, a temporal trend and periodic analysis layer, and a score fusion output layer. The input layer receives the temporal semantic trajectory data. The modality adaptation embedding layer designs a small fully connected network for each modality feature, maps the key semantic features to an embedding space of the same dimension, and outputs the temporal embedding features of each modality. The basic temporal feature extraction layer uses a bidirectional LSTM network to concatenate the embedded features of each modality and output the hidden state H(t) at each time step. The temporal attention layer calculates attention weights to obtain the weights for each time step. The multimodal consistency verification layer detects the logical consistency of different modal features within the same time step and adjusts the weights corresponding to that time step based on the detection results. The multimodal consistency check layer detects the logical consistency of different modal features within the same time step and outputs the strength of contradictory signals. The time-series trend and cycle analysis layer extracts slope features from the attention-weighted time-series features through convolutional layers, outputs the trend degradation degree, and compares the difference between the key time intervals where attention is focused and historical cycles, outputting the cycle disruption degree. The scoring fusion output layer outputs a score indicating the likelihood of abandonment.

4. The method for constructing a dynamic database of remote sensing maps by fusing multi-source heterogeneous data according to claim 3, characterized in that, include: The difference between the key time interval for comparing attention focus and historical cycles is used to output the degree of cycle disruption. The process includes: inputting the original temporal semantic trajectory without attention weighting; converting the temporal features to the frequency domain through a Fourier transform layer; identifying the dominant periodic component; normalizing the amplitude of the dominant periodic component to obtain the periodic signal intensity; extracting the high-weight time interval focused by the temporal attention layer; comparing the periodic features of this interval with historical periods to obtain the signal intensity decrease ratio and phase shift amplitude; and weighted fusing the signal intensity decrease ratio and phase shift amplitude to obtain the periodicity disruption degree.

5. The method for constructing a dynamic database of remote sensing maps by fusing multi-source heterogeneous data according to claim 1, characterized in that, include: The method for reasoning about the causal credibility of abandoned land parcels based on the graph neural network model includes: traversing all directed paths in the causal graph pointing to the endpoint, generating candidate causal paths, calculating the causal strength of each candidate path to obtain the key path, calculating the path contribution based on the graph neural network model, and summing the contribution of each key path by its corresponding weight to obtain a preliminary causal credibility score.

6. A system for constructing a dynamic database for remote sensing maps by fusing multi-source heterogeneous data, characterized in that, include: The first acquisition module is used to acquire multi-source heterogeneous remote sensing data of the target area, and preprocess the remote sensing data to extract plot remote sensing time series data; The first construction module is used to construct a temporal semantic trajectory for each of the land parcels, wherein the temporal semantic trajectory includes a sequential representation of the key semantic features of the land parcel attributes in the time dimension; The first generation module is used to identify semantic evolution trends based on the temporal semantic trajectory using a semantic drift detection model, and generate a suspected abandonment score. The first reasoning module is used to construct a structural causal graph based on static factors, dynamic factors and prior knowledge of the land parcel through a causal discovery algorithm, and to reason about the credibility score of the land parcel's abandonment causality based on the structural causal graph through a graph neural network model. The static factors reflect the inherent attributes of the land parcel and its long-term stable geographical conditions, the dynamic factors are feature change indicators extracted based on time-series semantic trajectories, and the prior knowledge is auxiliary information from external data sources. The tag generation module is used to automatically generate abandoned tags for plots of land whose abandonment suspicion score and abandonment causal credibility both exceed the set thresholds. The writing module is used to write land parcel information into a dynamic database. The land parcel information includes land parcel identifier, abandoned land tag, semantic trajectory, identification source, scoring result or timestamp.

7. The remote sensing map dynamic database construction system based on multi-source heterogeneous data fusion according to claim 6, characterized in that, include: The construction of a temporal semantic trajectory for each land parcel includes extracting key semantic features that characterize the attributes of the land parcel. These key semantic features include vegetation index, texture features, thermal infrared features, and SAR intensity features. The key semantic features are then organized in chronological order, and the key semantic features corresponding to each time node are combined at each time node.

8. An electronic device, the electronic device comprising: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, characterized in that: the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1-5.

9. A computer-readable medium storing computer program instructions thereon, characterized in that: The computer program instructions can be executed by a processor to implement the method as described in any one of claims 1-5.

10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, it implements the method as described in any one of claims 1-5.

Images