Dynamic monitoring system for forest and grass resources

Abstract

The invention relates to the technical field of forestry management, in particular to a forest and grass resource dynamic monitoring system which comprises the steps that a graph attention network is adopted to conduct high-low weight recognition processing on the structure difference value between node pairs, feature vectors are established through the canopy height difference and canopy density difference between adjacent nodes, and the canopy height difference and the canopy density difference between adjacent nodes are obtained; an edge weight scoring system is constructed in the form of segmented statistics and proportion weighting, node edge pairs with high influence relation strength are dynamically screened, ordered aggregation of spatial communication strength is realized, a graph neural network is introduced in a parameter fusion stage to construct a node parameter representation mechanism, and the spatial communication strength is improved. The tree species proportion, the grade of diameter at breast height and the community vertical structure in a regional sample plot are used as input features, unified mapping of node features of each region is completed in multiple rounds of iteration, cross-regional difference analysis is executed on model output in combination with biomass change frequency, parameter items with the difference proportion lower than a set threshold value are replaced with unified expression, and the model output is obtained. And parameter synchronization is realized to realize space nesting and feature integration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forestry management, and in particular to a forest and grass resource dynamic monitoring system. BACKGROUND

[0002] The technical field of forestry management refers to the technical activities of organizing, planning, monitoring, evaluating and scheduling of natural ecological resources such as forests and grasslands, including forest resource information collection, forest and grass growth dynamic monitoring, ecological quality evaluation, pest warning, resource utilization control, fire risk control, etc. The continuous grasp and efficient management of forest and grass resource status support scientific decision-making and ecological safety guarantee.

[0003] A forest and grass resource dynamic monitoring system is an integrated information system for realizing real-time monitoring, data analysis and early warning feedback of the state of natural resources such as forest land and grassland. The design purpose of the system is to improve the efficiency, accuracy and timeliness of forest and grass resource monitoring, realize quantitative evaluation of resource change trend and timely discovery of abnormal state. The system realizes resource protection refinement, management digitization and decision-making scientization by building multi-source data fusion, spatial distribution monitoring and index evaluation model, so as to improve the ecological governance ability and the level of sustainable utilization of regional forest and grass resources.

[0004] The prior art generally relies on a single data source or a static structure model for periodic processing in forest and grass resource dynamic monitoring. The monitoring behavior mainly focuses on change comparison and parameter difference calculation at the polygon level, and lacks the ability of continuous fusion across time scales and spatial structures. The data collection method is executed independently in stages, and it is difficult to establish the relationship mapping between the unified timeline and multi-source data, resulting in a break in the analysis results caused by the time sequence difference. The judgment of the relationship between nodes cannot introduce the similarity measurement between regional structures, which may cause adjacent forest plots to be incorrectly determined as trend anomalies when the coverage difference is small. Moreover, the existing method is often realized by substituting the average value or using a fixed coefficient adjustment method, which is difficult to express the unique statistical difference between different regions, and may produce significant deviation in the plot structure with obvious differences, resulting in reduced model generalization ability. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art and to provide a forest and grass resource dynamic monitoring system.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme: a forest and grass resource dynamic monitoring system comprises: A space-time graph construction module: based on the forest plot block canopy height, canopy density and soil moisture data obtained by remote sensing image interpretation, the nodes are numbered and mapped, the center of gravity is calculated and threshold judged, the time label is aligned, the node and edge set are established, and the space-time graph structure is generated; The weight distribution module: based on the space-time graph structure, calling the difference value of the forest block crown height generated by aerial remote sensing and the adjacent forest block canopy density, using graph attention network, threshold judgment, scoring and sorting adjacent edges by piecewise statistics and ratio operation, extracting high-weight edges and aggregating, obtaining a dynamic weight matrix; The collaborative updating module: based on the dynamic weight matrix, calling the tree species distribution statistics and community structure parameters generated by the forestry resource survey database, using graph neural network, adopting sample plot statistics and biomass change coefficient to count iteration times, accumulating the difference proportion of cross-regional parameters, comparing with the original parameter difference value, replacing the items less than the threshold value, and establishing a unified model parameter; The trend analysis module: based on the unified model parameter, calling the monthly remote sensing interpretation of the crown coverage and the ground collection of air temperature and humidity, using fixed window accumulation difference and calculating sliding average, comparing with the global average value to screen forest blocks, and obtaining a growth trend list; The abnormality judgment module: based on the growth trend list and the unified model parameter, calling the measured crown coverage and the predicted coverage, calculating the absolute value of the difference one by one and sorting, executing threshold comparison to screen the over-limit forest blocks, and obtaining an abnormal block set.

[0007] As a further scheme of the present application, the space-time graph construction module comprises: The node mapping submodule: based on the forest block crown height, canopy density and soil humidity data obtained by remote sensing image interpretation, the state vector is constructed by extracting the grid value within the boundary of each forest block, the state vector is matched with the node label, the threshold value is judged after the difference value of the longitude and latitude coordinates is calculated, the time label value is sorted and aligned, and the node set is established; The edge set construction submodule: based on the node set, the boundary coordinates of the node pairs are read to perform distance difference calculation and threshold judgment, the node pairs meeting the conditions are connected to form edges and integrated into the edge set, and the edge set is de-duplicated and renumbered to obtain the space-time graph structure.

[0008] As a further scheme of the present application, the weight distribution module comprises: The difference calculation submodule: based on the space-time graph structure, calling the difference value of the forest block crown height generated by aerial remote sensing and the adjacent forest block canopy density, using graph attention network, reading node pair data one by one, extracting the crown height and canopy density values of the two nodes, calculating the height difference and canopy density difference between the nodes, performing spatial adjacency verification combined with the geographical location of the block, removing the node pairs that do not meet the connectivity relationship, and generating a difference matrix; The weight evaluation submodule: based on the difference matrix, dividing intervals according to the difference value range and counting the occurrence frequency of each interval, calculating the weighted ratio combined with the spatial distribution characteristics of the node pairs, assigning weights to each edge, extracting all edge weight values to form a list, sorting and marking the weight ranking according to the weight size, and generating a weight sequence; Edge aggregation sub-module: based on the weight sequence, filtering the edge weights in the front of the ranking, combining the corresponding node index and weight value into edge weight mapping, establishing an edge set and removing redundant edges, integrating to generate a weighted connection relationship matrix according to the node number, and obtaining a dynamic weight matrix.

[0009] As a further scheme of the present application, the graph attention network first reads the canopy height and canopy density values of the nodes corresponding to each edge based on the spatio-temporal graph structure, then maps the features of each node to a unified embedding space through a shared linear transformation module, extracts the mapped features according to the edge pairs, then inputs a learnable attention mechanism module, calculates the feature similarity score between nodes, generates attention weights after normalization processing, and finally aggregates the adjacent node features to the target node according to the attention weights, and outputs the updated node representation.

[0010] As a further scheme of the present application, the graph attention network is according to the formula: Where: Δij' represents the comprehensive difference value between forest plot i and forest plot j, α represents the weight coefficient of the canopy height difference term, β represents the weight coefficient of the canopy density difference term, γ represents the weight coefficient of the Euclidean space distance term, δ represents the weight coefficient of the remote sensing collection time difference term, ε represents the weight coefficient of the environmental factor difference term, hi represents the canopy height value of forest plot i, hj represents the canopy height value of forest plot j, ci represents the canopy density value of forest plot i, cj represents the canopy density value of forest plot j, dij represents the Euclidean space distance between forest plot i and forest plot j, represents the collection time difference of the remote sensing data corresponding to forest plot i and forest plot j, and eij represents the normalized environmental factor difference value between forest plot i and forest plot j.

[0011] As a further scheme of the present application, the collaborative updating module comprises: Parameter integration sub-module: based on the dynamic weight matrix, calling the tree species distribution statistics and community structure parameters generated by the forestry resource survey database, importing the tree species proportion, diameter grade and community vertical structure index in each regional sample plot using graph neural network, unifying the field format, merging according to the regional dimension, forming a multi-dimensional parameter set across regions, and generating a parameter summary table; Difference calculation sub-module: based on the parameter summary table, comparing the same type of parameters in different regions field by field, calculating the maximum and minimum difference and recording, combining with the historical biomass change coefficient to analyze the update frequency of each region, marking the fields with a difference greater than the average difference as difference key items, and generating a difference identification set; Parameter replacement submodule: based on the difference identification set, the difference value of each item in the mark field and the original parameter is distinguished, the parameter item with a difference value lower than the set range is locked and positioned in the corresponding position in the unified parameter set, the original parameter is replaced with a unified value, the updated comprehensive parameter set is reconstructed, and a unified model parameter is established.

[0012] As a further scheme of the present application, the graph neural network first reads the tree species proportion, diameter class level and community vertical structure index of each regional sample plot, constructs a node feature vector, and imports edge information according to the regional boundary relationship to form a graph structure, then relies on the graph structure to perform feature embedding on each node, converts the original features into low-dimensional representations through a shared mapping mechanism, in each iteration, the feature vectors of adjacent nodes are weighted and aggregated to the target node through an information transmission mechanism between adjacent nodes, realizing local structure information fusion, and then through a nonlinear transformation to improve the node representation ability, repeating the propagation and aggregation process for multiple rounds to obtain the updated table of each node considering the structural differences between regions.

[0013] As a further scheme of the present application, the graph neural network is according to the formula: Wherein: represents the multi-dimensional forest structure fusion evaluation value of region r, μ represents the weight coefficient of the tree species proportion index, pr represents the tree species composition proportion of the sample plot in region r, represents the weight coefficient of the diameter class distribution index, dr represents the diameter class structure distribution of the sample plot in region r, represents the weight coefficient of the community vertical structure index, vr represents the community vertical stratification index of the sample plot in region r, ρ represents the weight coefficient of the community complexity index, cr represents the community structure complexity index of the sample plot in region r, κ represents the weight coefficient of the Shannon diversity index, and sr represents the Shannon tree species diversity index of region r. represents the weight coefficient of the biomass density index, br represents the forest biomass density per unit area of region r, ϕ represents the weight coefficient of the landscape fragmentation index, fr represents the landscape fragmentation index of region r, ψ represents the weight coefficient of the terrain ruggedness index, and tr represents the terrain ruggedness value of region r.

[0014] As a further scheme of the present application, the trend analysis module comprises: Data aggregation submodule: based on the unified model parameter, the crown coverage rate interpreted by monthly remote sensing and the air temperature and humidity collected on the ground are called, the monthly image coverage rate file is compared with the meteorological monitoring log, the coverage rate and the temperature and humidity difference at the same time point are calculated and recorded, the difference value is accumulated and summed according to the fixed window interval, and the cumulative difference value set is generated. Trend filtering submodule: Based on the accumulated difference set, the sliding window length is set to gradually extract the difference sequence, the sum of the differences within the window is compared with the overall average value, the over-limit land parcel identifiers are collected and the index is converted to obtain the growth trend list.

[0015] As a further aspect of the present invention, the anomaly determination module includes: The difference calculation submodule: Based on the growth trend list and unified model parameters, it calls the measured canopy coverage rate and the model predicted coverage rate, reads the corresponding time coverage records in the monitoring database, subtracts the prediction results one by one, converts the absolute value and stores it, and generates a coverage difference set; Anomaly Filtering Submodule: Based on the coverage difference set, sort the differences in descending order, compare them with a set threshold, extract the IDs of the over-limit plots, organize them, and obtain the anomaly plot set.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by encoding and mapping the raster values ​​of canopy height, canopy closure and soil moisture in remote sensing interpretation data into nodes, and combining the judgment of centroid offset distance and time label sorting, the coupling construction of temporal and spatial features is completed, forming a dynamic node relationship. In this invention, a graph attention network is used to identify the high and low weights of the structural differences between node pairs. Feature vectors are established by the difference in canopy height and canopy closure between adjacent nodes. An edge weight scoring system is constructed in the form of segmented statistics and proportional weighting. Node edge pairs with high influence relationship strength are dynamically screened to achieve ordered aggregation of spatial connectivity strength and effectively suppress redundant path interference in edge weight distribution. In this invention, a graph neural network is introduced in the parameter fusion stage to construct a node parameter representation mechanism. The tree species ratio, diameter at breast height (DBH) grade, and vertical structure of the community in the regional sample plots are used as input features. In multiple iterations, the unified mapping of node features in each region is completed. Combined with the frequency of biomass change, cross-regional difference analysis is performed on the model output. Parameters with a difference ratio below a set threshold are replaced with a unified expression, so that the parameters can achieve spatial nesting and feature integration simultaneously. Attached Figure Description

[0017] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, in the description of the present application, "a plurality of" means two or more, unless otherwise explicitly and specifically limited.

[0020] Embodiment one Please refer to Figure 1 The present application provides a technical solution: a forest and grass resource dynamic monitoring system comprises: The spatiotemporal graph construction module: based on the forest block crown height, canopy density and soil moisture data obtained by remote sensing image interpretation, the nodes are numbered and mapped, the barycentric offset is calculated and threshold judgment is performed, the time label is aligned, the node and edge set are established, and the spatiotemporal graph structure is generated; The weight distribution module: based on the spatiotemporal graph structure, the difference between the forest block crown height generated by aerial remote sensing and the adjacent forest block canopy density is called, the graph attention network is used, the threshold judgment is performed, the adjacent edge is scored and sorted by piecewise statistics and ratio operation, the high-weight connected edge is extracted and aggregated, and the dynamic weight matrix is obtained; The collaborative updating module: based on the dynamic weight matrix, the tree species distribution statistics and community structure parameters generated by the forestry resource survey database are called, the graph neural network is used, the sample plot statistics value and the biomass change coefficient are counted and iterated, the difference proportion of cross-regional parameters is accumulated and compared with the original parameter difference, the items less than the threshold value are replaced, and the unified model parameters are established; The trend analysis module: based on the unified model parameters, the crown cover rate interpreted by monthly remote sensing and the air temperature and humidity collected on the ground are called, the difference is accumulated by fixed window and the moving average is calculated, the forest blocks are compared with the global average to screen, and the growth trend list is obtained; The anomaly determination module: based on the growth trend list and the unified model parameters, the measured crown cover rate and the predicted cover rate are called, the difference absolute value is calculated and sorted one by one, the threshold comparison is performed to screen the over-limit forest blocks, and the abnormal plot set is obtained.

[0021] The spatiotemporal graph construction module comprises: The node mapping submodule: based on the forest block crown height, canopy density and soil moisture data obtained by remote sensing image interpretation, the state vector is constructed by extracting the grid value within the boundary of each forest block, the state vector is matched with the node label, the threshold judgment is performed after the difference value of the longitude and latitude coordinates is calculated, the time label value is sorted and aligned, and the node set is established; Edge set construction submodule: based on the node set, traverse the node pair to read the boundary coordinates to perform distance difference calculation and threshold judgment, generate edge connection for the node pairs meeting the conditions and integrate them into the edge set, remove duplicates and renumber and arrange, obtain the space-time graph structure; Node mapping submodule: based on the forest block crown height, canopy density and soil moisture data obtained by interpreting remote sensing images, using the sliding window pixel extraction method, the remote sensing image within the boundary of each forest block is tiled and scanned with a window size of 32x32 pixels, the grid values in each window are selected and flattened into a vector in row priority order, the mean value of the vector is calculated to obtain the mean values of the crown height, canopy density and soil moisture, respectively, the three mean values are sequentially spliced into a state vector, the field mapping numbering method is used to traverse the identification field in the forest block vector label file, the state vector number is indexed and bound with the forest block ID field, the projection coordinate difference calculation method is used to read the centroid coordinates of each forest block, the two-dimensional Euclidean distance function is used to calculate the coordinate difference between the centroids, the threshold is set to 20 meters, the difference is filtered, and all coordinate differences in the filtering result less than the threshold are marked as legal nodes, the time field sorting method is used to convert the time labels attached to each grid image into a unified format, and the numerical values are arranged in ascending order of year, month and day after conversion, and the node index field is rearranged and output after rearrangement, generating a node set; Edge set construction submodule: based on the node set, using the nearest neighbor boundary screening method, for any two node combinations, the boundary vector data reading tool is called to import its boundary coordinate set, the boundary envelope box center point difference calculation method is used to extract the maximum and minimum x, y coordinate values of each boundary, the coordinate difference of the center point position is calculated, the filtering distance threshold is set to 1 km, the Boolean judgment filtering operation is used to perform conditional judgment operation on all difference value data, and the node pairs whose center point difference is within the threshold are set as connectable state, the node pair writing method is used to combine and encode the legal node pairs according to the index value, and write them into the edge set structure list in order, the edge set de-duplication operation is performed, the value comparison is performed on all generated edge structure information to exclude redundant items, and the remaining edge structure is renumbered in ascending order through the position renumbering operation, and finally the space-time graph structure is generated.

[0022] The weight distribution module includes: Difference calculation submodule: based on the space-time graph structure, the forest block crown height generated by aerial remote sensing and the adjacent forest block canopy density are called, the graph attention network is used, the node pair data is read one by one, the crown height and canopy density values of the two nodes are extracted, the height difference and canopy density difference between the nodes are calculated, the spatial adjacency verification is performed combined with the geographical position of the block, the node pairs that do not meet the connectivity relationship are excluded, and the difference matrix is generated; The weight evaluation submodule: based on the difference matrix, the interval is divided according to the difference range and the frequency of each interval is counted, the weighted ratio is calculated combined with the spatial distribution characteristics of the node pair, the weight is distributed to each edge, the weight value of all edges is extracted to form a list, the weight is ranked according to the weight size and marked, and the weight sequence is generated; The edge aggregation submodule: based on the weight sequence, the edges with the top ranking weights are screened, the corresponding node index and weight value are combined into an edge weight mapping, the edge set is established and the redundant edges are removed, the weighted connection relationship matrix is integrated according to the node number, and the dynamic weight matrix is obtained; The difference calculation submodule: based on the space-time graph structure, the crown layer height generated by the aviation remote sensing and the adjacent forest block canopy density data are called, the graph attention network method is adopted, each group of node pair data is read and input, each node contains two features, crown layer height and canopy density, the crown layer height value and canopy density value of the source node and the target node are extracted in turn, the crown layer height difference and the canopy density difference between the two nodes are calculated, the centroid coordinates of the corresponding block boundary are called for each group of node pairs, the two-dimensional plane coordinate distance between the two centroids is calculated, the spatial adjacency judgment threshold is set to 500 meters, the node pairs with a distance exceeding 500 meters are removed, and two numerical tables are constructed in the remaining node pair set to store the crown layer height difference and the canopy density difference. The two difference tables are spliced into a difference combination table according to the column, and finally a difference matrix is generated; The weight evaluation submodule: based on the difference matrix, the segmented statistical weighted evaluation method is adopted, the crown layer height difference and the canopy density difference are divided into 10 fixed width numerical intervals, each interval is set to 0.2 meters, the frequency of each edge structure corresponding to the difference value falling into each interval is counted, the total number of frequencies of each interval is recorded, the spatial distance value between the corresponding nodes of each edge structure is extracted, and the weight of each edge is calculated as the frequency of the interval it belongs to divided by its spatial distance. After traversing all edge structures to complete weight calculation, an edge weight list is established, the edge structures are sorted by weight value from large to small, and each edge is labeled with a serial number marker after sorting, and finally a weight sequence is generated; The edge aggregation submodule: based on the weight sequence, the top ranking screening aggregation method is adopted, the screening threshold is set to the top 30%, the edge structures with the top 30% weight ranking are selected, the corresponding source node number, target node number and weight value of each edge are extracted, the three contents are combined to form an edge weight mapping structure, the existing duplicate edge structures are numbered and merged and the weight is averaged, the duplicate items are deleted, and all edge structures are renumbered according to the source node number from small to large. When constructing the connection relationship matrix, the matrix size is set to the total number of nodes multiplied by the total number of nodes, and the weight value of each edge structure is written into the corresponding position according to the node number, and finally a dynamic weight matrix is generated.

[0023] The graph attention network first reads the crown height and canopy density values of the nodes corresponding to each edge based on the spatio-temporal graph structure, then maps the node features to a unified embedding space through a shared linear transformation module, extracts the mapped features according to the edges, inputs them into a learnable attention mechanism module, calculates the feature similarity scores between nodes, normalizes them to generate attention weights, and finally aggregates the adjacent node features to the target node according to the attention weights to output the updated node representation; The graph attention network is according to the formula: Wherein: Δij' represents the comprehensive difference value between forest plot i and forest plot j, represents the weight coefficient of the crown height difference term, represents the weight coefficient of the canopy density difference term, represents the weight coefficient of the Euclidean space distance term, represents the weight coefficient of the remote sensing collection time difference term, and ε represents the weight coefficient of the environmental factor difference term. hi represents the crown height value of forest plot i, hj represents the crown height value of forest plot j, ci represents the canopy density value of forest plot i, and cj represents the canopy density value of forest plot j. dij represents the Euclidean space distance between forest plot i and forest plot j, represents the collection time difference of the remote sensing data corresponding to forest plot i and forest plot j, and eij represents the normalized environmental factor difference value between forest plot i and forest plot j. The execution process is as follows: first, the crown height hi and hj of forest plot i and forest plot j are extracted based on the aerial remote sensing image data, the absolute value of the difference hi-hj is calculated, and the weight coefficient is multiplied, then the canopy density data ci and cj of the corresponding forest plot are read, the absolute value of the difference ci-cj is calculated, and the weight coefficient is multiplied, then the distance dij is obtained according to the geographical coordinate information of the forest plot using the Euclidean distance calculation formula, and the weight coefficient is multiplied, then the time difference is calculated by taking the absolute value of the collection time stamp of the aerial remote sensing data of the two regions and multiplying the coefficient δ, and finally the environmental factors such as leaf area index and net primary productivity corresponding to each forest plot are read from the ecological environment database, the difference eij is calculated after normalization, and the weight coefficient is multiplied. The above five items respectively measure the difference between forest plots in terms of biophysical structure, spatial layout, time update frequency and environmental ecological response, and the comprehensive difference value Δij' is generated by weighted summation, which is used to construct the edge weight matrix between nodes, so as to guide the monitoring model to identify abnormal forest change trend and realize dynamic and accurate monitoring of forest and grass resources.

[0024] The collaborative updating module includes: The parameter integration submodule: based on the dynamic weight matrix, calling the tree species distribution statistics and community structure parameters generated by the forestry resource survey database, using graph neural network, importing the tree species proportion, diameter class and community vertical structure index in each regional sample plot, unifying the field format, merging according to the regional dimension, forming the multi-dimensional parameter set across regions, generating the parameter summary table; The difference calculation submodule: based on the parameter summary table, comparing the same type parameters of different regions field by field, calculating the maximum and minimum difference and recording, combining the historical biomass change coefficient to analyze the update frequency of each region, marking the fields with a difference greater than the average difference as the difference key items, generating the difference identification set; The parameter replacement submodule: based on the difference identification set, the marked fields are compared with the original parameters one by one to determine the difference value, the parameter items with a difference value lower than the set range are locked and positioned in the corresponding position in the unified parameter set, the original parameters are replaced by the unified value, and the updated comprehensive parameter set is reconstructed to establish the unified model parameters; The parameter integration submodule: based on the dynamic weight matrix, calling the tree species distribution statistics and community structure parameters generated by the forestry resource survey database, using graph neural network method, setting the input format of graph structure as node feature matrix and adjacency matrix, where the node feature matrix contains sample plot number, tree species proportion, diameter class, community vertical structure index, and the adjacency matrix is constructed by direct conversion of dynamic weight matrix, the input dimension of graph neural network is initialized as 4, the hidden layer dimension is 16, the output dimension is 4, and the iteration layer is set as 3 layers, the full connection structure is used for each layer network, and the input matrix is processed by layer-by-layer feature aggregation and activation function, the node representation result is recorded after each iteration and written into the intermediate representation matrix, after the convergence of the node features of all sample plots, the representation matrix is grouped according to the regional label field, the unified field format operation includes field renaming, data type unification, missing value zero padding, and decimal precision unification to 2 decimal places, the multi-dimensional feature matrix in regional dimension is generated by horizontal splicing for each group of sample plot data, then vertical splicing is performed to combine into a complete cross-regional multi-dimensional parameter set, and finally the parameter summary table is generated; The difference calculation submodule: based on the parameter summary table, using the field-by-field statistical analysis method, traversing all feature fields, extracting the corresponding value column in all regional sample plots for each field, finding the maximum and minimum value by column and calculating the difference value, recording each field difference value result to establish the field difference list, arranging according to the field index order, then calling the historical biomass change sequence data of the region, calculating the unit area biomass increment value for each region per year, summarizing all sample plot corresponding time series values, setting the analysis time period as 10 years, calculating the average variation amplitude for each year, calculating the average biomass change of each region, identifying the fields with a difference higher than the average difference as the difference key items according to the field difference and change frequency, establishing the difference identification marker array and recording the corresponding field name and index position, and finally generating the difference identification set; Parameter replacement submodule: based on the difference identification set, the interval judgment matching replacement method is adopted, the original parameter value of each marked field is read in turn and the corresponding field value in the unified parameter set is calculated, the absolute difference between the two values is calculated, the difference threshold range is set as ±5%, the parameter items with the difference in the range are replaced, the replacement operation is positioned by the field position index to find the coordinate position of the original parameter in the matrix, the corresponding value in the unified parameter set is extracted and written into the corresponding coordinate of the original matrix, after the replacement of all fields, the updated matrix is checked, the integrity verification, field uniqueness verification and field order rearrangement are performed, and finally the unified model parameters are generated.

[0025] The graph neural network first reads the tree species proportion, diameter class and community vertical structure index of each regional sample plot, constructs a node feature vector, and imports edge information according to the regional boundary relationship to form a graph structure. Then, relying on the graph structure, feature embedding is performed on each node. The original features are converted into low-dimensional representations through a shared mapping mechanism. In each iteration, the feature vectors of adjacent nodes are weighted and aggregated to the target node through an information transmission mechanism between adjacent nodes to realize local structure information fusion. Then, the node representation ability is improved through a nonlinear transformation. The propagation and aggregation process is repeated for multiple rounds to obtain the updated table of each node considering the structural differences between regions.

[0026] The graph neural network is according to the formula: Wherein: represents the multi-dimensional forest structure fusion evaluation value of region r, represents the weight coefficient of the tree species proportion index, pr represents the tree species composition proportion of the sample plot in region r, represents the weight coefficient of the diameter class distribution index, dr represents the diameter class structure distribution of the sample plot in region r, represents the weight coefficient of the community vertical structure index, vr represents the community vertical stratification index of the sample plot in region r, represents the weight coefficient of the community complexity index, cr represents the community structure complexity index of the sample plot in region r, κ represents the weight coefficient of the Shannon diversity index, sr represents the Shannon tree species diversity index of region r, λ represents the weight coefficient of the biomass density index, br represents the forest biomass density per unit area of region r, φ represents the weight coefficient of the landscape fragmentation index, fr represents the landscape fragmentation index of region r, ψ represents the weight coefficient of the terrain ruggedness index, tr represents the terrain ruggedness value of region r. Execution process: first, based on the forestry resource survey database, the tree species composition proportion pr of the sample plot in region r is extracted, corresponding to the diameter class distribution dr, the community vertical stratification index vr and the structure complexity index cr, the four types of structure parameters are input and multiplied by the weight coefficients Subsequently, combining aerial remote sensing imagery and field monitoring data, the Shannon species diversity index (sr) of the sample plots within the region was calculated, the forest biomass density per unit area (br) was estimated, and the landscape fragmentation index (fr) and the terrain ruggedness index (tr) calculated by the digital elevation model were obtained from the geographic information system and multiplied by weighting coefficients. Then, the above eight items are weighted and summed to obtain the fusion evaluation value Sr′, which is used to characterize the dynamic comprehensive characteristics of forest and grassland resources in multiple dimensions such as structural stability, biodiversity, ecological quality and topographic adaptability. The weight coefficients are calculated by principal component analysis to calculate the variance contribution rate of each indicator in the sample set and normalize the values ​​to control the influence of each parameter in the fusion calculation.

[0027] The trend analysis module includes: Data aggregation submodule: Based on unified model parameters, it calls the canopy coverage interpreted by monthly remote sensing and the temperature and humidity collected on the ground, imports the monthly image coverage file and meteorological monitoring log for comparison, calculates and records the difference between coverage and temperature and humidity at the same time point, and accumulates the difference at fixed window intervals to generate a cumulative difference set. Trend filtering submodule: Based on the cumulative difference set, the difference sequence is gradually truncated by setting the sliding window length, the sum of the differences within the window is calculated and compared with the overall average, the over-limit land parcels are collected and converted into indexes, and a list of growth trends is obtained; The data aggregation submodule, based on unified model parameters, calls the canopy coverage data interpreted monthly by remote sensing and the temperature and humidity data collected from the ground. It adopts a time-synchronized difference calculation method, imports monthly image coverage files, sets the file naming format to a unified encoding structure of year, month, and day, iterates through each remote sensing file to extract the date tag field, performs a unified format conversion operation with the time field in the ground meteorological monitoring log record, constructs a unified year and month index tag using string truncation and integer concatenation, performs index matching on the two datasets according to the year and month tags, obtains the coverage value, temperature value, and humidity value at the same time point, and calculates the difference between coverage and temperature, and the difference between coverage and humidity using direct subtraction operations. The difference results are stored in the record table with the plot ID as the key. The difference list is arranged in chronological order for each plot, and a fixed window interval length of 3 months is set. The difference sequence of each plot is segmented in a sliding window manner, and the difference within each segment is accumulated. During the sliding process, the start and end time index and plot number of each segment are retained, and finally, a cumulative difference set is generated. Trend screening submodule: based on the cumulative difference set, a sliding window sequence comparison method is adopted, the sliding window length is set to 3, the difference value sequence corresponding to each plot is processed by sliding traversal, the difference value sum in each window segment is extracted, the value is compared with the average value of all window segments of the current plot, the average value is calculated by dividing the sum of all sliding window segments in the current plot difference value sequence by the number of segments, the over-limit judgment condition is set as the difference value sum of the current window segment is greater than 1.2 times the average value, the starting time label and plot number of all window segments meeting the condition are recorded, the plot number is uniformly converted into plot index format, the starting time and index value combination structure of each plot is retained and written into the identification table, and finally the growth trend list is generated.

[0028] The anomaly determination module includes: Difference calculation submodule: based on the growth trend list and unified model parameters, the measured canopy coverage and the model predicted coverage are called, the corresponding time coverage records in the monitoring database and the prediction results are subtracted one by one, the absolute value is converted and stored, and the coverage difference set is generated; Abnormal screening submodule: based on the coverage difference set, the difference values are arranged in descending order according to the size, compared with the set threshold, the over-limit plot ID is extracted and sorted, and the abnormal plot set is obtained; Difference calculation submodule: based on the growth trend list and unified model parameters, the measured canopy coverage and the model predicted coverage data are called, the corresponding time point difference calculation method is adopted, the measured coverage value recorded in the monitoring database of each plot is read, and the model predicted coverage value at the same time node of the plot is read synchronously, the read field name is set as plot ID, year-month label, measured coverage, predicted coverage, the matching condition is that the plot ID and the year-month label are consistent at the same time, the measured value of each record is subtracted from the predicted value, the obtained difference is uniformly converted into absolute value, the conversion method is that each value is multiplied by negative one if it is negative, and the original value is retained if it is positive, the generated absolute difference still retains the plot ID and time label fields corresponding to the original record, all difference records are aggregated into separate difference sequences according to the plot ID, the difference sequence of each plot is written into the difference summary table and arranged in order according to the time label, and finally the coverage difference set is generated; Abnormal screening submodule: based on the coverage difference set, a threshold difference filtering sorting method is adopted, the sorting basis is set as the difference value from large to small, the maximum value of the difference sequence in each plot is taken as the abnormal determination value corresponding to the plot, a maximum difference list of all plots is established, the abnormal threshold is set as 0.25, the unit is percentage difference, the size of each record in the difference list is compared and judged, if the current difference is greater than the threshold, the corresponding plot ID is extracted and written into the result list, a structured sorting operation is performed to sort all over-limit plot IDs in order, the unique ID sequence is generated after removing the repeatedly appearing plot IDs, and finally the abnormal plot set is generated.

[0029] The above merely provides the preferred embodiment of the present application and not intended to limit the present application to other forms. Any person skilled in the art can make modifications and variations to the present application without departing from the spirit and scope of the present application. Therefore, the scope of the present application shall be defined and protected by the appended claims.

Claims

1. A forest and grass resource dynamic monitoring system, characterized in that, The system comprises: The spatiotemporal graph construction module: based on the forest block crown height, canopy density and soil moisture data obtained by remote sensing image interpretation, the node mapping is performed, the center of gravity offset is calculated and threshold judgment is performed, the time label is aligned, the node and edge set are established, and the spatiotemporal graph structure is generated; The weight distribution module: based on the spatiotemporal graph structure, the difference between the forest block crown height generated by aerial remote sensing and the adjacent forest block canopy density is called, the graph attention network is adopted, the threshold judgment is performed, the adjacent edge is scored and sorted by segment statistics and ratio operation, the high-weight connected edge is extracted and aggregated, and the dynamic weight matrix is obtained; The collaborative updating module: based on the dynamic weight matrix, the tree species distribution statistics and community structure parameters generated by the forestry resource survey database are called, the graph neural network is used, the sample plot statistics value and the biomass change coefficient are counted, the iteration times are counted, the difference between the parameters is compared, the parameters less than the threshold value are replaced, and the unified model parameters are established; The trend analysis module: based on the unified model parameters, the monthly remote sensing interpretation crown cover and the ground collected temperature and humidity are called, the fixed window difference value is accumulated and the moving average is calculated, the forest block is screened by comparing with the global average value, and the growth trend list is obtained; The abnormality judgment module: based on the growth trend list and the unified model parameters, the measured crown cover and the predicted cover are called, the difference absolute value is calculated and sorted one by one, the threshold comparison is performed to screen the over-limit forest block, and the abnormal block set is obtained.

2. The system according to claim 1, wherein, The spatiotemporal graph construction module comprises: The node mapping submodule: based on the forest block crown height, canopy density and soil moisture data obtained by remote sensing image interpretation, the state vector is constructed by extracting the grid value within the boundary of each forest block, the state vector is matched with the node label, the threshold judgment is performed after the difference value of the longitude and latitude coordinates is calculated, the time label value is aligned, and the node set is established; The edge set construction submodule: based on the node set, the boundary coordinates of the node pairs are read to perform distance difference calculation and threshold judgment, the node pairs meeting the conditions are connected to form edges and integrated into the edge set, and the time-space graph structure is obtained after removing the duplicates and renumbering.

3. The system according to claim 1, wherein, The weight distribution module comprises: The difference calculation submodule: based on the spatiotemporal graph structure, the forest block crown height generated by aerial remote sensing and the adjacent forest block canopy density are called, the graph attention network is adopted, the node pair data is read one by one, the crown height and canopy density values of the two nodes are extracted, the height difference and canopy density difference between the nodes are calculated, the spatial adjacency verification is performed combined with the geographical location of the block, the node pairs not meeting the connectivity relationship are removed, and the difference matrix is generated; The weight evaluation submodule: based on the difference matrix, the intervals are divided according to the difference value range and the occurrence frequency of each interval is counted, the weighted ratio is calculated combined with the spatial distribution characteristics of the node pairs, the weight is distributed to each connected edge, the weight values of all connected edges are extracted to form a list, the weight ranking is marked according to the weight size, and the weight sequence is generated; Edge aggregation submodule: based on the weight sequence, the edge with a high weight is screened, the node index and the weight value are combined as edge weight mapping, the edge set is established and the redundant edges are removed, the node number is integrated to generate a weighted connection relationship matrix, and a dynamic weight matrix is obtained.

4. The forest resource dynamic monitoring system according to claim 3, characterized in that, The graph attention network first reads the crown height and canopy density values of the nodes corresponding to each edge based on the spatio-temporal graph structure, then maps the node features to a unified embedding space through a shared linear transformation module, extracts the mapped features according to the edge pairs, and then inputs them into a learnable attention mechanism module to calculate the feature similarity score between nodes. After normalization, the attention weight is generated. Finally, the adjacent node features are aggregated to the target node according to the attention weight, and the updated node representation is output.

5. The system according to claim 3, wherein, The graph attention network is according to the formula: wherein: Δij' represents the comprehensive difference value between the forest plot i and the forest plot j, α represents the weight coefficient of the crown height difference value term, β represents the weight coefficient of the canopy density difference value term, γ represents the weight coefficient of the Euclidean space distance term, δ represents the weight coefficient of the remote sensing collection time difference term, ε represents the weight coefficient of the environmental factor difference value term, hi represents the crown height value of the forest plot i, hj represents the crown height value of the forest plot j, ci represents the canopy density value of the forest plot i, cj represents the canopy density value of the forest plot j, dij represents the Euclidean space distance between the forest plot i and the forest plot j, represents the collection time difference of the corresponding remote sensing data between the forest plot i and the forest plot j, and eij represents the normalized environmental factor difference value between the forest plot i and the forest plot j.

6. The system according to claim 1, wherein, The collaborative updating module comprises: Parameter integration submodule: based on the dynamic weight matrix, the tree species distribution statistics and community structure parameters generated by the forestry resource survey database are called, the graph neural network is adopted, the tree species proportion, diameter class and community vertical structure index in each regional sample plot are imported, the field format is unified, and the multi-dimensional parameter set across regions is formed by merging according to the regional dimension, and a parameter summary table is generated; Difference calculation submodule: based on the parameter summary table, the same type of parameters in different regions are compared field by field, the maximum and minimum differences are calculated and recorded, the historical biomass change coefficient is analyzed to analyze the update frequency of each region, and the fields with a difference greater than the average difference are marked as difference key items to generate a difference identification set; Parameter replacement submodule: based on the difference identification set, the marked fields are sequentially differentiated from the original parameters, the parameter items with a difference value lower than a set range are locked and positioned in the corresponding position in the unified parameter set, the original parameters are replaced with unified values, the updated comprehensive parameter set is reconstructed, and the unified model parameters are established.

7. The system according to claim 6, wherein, The graph neural network first reads the tree species proportion, diameter class and community vertical structure index of each regional sample plot, constructs a node feature vector, and imports edge information to form a graph structure according to the regional boundary relationship, then performs feature embedding on each node based on the graph structure, converts the original features to low-dimensional representations through a shared mapping mechanism, in each iteration, the feature vectors of adjacent nodes are weighted and aggregated to the target node through an information transmission mechanism between adjacent nodes, local structure information fusion is realized, and the node representation ability is improved through a nonlinear transformation. Repeat the propagation and aggregation process for multiple rounds to obtain the updated representation of each node considering the structural differences between regions.

8. The system according to claim 6, wherein, The graph neural network is according to the formula: wherein: denotes the multi-dimensional forest structure fusion evaluation value of region r, μ denotes the weight coefficient of the tree species proportion index, pr denotes the tree species composition proportion of the sample plot within region r, v denotes the weight coefficient of the diameter class distribution index, dr denotes the diameter class structure distribution of the sample plot within region r, ξ denotes the weight coefficient of the community vertical structure index, vr denotes the community vertical stratification index of the sample plot within region r, denotes the weight coefficient of the community complexity index, cr denotes the community structure complexity index of the sample plot within region r, κ denotes the weight coefficient of the Shannon diversity index, sr denotes the Shannon tree species diversity index of region r, λ denotes the weight coefficient of the biomass density index, br denotes the forest biomass density per unit area of region r, φ denotes the weight coefficient of the landscape fragmentation index, fr denotes the landscape fragmentation index of region r, ψ denotes the weight coefficient of the terrain ruggedness index, tr denotes the terrain ruggedness value of region r. 9.The forest resource dynamic monitoring system according to claim 1, characterized in that, The trend analysis module comprises: Data aggregation submodule: based on the unified model parameters, the crown coverage rate interpreted from monthly remote sensing and the air temperature and humidity collected on the ground are called, the monthly image coverage rate file and the meteorological monitoring log are compared, the coverage rate and temperature and humidity difference at the same time point are calculated and recorded, the difference value is summed up by adding up the difference value at a fixed window interval, and a cumulative difference value set is generated; Trend filtering submodule: based on the cumulative difference value set, set the sliding window length to gradually intercept the difference value sequence, compare the window sum with the overall average value, collect the out-of-limit plot identifiers and convert the index, and obtain a growth trend list. 10.The forest resource dynamic monitoring system according to claim 1, characterized in that, The anomaly determination module comprises: Difference calculation submodule: based on the growth trend list and unified model parameters, the measured canopy coverage and the model predicted coverage are called, the corresponding time coverage records in the monitoring database and the prediction results are subtracted piece by piece, the absolute values are converted and stored, and the coverage difference set is generated; Abnormal screening submodule: based on the coverage difference set, the difference values are arranged in descending order, compared with the set threshold, the over-limit plot ID is extracted and sorted, and the abnormal plot set is obtained.

Images