A method and system for automatically extracting based on natural resource comprehensive investigation data

Abstract

This invention discloses a method and system for automatic extraction based on comprehensive natural resource survey data, comprising the following steps: Step 1: Unifying coordinates, resolution, and time stamps to establish block indexes and overlapping area indexes; Step 2: Encoding the comprehensive survey data prior to generate prior token sequences and confidence sequences; Step 3: Performing improved SegFormer inference on remote sensing image blocks to generate block probability maps; Step 4: Seamlessly stitching the block results based on the overlapping area index to generate a full-map probability map; Step 5: Performing consistent projection on the full-map probability map based on survey rules; Step 6: Decoding to generate a planar map structure; Step 7: Constructing patch vectors to achieve automatic extraction of natural resource elements. This invention improves the topological consistency of patch vectors.

Description

Technical Field

[0001] This invention relates to the field of geographic information technology, and in particular to a method and system for automatic extraction based on comprehensive natural resource survey data. Background Technology

[0002] Comprehensive natural resource surveys typically rely on the integration of remote sensing image interpretation and comprehensive survey data to complete element identification, boundary delineation, and vectorization of map features. Remote sensing images provide continuous spatial observation information, while comprehensive survey data provides structured prior information such as vector boundaries, attribute ledgers, field sampling points, and topographic factors. In actual production, common processes include coordinate and resolution unification, time stamp alignment, block processing and full-frame stitching, rule verification, and vector output. In existing technologies, remote sensing image interpretation often uses deep learning semantic segmentation networks to perform pixel-level classification of images, and then obtains full-frame results through stitching and post-processing. At the same time, survey rules are used to constrain and correct the classification results, and map feature boundaries are generated based on contour extraction, skeletonization, or graph structure construction to meet the expression requirements of natural resource element results.

[0003] Existing technologies generally suffer from insufficient uniformity in the fusion of remote sensing imagery and comprehensive survey data. Comprehensive survey data often uses vector boundaries and ledger fields as the core carriers, while remote sensing images use raster pixels as the core carriers. There is a lack of a block index and overlapping area index system with geographic location coding as the link, which leads to unstable mapping relationships of prior information in the block space. Prior data is difficult to form prior token expressions that can be directly used in model inference, and differences in prior reliability are also difficult to modulate through confidence. Taking deep learning segmentation networks as an example, the inference process relies more on image texture and spectral features. When encountering objects with the same spectrum, shadow occlusion, seasonal changes, thin linear features, and low-contrast boundaries, category confusion and boundary drift are likely to occur. In addition, there is a lack of a routing mechanism with evidence residuals as trigger signals, which makes it difficult for attention calculation and prior fusion to focus on conflict areas in space. The contradictions between the block probability map and prior information are difficult to be identified and resolved in a timely manner.

[0004] Existing technologies also suffer from link breakage issues in the block stitching, rule-consistent projection, and vector map generation stages. Block stitching results are usually completed by simply weighting or directly covering overlapping areas, lacking a seamless stitching memory library to continuously remember and align the features and probabilities of overlapping areas. This easily leads to probability abrupt changes and boundary breaks in overlapping areas, which can then be magnified to the full probability map. Rule constraints are often implemented by discrete rule filtering or hard threshold repair, making it difficult to form a differentiable update process for category assignment variables and area constraints at the rule-consistent projection layer. This results in unstable coupling between rule correction and model output. Furthermore, the vectorization process relies heavily on morphology and contour tracking to obtain closed boundaries, but lacks a unified expression that simultaneously completes node prediction and edge prediction and constructs a planar map at the planar graph decoder level. Closure constraints, non-self-intersection constraints, and shared edge consistency constraints are difficult to execute collaboratively on the same graph structure, easily generating geometric problems such as self-intersection, gaps, and inconsistent shared edges, affecting the topological consistency of the vector map results and the completeness of automatic feature extraction.

[0005] Therefore, how to provide a method and system for automatic extraction based on comprehensive natural resource survey data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method and system for automatic extraction of natural resource element patches based on comprehensive natural resource survey data. This invention constructs a unified geographic location coding system for remote sensing images and comprehensive survey data, introduces prior token sequences and prior confidence sequences, and combines improved SegFormer evidence residual-driven reasoning, seamless splicing memory mechanism, rule-consistent differentiable projection, and planar map constraint decoding to achieve automatic generation and vectorized expression of natural resource element patches. It has the advantages of stable prior fusion, strong boundary continuity, high rule consistency, and high accuracy of automatic element extraction.

[0007] A method for automatically extracting data based on comprehensive natural resource survey data according to an embodiment of the present invention includes the following steps:

[0008] Step 1: Perform coordinate unification, resolution unification, and time stamp alignment on remote sensing images and comprehensive survey data, generate geographic location codes, and establish block indexes and overlapping area indexes;

[0009] Step 2: Perform prior coding on the comprehensive survey data to generate a prior token sequence and a prior confidence sequence, and establish a mapping relationship between the prior token sequence and the block index;

[0010] Step 3: Perform improved SegFormer encoding inference on the remote sensing image blocks, calculate the evidence residual map, trigger evidence routing attention to read the prior token sequence according to the evidence residual map, and generate block features and block probability maps;

[0011] Step 4: Establish a seamless stitching memory based on the overlap region index, perform alignment and fusion on the block features and block probability maps to generate full-frame features and full-frame probability maps;

[0012] Step 5: Perform a differentiable projection update on the full probability map at the rule consistency projection layer, investigate the set of rules to limit the category assignment variables and area constraints, and generate the projected probability map;

[0013] Step 6: Perform node prediction and edge prediction on the projected probability map and full-frame features in the planar graph decoder to generate node set and edge set, and construct planar graph;

[0014] Step 7: Based on the planar graph, apply closure constraints, no self-intersection constraints, and shared edge consistency constraints to generate map patch vectors and complete the automatic extraction of natural resource elements.

[0015] Optionally, step one specifically includes:

[0016] Coordinate and resolution unification is performed on remote sensing images and comprehensive survey data to generate raster row and column indexes. The comprehensive survey data is then mapped and written into raster attribute fields according to the raster row and column indexes. The raster attribute fields include vector boundary identifiers, attribute ledger field values, field sampling point identifiers, and topographic factor field values. The vector boundary identifier is the association identifier of the vector boundary record on the raster row and column index. The attribute ledger field value consists of a field identifier and a field value. The field identifier is the attribute ledger field name code, and the field value is the value corresponding to the attribute ledger field name. The field sampling point identifier is the association identifier of the field sampling point record on the raster row and column index. The topographic factor field value is the numerical value corresponding to the topographic factor field name.

[0017] Time stamp alignment is performed between the time stamps of remote sensing images and the time stamps of comprehensive survey data. A baseline time axis is established and a time index is generated. The baseline time axis is determined by the baseline time start point and the time step, and the time index is determined by the position of the time stamp on the baseline time axis.

[0018] Geolocation codes are generated based on raster row and column indexes, and block indexes and overlap indexes are established based on geolocation codes. Block indexes are used to identify the blocks corresponding to geolocation codes, and overlap indexes are used to identify the set of geolocation codes where blocks overlap.

[0019] Optionally, step two specifically involves:

[0020] Based on the block index, determine the set of raster row and column indexes corresponding to the block coverage area, merge the raster attribute fields according to the time index, and collect the vector boundary identifier, attribute ledger field value, field sampling point identifier, and terrain factor field value in the raster attribute fields to form a block prior record set, with the block prior record set using the block index as the record key;

[0021] Prior coding is performed on the block prior record set, mapping the vector boundary identifier to the boundary embedding vector, the attribute ledger field value to the ledger embedding vector, the field sampling point identifier to the sampling embedding vector, and the terrain factor field value to the terrain embedding vector. The prior token sequence is generated by splicing and mapping the data in a preset order.

[0022] A prior confidence sequence is generated based on the block prior record set. The prior confidence sequence is calculated and normalized based on the data source identifier, time index consistency identifier, and spatial coverage identifier. A mapping relationship is established between the prior token sequence and the block index. The mapping relationship associates the block index, the prior token sequence, and the prior confidence sequence.

[0023] Optionally, step three specifically includes:

[0024] Based on the block index, remote sensing image blocks are extracted from the remote sensing image, and the prior token sequence and prior confidence sequence are read based on the block index.

[0025] The improved SegFormer performs encoding on remote sensing image blocks to obtain encoded features, and performs decoding based on the encoded features to obtain an initial block probability map. The initial block probability map contains probability values ​​arranged by category at pixel positions.

[0026] Prior category scores are calculated based on prior token sequences, and the prior category scores are mapped to the remote sensing image block space according to the block index to generate a prior guided probability map. The prior guided probability map contains probability values ​​arranged by category at the pixel position.

[0027] The evidence residual map is calculated based on the initial block probability map and the prior guided probability map. The evidence residual value at the pixel position is obtained by summing the absolute values ​​of the probability differences in the category dimension.

[0028] An evidence routing mask is generated based on the evidence residual map and the residual threshold. The evidence routing mask is marked as valid at positions where the evidence residual value is greater than the residual threshold, and as invalid at positions where the evidence residual value is less than or equal to the residual threshold.

[0029] Evidence routing attention is performed at the valid location marked by the evidence routing mask. Evidence routing attention uses the encoded features as the query vector, the prior token sequence as the key vector and value vector to generate attention output features, and the attention output features are weighted by the confidence weights obtained by mapping the prior confidence sequence.

[0030] The encoded features are fused with the weighted attention output features to generate block features, and the initial block probability map is updated based on the block features to generate a new block probability map.

[0031] Optionally, step four specifically involves:

[0032] The write position of the block feature and block probability map in the full feature and full probability map is determined based on the block index. The write position is determined by the raster row and column index obtained by the geolocation coding mapping associated with the block index.

[0033] A seamless splicing memory is built based on the overlapping area index. Memory entries are built according to the overlapping area index. The memory entries contain the set of geographic location codes associated with the overlapping area index, the overlapping area feature tensor, and the overlapping area probability tensor.

[0034] The block features and block probability maps are written to the full features and full probability maps according to the writing position. The non-overlapping area index coverage range within the block index coverage range is covered by the overwrite write.

[0035] Within the coverage area of ​​the overlapping area index, the block features and block probability maps are cropped based on the geographic location encoding set associated with the overlapping area index to obtain aligned block features and aligned block probability maps. The overlapping area feature tensor and overlapping area probability tensor are read from the seamless splicing memory based on the overlapping area index.

[0036] The alignment block feature and the overlapping area feature tensor are weighted and summed position by position according to the fusion weight to generate the fused overlapping area feature. The alignment block probability map and the overlapping area probability tensor are weighted and summed position by position according to the fusion weight to generate the fused overlapping area probability. The fusion weight is determined and normalized by the relative position from the raster row and column index to the block index boundary within the overlapping area index coverage area.

[0037] The fused overlapping area features and fused overlapping area probabilities are written into the overlapping area index coverage of the full-scale feature and full-scale probability maps, and the fused overlapping area features and fused overlapping area probabilities are written into the corresponding memory entries of the seamless splicing memory.

[0038] Optionally, step five specifically includes:

[0039] The full probability map is variableized at the rule-consistent projection layer to generate category-assigned variables. The category-assigned variables contain probability components arranged by category at the raster row and column index positions.

[0040] The rule coverage of the survey rule set is determined based on geolocation coding. Within the rule coverage, category constraints and area constraints are applied to the category allocation variables. Category constraints limit the set of category values ​​for the category allocation variables, and area constraints limit the category area statistics of the category allocation variables within the rule coverage.

[0041] Constraint violation is calculated based on category constraints and area constraints. The constraint violation includes category constraint violation and area constraint violation. Category constraint violation is obtained by summing the probability components of the set of violation category values ​​within the rule coverage area. Area constraint violation is obtained by summing the statistical values ​​of category area and the difference between area constraints within the rule coverage area.

[0042] Differentiable projection updates are performed on the category assignment variables based on the constraint violation amount. The differentiable projection update includes truncating the probability components of the set of violation category values ​​and normalizing the category dimension, and scaling the category probability components of the violation area constraint and normalizing the category dimension. The update terminates when the constraint violation amount is less than a threshold.

[0043] The class assignment variables updated by differentiable projection are reconstructed to generate a projective probability map.

[0044] Optionally, step six specifically includes:

[0045] The projection probability map and the full-frame feature are input into the planar graph decoder to generate a node score map on the full-frame feature. Peak extraction is performed on the node score map to obtain node candidate coordinates. The nodes are sorted based on their scores and filtered based on the minimum spacing constraint to obtain a node set. The node set contains node identifiers and node coordinates.

[0046] In the planar graph decoder, an edge score tensor is generated based on the full-frame features. The edge score tensor uses the node identifier pairs of the node set as indices, and the edge score tensor takes the edge score of the node identifier pair.

[0047] An edge set is generated by performing threshold filtering based on the edge score tensor and node identifier. The edge set includes edge identifier, start node identifier, and end node identifier.

[0048] Adjacency records are generated based on the set of nodes and the set of edges. The adjacency records are associated with edge identifiers using node identifiers to construct a planar graph.

[0049] Optionally, step seven specifically includes:

[0050] Based on the planar graph, a closed boundary chain is generated by tracking and recording the adjacency relationship in the edge set. The closed boundary chain satisfies that the starting node identifier and the ending node identifier are consistent. The closed boundary chain is written into the closed loop set.

[0051] Perform a no-self-intersection constraint check on the set of closed loops. The no-self-intersection constraint check performs intersection detection on the edge segments contained in the closed boundary chain. If an intersection is detected, delete the closed boundary chain where the intersection is detected.

[0052] Perform shared edge consistency constraint verification on the closed loop set. The shared edge consistency constraint verification extracts the shared edges based on the edge identifier and determines the consistency between the start node identifier and the end node identifier of the shared edge. If the consistency determination is not satisfied, delete the closed boundary chain that does not satisfy the consistency determination.

[0053] Based on the closed loop set that passes the self-intersection constraint verification and the shared edge consistency constraint verification, a patch vector is generated. The patch vector contains patch identifiers and vertex sequences. The vertex sequences are formed by arranging the coordinates of the corresponding nodes of the closed boundary chain in the order of the closed boundary chain, thus completing the automatic extraction of natural resource elements.

[0054] A system for automatic extraction based on comprehensive natural resource survey data according to an embodiment of the present invention includes:

[0055] The data alignment index module is used to perform coordinate unification, resolution unification, and time stamp alignment on remote sensing images and comprehensive survey data, generate geolocation codes, establish block indexes and overlapping area indexes, and generate raster row and column indexes, baseline time axis, time index, and raster attribute fields.

[0056] The prior coding module is used to perform prior coding on the comprehensive survey data, generate a prior token sequence and a prior confidence sequence, and establish a mapping relationship between the prior token sequence and the block index.

[0057] The block interpretation module is used to perform improved SegFormer encoding inference on remote sensing image blocks, calculate evidence residual maps, trigger evidence routing attention to read prior token sequences according to the evidence residual maps, and generate block features and block probability maps.

[0058] The seamless stitching module is used to build a seamless stitching memory based on the overlap area index, perform alignment and fusion on the block features and block probability maps, and generate full-frame features and full-frame probability maps.

[0059] The rule projection module is used to perform differentiable projection updates on the full-scale probability map at the rule consistency projection layer, investigate the rule set to limit category assignment variables and area constraints, and generate the projected probability map;

[0060] The planar graph decoding module is used to perform node and edge prediction on the projected probability map and full-frame features in the planar graph decoder, generate node and edge sets, and construct a planar graph.

[0061] The feature generation module is used to generate feature vectors based on planar maps by applying closure constraints, no self-intersection constraints, and shared edge consistency constraints, thereby completing the automatic extraction of natural resource elements.

[0062] The beneficial effects of this invention are:

[0063] This invention achieves coordinate unification, resolution unification, and time stamp alignment between remote sensing imagery and comprehensive survey data, constructing a data organization method centered on geographic location coding. This enables comprehensive survey data to be accurately mapped to the remote sensing image space in the form of raster attribute fields. Spatially consistent data alignment is achieved under the constraints of block indexing and overlapping area indexing. Combined with explicit modeling methods of prior token sequences and prior confidence sequences, survey attributes, field sampling information, and topographic factors participate in semantic reasoning during the interpretation stage. This fundamentally improves the ability to express attribute consistency and spatial constraints in the process of natural resource element identification, avoiding the category confusion and boundary drift problems caused by relying solely on image texture features.

[0064] This invention introduces an evidence routing attention mechanism based on evidence residual maps in the segmented interpretation stage. This allows the improved SegFormer to introduce prior information only at locations where there is a significant deviation between prior and image reasoning, effectively suppressing prior interference and enhancing the ability to distinguish key areas. Combined with a seamless splicing memory, it enables continuous fusion of segmented results within overlapping areas. Furthermore, through a rule-consistent projection layer and planar map decoding constraints, survey rules, area constraints, and geometric consistency are directly integrated into the result generation process. This achieves boundary closure, topological consistency, and map vector output that meets survey specifications, significantly improving the overall consistency, structural reliability, and practical usability of the automatic extraction results of natural resource elements. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 This is a flowchart of a method for automatic extraction based on comprehensive natural resource survey data proposed in this invention;

[0067] Figure 2 This is a schematic diagram of the improved SegFormer block interpretation method for automatic extraction of natural resource comprehensive survey data proposed in this invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figures 1-2 A method for automatically extracting data based on comprehensive natural resource survey data includes the following steps:

[0070] Step 1: Perform coordinate unification, resolution unification, and time stamp alignment on remote sensing images and comprehensive survey data, generate geographic location codes, and establish block indexes and overlapping area indexes;

[0071] Step 2: Perform prior coding on the comprehensive survey data to generate a prior token sequence and a prior confidence sequence, and establish a mapping relationship between the prior token sequence and the block index;

[0072] Step 3: Perform improved SegFormer encoding inference on the remote sensing image blocks, calculate the evidence residual map, trigger evidence routing attention to read the prior token sequence according to the evidence residual map, and generate block features and block probability maps;

[0073] Step 4: Establish a seamless stitching memory based on the overlap region index, perform alignment and fusion on the block features and block probability maps to generate full-frame features and full-frame probability maps;

[0074] Step 5: Perform a differentiable projection update on the full probability map at the rule consistency projection layer, investigate the set of rules to limit the category assignment variables and area constraints, and generate the projected probability map;

[0075] Step 6: Perform node prediction and edge prediction on the projected probability map and full-frame features in the planar graph decoder to generate node set and edge set, and construct planar graph;

[0076] Step 7: Based on the planar graph, apply closure constraints, no self-intersection constraints, and shared edge consistency constraints to generate map patch vectors and complete the automatic extraction of natural resource elements.

[0077] In this embodiment, step one specifically includes:

[0078] Coordinate and resolution unification are performed on remote sensing imagery and comprehensive survey data. Coordinate unification includes selecting a unified coordinate reference system and transforming the pixel center coordinates of the remote sensing imagery and the geometric coordinates of the comprehensive survey data to the unified coordinate reference system. Resolution unification includes selecting a target resolution and resampling the remote sensing imagery, discretizing the spatial extent under the unified coordinate reference system into rasters, and establishing raster row and column indices. The raster row and column indices are determined by the raster origin coordinates, target resolution, row direction index, and column direction index. The comprehensive survey data is mapped and written into the raster attribute fields according to spatial coverage relationships. Mapping and writing includes mapping vector boundaries according to coverage relationships. The raster row and column index of the cover is written to the vector boundary identifier. The attribute ledger is written to the attribute ledger field value according to spatial association. The field sampling point is written to the field sampling point identifier according to the raster row and column index it falls into. The terrain factor is resampled according to the target resolution and written to the terrain factor field value according to the raster column index. The vector boundary identifier is used to associate the vector boundary record. The attribute ledger field value consists of the field identifier and the field value. The field identifier is the attribute ledger field name code. The field value is the value corresponding to the field identifier. The field sampling point identifier is used to associate the field sampling point record. The terrain factor field value is the numerical value corresponding to the terrain factor field name.

[0079] Time stamp alignment is performed between the time stamps of remote sensing images and the time stamps of comprehensive survey data. Time stamp alignment includes establishing a reference time axis and determining the reference time start and time step. The time stamps of remote sensing images and comprehensive survey data are mapped to the reference time start and time step to generate a time index. The time index is associated with the raster row and column index, written into the record, and used for time stamp alignment retrieval.

[0080] Geolocation codes are generated based on the raster row and column indexes, and block indexes and overlap indexes are established based on the geolocation codes. The geolocation codes are calculated from the row direction index, column direction index, and the number of raster columns. The block index is determined by the geolocation codes and the block size and is used to identify the block spatial range. The overlap index is determined by the block spatial range and the overlap width and records the set of geolocation codes covered by the overlap area.

[0081] In this embodiment, step two specifically includes:

[0082] Based on the block index, the set of raster row and column indexes corresponding to the block coverage area is determined. The set of raster row and column indexes consists of all raster row and column indexes within the block space indicated by the block index. The raster attribute fields are merged according to the time index. The merging includes grouping the raster attribute field records according to the time index and performing aggregation within the group. The vector boundary identifiers are aggregated according to the coverage raster row and column index. The attribute ledger field values ​​are aggregated according to the field identifier. The field sampling point identifiers are aggregated according to the landing point raster column index. The terrain factor field values ​​are aggregated according to the raster column index, forming a block prior record set. The block prior record set uses the block index as the record key.

[0083] Prior coding is performed on the block prior record set. Prior coding includes establishing a boundary embedding mapping for vector boundary identifiers and outputting boundary embedding vectors, establishing a ledger embedding mapping for attribute ledger field values ​​according to field identifiers and outputting ledger embedding vectors, establishing a sampling embedding mapping for field sampling point identifiers and outputting sampling embedding vectors, establishing a terrain embedding mapping for terrain factor field values ​​according to terrain factor field names and outputting terrain embedding vectors, and concatenating the boundary embedding vector, ledger embedding vector, sampling embedding vector, and terrain embedding vector in a preset order. The concatenated vectors are mapped to generate a prior token sequence.

[0084] A prior confidence sequence is generated based on a block-based prior record set. The generation of the prior confidence sequence includes associating a data source identifier, a time index consistency identifier, and a spatial coverage identifier with each prior token in the prior token sequence, and calculating a confidence value. The data source identifier is determined by the source of comprehensive survey data. The time index consistency identifier is determined by the matching relationship between the time index of the block-based prior record set and the time index within the block coverage area. The spatial coverage identifier is determined by the coverage relationship of the raster row and column index set corresponding to the block coverage area. The confidence value is normalized to generate a prior confidence sequence. A mapping relationship between the prior token sequence and the block index is established. The mapping relationship records the set of positions of the prior token sequence corresponding to the block index and associates it with the prior confidence sequence.

[0085] In this embodiment, step three specifically includes:

[0086] Based on the block index, remote sensing image blocks are extracted from the remote sensing image. The remote sensing image blocks are composed of pixel values ​​corresponding to the raster row and column indices within the block space indicated by the block index. The prior token sequence and prior confidence sequence are read based on the block index. The reading includes locating the prior token sequence position set by the mapping relationship between the prior token sequence and the block index and retrieving the token vector corresponding to the position set. At the same time, the prior confidence sequence value corresponding to the same position set is retrieved.

[0087] The improved SegFormer performs encoding on remote sensing images in blocks to obtain encoded features. The encoded features maintain the pixel position correspondence within the block space. Based on the encoded features, decoding is performed to obtain an initial block probability map. The initial block probability map contains probability values ​​arranged by category at pixel positions, and the probability values ​​satisfy normalization constraints in the category dimension.

[0088] The prior category score is calculated based on the prior token sequence. The prior category score is obtained by linearly mapping the prior token sequence to the category dimension score vector. The prior category score is then mapped to the remote sensing image block space according to the block index to generate a prior guidance probability map. The mapping includes broadcasting the prior category score to the pixel positions within the block space and normalizing it according to the category dimension to obtain the pixel position probability value. The prior guidance probability map contains probability values ​​arranged by category at the pixel position.

[0089] The evidence residual map is calculated based on the initial block probability map and the prior guided probability map. The evidence residual value at the pixel position is obtained by subtracting the category dimension probability values ​​of the initial block probability map and the category dimension probability values ​​of the prior guided probability map category by category, taking the absolute value, and then summing them in the category dimension.

[0090] An evidence routing mask is generated based on the evidence residual map and the residual threshold. The value of the evidence routing mask at the pixel position is obtained by comparing the evidence residual value with the residual threshold. If the evidence residual value is greater than the residual threshold, it is set as a valid value. If the evidence residual value is less than or equal to the residual threshold, it is set as an invalid value.

[0091] Evidence routing attention is performed at the valid locations marked by the evidence routing mask. Evidence routing attention uses the vector corresponding to the location of the encoded feature pixel as the query vector, and uses the token vector of the prior token sequence as the key vector and the value vector to calculate the attention weight. The value vector is then weighted and summed to obtain the attention output feature. The prior confidence sequence is mapped to obtain the confidence weight, which is then multiplied by the attention weight to obtain the weighted attention weight. The weighted attention weight is used to calculate the weighted attention output feature.

[0092] The encoded features and the weighted attention output features are fused to generate block features. The fusion includes performing a concatenated linear mapping on the encoded features and the weighted attention output features while maintaining the correspondence between pixel positions. The block probability map is then updated based on the block features to generate a new block probability map. The update includes inputting the block features into the classification mapping layer to obtain the category dimension update score and performing normalization to obtain the block probability map.

[0093] This invention introduces prior token sequences and prior confidence sequences on top of SegFormer. It uses linear mapping to obtain prior class scores and generate a prior guided probability map. Then, it constructs an evidence residual map based on the difference between the initial block probability map and the prior guided probability map. An evidence routing mask is formed based on the residual threshold, triggering evidence routing attention only at locations with large residuals. Encoded features are used as query vectors, and the prior token sequence is used as key and value vectors to calculate attention output features. Simultaneously, a prior confidence sequence is used to generate confidence weights for weighting. The weighted attention output features and encoded features are fused to obtain block features and update the block probability map. Compared to the unmodified SegFormer's decoding method that relies solely on image texture, this invention achieves prior constraint-driven differentiated attention and confidence modulation, reducing class confusion and spurious responses, improving boundary continuity and spatial consistency, and reducing ineffective attention calculations under the constraint of the residual mask to maintain inference stability. This makes the block probability map and the prior guided probability map complementary, enhancing the ability to distinguish low-contrast features, thin linear elements, and noisy scenes, mitigating probability abrupt changes at stitching boundaries, and providing more reliable input to the full-frame probability map.

[0094] In this embodiment, step four specifically includes:

[0095] The block index determines the writing position of the block features and block probability maps in the full-size features and full-size probability maps. The writing position determination includes locating the geographic location code set corresponding to the block spatial range by the block index, and mapping the geographic location code set to a raster row and column index set. The raster row and column index set is used to identify the block writing area in the full-size features and full-size probability maps.

[0096] A seamless splicing memory is established based on the overlapping area index. The memory includes generating memory entries according to the overlapping area index and using the overlapping area index as the retrieval key. The memory entries contain the overlapping area index-associated geographic location encoding set, the overlapping area feature tensor, and the overlapping area probability tensor. The spatial dimensions of the overlapping area feature tensor and the overlapping area probability tensor correspond to the overlapping area index-associated geographic location encoding set.

[0097] The block features and block probability maps are written to the full-frame features and full-frame probability maps according to the writing position. The writing includes writing the raster row and column index positions corresponding to the geolocation encoding set within the block index coverage area into the block feature pixel position vector and the pixel position category probability of the block probability map. The writing of the non-overlapping area index coverage area within the block index coverage area adopts the overlay writing rule.

[0098] Within the coverage area of ​​the overlapping area index, the block features and block probability maps are cropped based on the overlapping area index and the associated geolocation coding set to obtain aligned block features and aligned block probability maps. The cropping includes locating the pixel position vectors of the block features and the pixel position category probabilities of the block probability maps according to the geolocation coding set and forming a cropping tensor. The overlapping area feature tensor and the overlapping area probability tensor are read from the seamless splicing memory based on the overlapping area index. The memory entries are retrieved using the overlapping area index and the memory entry storage tensor is output.

[0099] The alignment block feature and the overlapping area feature tensor are weighted and summed position by position according to the fusion weight to generate the fused overlapping area feature. The alignment block probability map and the overlapping area probability tensor are weighted and summed position by position according to the fusion weight to generate the fused overlapping area probability. The fusion weight establishes boundary distance values ​​for the raster row and column indices within the coverage area of ​​the overlapping area index and normalizes the boundary distance values. The boundary distance values ​​are determined by the minimum distance from the raster row and column index to the block index boundary.

[0100] The fused overlapping area features and fused overlapping area probabilities are written into the overlapping area index coverage of the full-map feature and full-map probability map. The location is written using the raster row and column index mapping of the geographic location encoding set associated with the overlapping area index. The fused overlapping area features and fused overlapping area probabilities are written into the corresponding memory entries in the seamless splicing memory bank. The memory entries are retrieved using the overlapping area index and the memory entry storage tensor is updated and overwritten.

[0101] In this embodiment, step five specifically includes:

[0102] In the rule-consistent projection layer, the full probability map is subjected to variable processing to generate category assignment variables. Variable processing includes expanding the full probability map by raster row and column indices while maintaining the one-to-one correspondence between raster row and column indices and probability vectors. The category assignment variables contain probability components arranged by category at the raster row and column index positions and satisfy the category dimension normalization constraint.

[0103] The rule coverage of the survey rule set is determined based on geolocation coding. The rule coverage is obtained by mapping geolocation coding to the set of raster row and column indices. Within the rule coverage, category constraints and area constraints are applied to the category assignment variables. Category constraints limit the set of category values ​​of the category assignment variables by allowing the set of category identifiers. Area constraints limit the category area statistics of the category assignment variables within the rule coverage by the lower and upper area bounds. The category area statistics are obtained by summing the category probability components of the raster row and column index positions within the rule coverage and converting them according to the target resolution.

[0104] The constraint violation amount is calculated based on category constraints and area constraints. The category constraint violation amount is obtained by summing the category probability components that do not belong to the set of allowed category identifiers within the rule coverage area. The area constraint violation amount is obtained by summing the deviations of the category area statistics from the lower and upper bounds of the area. The constraint violation amount is obtained by weighted summation of the category constraint violation amount and the area constraint violation amount.

[0105] Based on the constraint violation, the category assignment variable is updated using a differentiable projection. The differentiable projection update includes truncating the category probability components that do not belong to the set of allowed category identifiers and renormalizing the category dimension. It also includes scaling the category probability components whose category area statistics exceed the lower or upper bound of the area and renormalizing the category dimension. The update process is performed iteratively, and the constraint violation is recalculated after each update. The update terminates when the constraint violation is less than the threshold.

[0106] The differentiable projection updated category assignment variables are reconstructed to generate a projected probability map. The reconstruction includes writing the category assignment variable probability vector back to the full-scale spatial location according to the raster row and column index while maintaining the consistent category arrangement order.

[0107] In this embodiment, step six specifically includes:

[0108] The projected probability map and the full-frame feature are input into the planar map decoder to generate a node score map on the full-frame feature. The node score map outputs the node confidence component at the raster row and column index position. The node candidate coordinates are determined by the local maximum position of the node score map in two-dimensional space. The node score is obtained by taking the value of the node score map at the node candidate coordinate position. The nodes are sorted in descending order based on the node score and deduplication is performed based on the minimum spacing constraint. The minimum spacing constraint requires that the Euclidean distance between the node candidate coordinates is not less than the spacing threshold. The filtering results are written into the node set, which is a sequence of node records. The node record contains the node identifier and the node coordinates.

[0109] In the planar graph decoder, an edge score tensor is generated based on the full-frame features. The edge score tensor uses node identifier pairs as indices. A node identifier pair consists of any two node identifiers in the node set. The value of the edge score tensor is calculated by the planar graph decoder based on the feature response of the direction of the connection line between the corresponding node coordinates of the node identifier pair. The edge score tensor outputs the edge score according to the node identifier pair.

[0110] Based on the edge score tensor, a threshold filtering is performed on the node identifier pair to generate an edge set. The threshold filtering limits the node identifier pair with the edge score to be written into the edge set. The edge set is a sequence of edge records, and the edge record contains the edge identifier, the starting node identifier, and the ending node identifier.

[0111] Adjacency records are generated based on the node set and edge set. For each node identifier, an associated edge identifier list is established in the adjacency record. The associated edge identifier list is obtained by summing the edge identifiers in the edge set whose starting node identifier or ending node identifier is equal to the node identifier. The planar graph is jointly represented by the node set, edge set, and adjacency record.

[0112] In this embodiment, step seven specifically includes:

[0113] Based on the planar graph, a closed boundary chain is generated by tracing the edge set according to the adjacency record. The tracing includes selecting the edge identifier in the edge set as the starting edge and reading the starting node identifier and the ending node identifier. The ending node identifier is used as the current node identifier, and the list of associated edge identifiers of the current node identifier is retrieved in the adjacency record. The edge identifier that has not been added to the closed boundary chain is selected from the list of associated edge identifiers as the next edge and the current node identifier is updated. The tracing terminates when the current node identifier is equal to the starting node identifier of the closed boundary chain. The closed boundary chain is represented by the edge identifier sequence and the corresponding node identifier sequence and written into the closed loop set.

[0114] Perform a self-intersection constraint check on the set of closed loops. The self-intersection constraint check performs pairwise intersection detection on the edge segments corresponding to the edge identifiers of the closed boundary chain. The edge segments are determined by the coordinates of the starting node and the ending node. The condition for the existence of an intersection point is that the edge segments intersect and the intersection point is not located at the shared endpoint coordinates of the adjacent edge segments. If the intersection point exists, delete the closed boundary chain where the intersection point exists.

[0115] Perform shared edge consistency constraint verification on the closed loop set. The shared edge consistency constraint verification is based on the edge identifier to count the occurrence of the edge identifier and the edge identifier with the occurrence greater than one is identified as the shared edge identifier. For the shared edge identifier, read the start node identifier and end node identifier corresponding to the shared edge identifier and determine the consistency between the start node identifier and the end node identifier. If the consistency determination is not met, it means that there are inconsistent records between the start node identifier and the end node identifier corresponding to the same shared edge identifier. If the consistency determination is not met, delete the closed boundary chain that does not meet the consistency determination.

[0116] Based on the closed loop set that passes the self-intersection constraint verification and the shared edge consistency constraint verification, a patch vector is generated. The patch vector contains patch identifiers and vertex sequences. The patch identifiers are generated by assigning values ​​according to the index order of the closed loop set. The vertex sequence is formed by mapping the node identifier sequence of the closed boundary chain to the node set to obtain the node coordinates and arranging them according to the node identifier sequence order, thus completing the automatic extraction of natural resource elements.

[0117] A system for automatic extraction based on comprehensive natural resource survey data includes:

[0118] The data alignment index module is used to perform coordinate unification, resolution unification, and time stamp alignment on remote sensing images and comprehensive survey data, generate geolocation codes, establish block indexes and overlapping area indexes, and generate raster row and column indexes, baseline time axis, time index, and raster attribute fields.

[0119] The prior coding module is used to perform prior coding on the comprehensive survey data, generate a prior token sequence and a prior confidence sequence, and establish a mapping relationship between the prior token sequence and the block index.

[0120] The block interpretation module is used to perform improved SegFormer encoding inference on remote sensing image blocks, calculate evidence residual maps, trigger evidence routing attention to read prior token sequences according to the evidence residual maps, and generate block features and block probability maps.

[0121] The seamless stitching module is used to build a seamless stitching memory based on the overlap area index, perform alignment and fusion on the block features and block probability maps, and generate full-frame features and full-frame probability maps.

[0122] The rule projection module is used to perform differentiable projection updates on the full-scale probability map at the rule consistency projection layer, investigate the rule set to limit category assignment variables and area constraints, and generate the projected probability map;

[0123] The planar graph decoding module is used to perform node and edge prediction on the projected probability map and full-frame features in the planar graph decoder, generate node and edge sets, and construct a planar graph.

[0124] The feature generation module is used to generate feature vectors based on planar maps by applying closure constraints, no self-intersection constraints, and shared edge consistency constraints, thereby completing the automatic extraction of natural resource elements.

[0125] Example 1:

[0126] To verify the feasibility of this invention in practice, it was applied to an automatic interpretation task for a comprehensive natural resource survey. This task, based on large-scale remote sensing imagery and long-term accumulated comprehensive survey data, aims to achieve automatic identification of natural resource elements, precise boundary positioning, and generation of vector map results. In this application scenario, the survey objects cover common natural resource types such as cultivated land, forest land, grassland, water areas, and construction land. The remote sensing imagery comes from uniformly acquired multispectral imagery data, and the comprehensive survey data comes from existing resource ledgers, field sampling records, and topographic analysis results. Existing technologies generally suffer from inconsistencies between image interpretation results and survey attributes, boundary breaks at patch stitching points, large deviations in area statistics, and high costs of manual post-processing in practical applications, seriously affecting the reliability and efficiency of natural resource survey results.

[0127] In this scenario, the remote sensing imagery and comprehensive survey data are first aligned in terms of coordinates, resolution, and time stamp. This establishes a one-to-one correspondence between image pixels and survey attributes within the same raster row and column indexing system, generating geolocation codes as the spatial basis for subsequent processing. Based on this, vector boundary markers, attribute ledger field values, field sampling point markers, and terrain factor field values ​​from the survey data are written into the raster attribute fields. Block indexes and overlap indexes are then established according to the geolocation codes, dividing the large-scale imagery into regular blocks and clearly defining the overlapping areas of these blocks. Subsequently, prior coding is performed on the comprehensive survey data within each block, mapping the survey attribute information into a priori token sequences. A priori confidence sequence is generated by combining data source consistency and spatial coverage, enabling the survey priors to participate in the remote sensing image interpretation process in a structured form.

[0128] In the segmented interpretation stage, the system encodes and decodes the remote sensing image segments into an improved SegFormer model to obtain an initial segmented probability map. A priori token sequence is introduced into the evidence residual calculation process. By comparing the image inference results with the prior guidance results, evidence routing attention is triggered at pixel locations with significant probability deviations, thus selectively introducing prior information and avoiding global interference with the image discrimination results. The segmented features and segmented probability maps generated through this process are written into a seamless stitching memory and continuously fused according to positional weights within the overlapping area to obtain spatially continuous full-frame features and a full-frame probability map. Subsequently, the full-frame probability map undergoes differentiable projection updates in the rule-consistent projection layer, combining the survey rule set to ensure that the category assignment results meet both category constraints and area statistics requirements. Finally, through planar map decoding and closure constraint processing, topologically consistent patch vector results are generated, enabling automatic extraction of natural resource elements.

[0129] To verify the beneficial effects of this invention, the method of this invention was compared with conventional deep learning interpretation methods that do not introduce prior tokens and evidence routing mechanisms. Key indicators such as automatic interpretation accuracy, map boundary integrity, area error, and manual correction workload were statistically analyzed. The comparison results are shown in Table 1.

[0130] Table 1 Comparison of Automatic Extraction Effects of Natural Resource Elements

[0131] Comparison indicators conventional methods Method of the present invention Pixel-level overall classification accuracy 86.4% 93.1% Average IoU of major land types 0.72 0.84 Image boundary closure rate 88.6% 97.3% Average error of single-class area 6.8% 2.1% Percentage of manual correction time 42% 15% Average number of fractures in the image 1.9 times / square kilometer 0.4 times / square kilometer

[0132] The results show that by introducing a comprehensive survey prior and evidence residual-driven mechanism, the interpretation results significantly improve both the accuracy of category discrimination and the consistency of boundary structure. Particularly in overlapping areas, the continuity of map patch boundaries is significantly improved, essentially eliminating the splicing misalignment and boundary overlap problems commonly found in traditional methods. Furthermore, because the rule-consistent projection layer explicitly controls area constraints, the interpretation results maintain a high degree of statistical consistency with survey specifications, greatly reducing the workload of post-processing and manual verification.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic extraction based on comprehensive natural resource survey data, characterized in that, Includes the following steps: Step 1: Perform coordinate unification, resolution unification, and time stamp alignment on remote sensing images and comprehensive survey data, generate geographic location codes, and establish block indexes and overlapping area indexes; Step 2: Perform prior coding on the comprehensive survey data to generate a prior token sequence and a prior confidence sequence, and establish a mapping relationship between the prior token sequence and the block index; Step 3: Perform improved SegFormer encoding inference on the remote sensing image blocks, calculate the evidence residual map, trigger evidence routing attention to read the prior token sequence according to the evidence residual map, and generate block features and block probability maps; Step 4: Establish a seamless stitching memory based on the overlap region index, perform alignment and fusion on the block features and block probability maps to generate full-frame features and full-frame probability maps; Step 5: Perform a differentiable projection update on the full probability map at the rule consistency projection layer, investigate the set of rules to limit the category assignment variables and area constraints, and generate the projected probability map; Step 6: Perform node prediction and edge prediction on the projected probability map and full-frame features in the planar graph decoder to generate node set and edge set, and construct planar graph; Step 7: Based on the planar graph, apply closure constraints, no self-intersection constraints, and shared edge consistency constraints to generate map patch vectors and complete the automatic extraction of natural resource elements.

2. The method for automatic extraction based on comprehensive natural resource survey data according to claim 1, characterized in that, Step one specifically involves: Coordinate and resolution unification is performed on remote sensing images and comprehensive survey data to generate raster row and column indexes. The comprehensive survey data is then mapped and written into raster attribute fields according to the raster row and column indexes. The raster attribute fields include vector boundary identifiers, attribute ledger field values, field sampling point identifiers, and topographic factor field values. The vector boundary identifier is the association identifier of the vector boundary record on the raster row and column index. The attribute ledger field value consists of a field identifier and a field value. The field identifier is the attribute ledger field name code, and the field value is the value corresponding to the attribute ledger field name. The field sampling point identifier is the association identifier of the field sampling point record on the raster row and column index. The topographic factor field value is the numerical value corresponding to the topographic factor field name. Time stamp alignment is performed between the time stamps of remote sensing images and the time stamps of comprehensive survey data. A baseline time axis is established and a time index is generated. The baseline time axis is determined by the baseline time start point and the time step, and the time index is determined by the position of the time stamp on the baseline time axis. Geolocation codes are generated based on raster row and column indexes, and block indexes and overlap indexes are established based on geolocation codes. Block indexes are used to identify the blocks corresponding to geolocation codes, and overlap indexes are used to identify the set of geolocation codes where blocks overlap.

3. The method for automatic extraction based on comprehensive natural resource survey data according to claim 1, characterized in that, Step two specifically involves: Based on the block index, determine the set of raster row and column indexes corresponding to the block coverage area, merge the raster attribute fields according to the time index, and collect the vector boundary identifier, attribute ledger field value, field sampling point identifier, and terrain factor field value in the raster attribute fields to form a block prior record set, with the block prior record set using the block index as the record key; Prior coding is performed on the block prior record set, mapping the vector boundary identifier to the boundary embedding vector, the attribute ledger field value to the ledger embedding vector, the field sampling point identifier to the sampling embedding vector, and the terrain factor field value to the terrain embedding vector. The prior token sequence is generated by splicing and mapping the data in a preset order. A prior confidence sequence is generated based on the block prior record set. The prior confidence sequence is calculated and normalized based on the data source identifier, time index consistency identifier, and spatial coverage identifier. A mapping relationship is established between the prior token sequence and the block index. The mapping relationship associates the block index, the prior token sequence, and the prior confidence sequence.

4. The method for automatic extraction based on comprehensive natural resource survey data according to claim 1, characterized in that, Step three specifically involves: Based on the block index, remote sensing image blocks are extracted from the remote sensing image, and the prior token sequence and prior confidence sequence are read based on the block index. The improved SegFormer performs encoding on remote sensing image blocks to obtain encoded features, and performs decoding based on the encoded features to obtain an initial block probability map. The initial block probability map contains probability values ​​arranged by category at pixel positions. Prior category scores are calculated based on prior token sequences, and the prior category scores are mapped to the remote sensing image block space according to the block index to generate a prior guided probability map. The prior guided probability map contains probability values ​​arranged by category at the pixel position. The evidence residual map is calculated based on the initial block probability map and the prior guided probability map. The evidence residual value at the pixel position is obtained by summing the absolute values ​​of the probability differences in the category dimension. An evidence routing mask is generated based on the evidence residual map and the residual threshold. The evidence routing mask is marked as valid at positions where the evidence residual value is greater than the residual threshold, and as invalid at positions where the evidence residual value is less than or equal to the residual threshold. Evidence routing attention is performed at the valid location marked by the evidence routing mask. Evidence routing attention uses the encoded features as the query vector, the prior token sequence as the key vector and value vector to generate attention output features, and the attention output features are weighted by the confidence weights obtained by mapping the prior confidence sequence. The encoded features are fused with the weighted attention output features to generate block features, and the initial block probability map is updated based on the block features to generate a new block probability map.

5. The method for automatic extraction based on comprehensive natural resource survey data according to claim 1, characterized in that, Step four specifically involves: The write position of the block feature and block probability map in the full feature and full probability map is determined based on the block index. The write position is determined by the raster row and column index obtained by the geolocation coding mapping associated with the block index. A seamless splicing memory is built based on the overlapping area index. Memory entries are built according to the overlapping area index. The memory entries contain the set of geographic location codes associated with the overlapping area index, the overlapping area feature tensor, and the overlapping area probability tensor. The block features and block probability maps are written to the full features and full probability maps according to the writing position. The non-overlapping area index coverage range within the block index coverage range is covered by the overwrite write. Within the coverage area of ​​the overlapping area index, the block features and block probability maps are cropped based on the geographic location encoding set associated with the overlapping area index to obtain aligned block features and aligned block probability maps. The overlapping area feature tensor and overlapping area probability tensor are read from the seamless splicing memory based on the overlapping area index. The alignment block feature and the overlapping area feature tensor are weighted and summed position by position according to the fusion weight to generate the fused overlapping area feature. The alignment block probability map and the overlapping area probability tensor are weighted and summed position by position according to the fusion weight to generate the fused overlapping area probability. The fusion weight is determined and normalized by the relative position from the raster row and column index to the block index boundary within the overlapping area index coverage area. The fused overlapping area features and fused overlapping area probabilities are written into the overlapping area index coverage of the full-scale feature and full-scale probability maps, and the fused overlapping area features and fused overlapping area probabilities are written into the corresponding memory entries of the seamless splicing memory.

6. The method for automatic extraction based on comprehensive natural resource survey data according to claim 1, characterized in that, Step five specifically involves: The full probability map is variableized at the rule-consistent projection layer to generate category-assigned variables. The category-assigned variables contain probability components arranged by category at the raster row and column index positions. The rule coverage of the survey rule set is determined based on geolocation coding. Within the rule coverage, category constraints and area constraints are applied to the category allocation variables. Category constraints limit the set of category values ​​for the category allocation variables, and area constraints limit the category area statistics of the category allocation variables within the rule coverage. Constraint violation is calculated based on category constraints and area constraints. The constraint violation includes category constraint violation and area constraint violation. Category constraint violation is obtained by summing the probability components of the set of violation category values ​​within the rule coverage area. Area constraint violation is obtained by summing the statistical values ​​of category area and the difference between area constraints within the rule coverage area. Differentiable projection updates are performed on the category assignment variables based on the constraint violation amount. The differentiable projection update includes truncating the probability components of the set of violation category values ​​and normalizing the category dimension, and scaling the category probability components of the violation area constraint and normalizing the category dimension. The update terminates when the constraint violation amount is less than a threshold. The class assignment variables updated by differentiable projection are reconstructed to generate a projective probability map.

7. The method for automatic extraction based on comprehensive natural resource survey data according to claim 1, characterized in that, Step six specifically involves: The projection probability map and the full-frame feature are input into the planar graph decoder to generate a node score map on the full-frame feature. Peak extraction is performed on the node score map to obtain node candidate coordinates. The nodes are sorted based on their scores and filtered based on the minimum spacing constraint to obtain a node set. The node set contains node identifiers and node coordinates. In the planar graph decoder, an edge score tensor is generated based on the full-frame features. The edge score tensor uses the node identifier pairs of the node set as indices, and the edge score tensor takes the edge score of the node identifier pair. An edge set is generated by performing threshold filtering based on the edge score tensor and node identifier. The edge set includes edge identifier, start node identifier, and end node identifier. Adjacency records are generated based on the set of nodes and the set of edges. The adjacency records are associated with edge identifiers using node identifiers to construct a planar graph.

8. The method for automatic extraction based on comprehensive natural resource survey data according to claim 1, characterized in that, Step seven specifically involves: Based on the planar graph, a closed boundary chain is generated by tracking and recording the adjacency relationship in the edge set. The closed boundary chain satisfies that the starting node identifier and the ending node identifier are consistent. The closed boundary chain is written into the closed loop set. Perform a no-self-intersection constraint check on the set of closed loops. The no-self-intersection constraint check performs intersection detection on the edge segments contained in the closed boundary chain. If an intersection is detected, delete the closed boundary chain where the intersection is detected. Perform shared edge consistency constraint verification on the closed loop set. The shared edge consistency constraint verification extracts the shared edges based on the edge identifier and determines the consistency between the start node identifier and the end node identifier of the shared edge. If the consistency determination is not satisfied, delete the closed boundary chain that does not satisfy the consistency determination. Based on the closed loop set that passes the self-intersection constraint verification and the shared edge consistency constraint verification, a patch vector is generated. The patch vector contains patch identifiers and vertex sequences. The vertex sequences are formed by arranging the coordinates of the corresponding nodes of the closed boundary chain in the order of the closed boundary chain, thus completing the automatic extraction of natural resource elements.

9. A system for automatically extracting data based on comprehensive natural resource survey data, comprising executing the method for automatically extracting data based on comprehensive natural resource survey data as described in any one of claims 1 to 8, characterized in that, include: The data alignment index module is used to perform coordinate unification, resolution unification, and time stamp alignment on remote sensing images and comprehensive survey data, generate geolocation codes, establish block indexes and overlapping area indexes, and generate raster row and column indexes, baseline time axis, time index, and raster attribute fields. The prior coding module is used to perform prior coding on the comprehensive survey data, generate a prior token sequence and a prior confidence sequence, and establish a mapping relationship between the prior token sequence and the block index. The block interpretation module is used to perform improved SegFormer encoding inference on remote sensing image blocks, calculate evidence residual maps, trigger evidence routing attention to read prior token sequences according to the evidence residual maps, and generate block features and block probability maps. The seamless stitching module is used to build a seamless stitching memory based on the overlap area index, perform alignment and fusion on the block features and block probability maps, and generate full-frame features and full-frame probability maps. The rule projection module is used to perform differentiable projection updates on the full-scale probability map at the rule consistency projection layer, investigate the rule set to limit category assignment variables and area constraints, and generate the projected probability map; The planar graph decoding module is used to perform node and edge prediction on the projected probability map and full-frame features in the planar graph decoder, generate node and edge sets, and construct a planar graph. The feature generation module is used to generate feature vectors based on planar maps by applying closure constraints, no self-intersection constraints, and shared edge consistency constraints, thereby completing the automatic extraction of natural resource elements.

Images