A dynamic recognition and hierarchical processing method for addition and subtraction hook problems based on image semantics

Abstract

This invention provides a dynamic identification and hierarchical processing method for land acquisition and depletion linkage issues based on image semantics. The method includes: unifying coordinates, aligning temporal relationships, and gridding encoding multi-source data of land acquisition and depletion linkage project areas; segmenting remote sensing images according to plot boundaries; extracting internal, boundary neighborhood semantic features and shape features; calculating boundary reliability and constructing plot semantic representations; extracting directional texture and hierarchical distribution features through a dual-path model; weighting and fusing them into structural semantic features according to boundary reliability; extracting multi-temporal temporal variation features by combining topographic elevation and historical data; generating spatiotemporal semantic vectors; identifying problems such as construction occupation, insufficient reclamation, boundary mismatch, and use deviation; updating boundary reliability; determining the plot problem level based on the identification results; analyzing adjacent similar plots as a group; determining boundary adjustment schemes; and outputting processing results including problem level and boundary optimization results.

Description

Technical Field

[0001] This application belongs to the field of recognition, and in particular relates to a dynamic recognition and hierarchical processing method for addition and subtraction hook problems based on image semantics. Background Technology

[0002] With the widespread implementation of land use quota trading projects, the need to monitor issues such as illegal construction and occupation, inadequate land reclamation, and deviation from land use within project areas is becoming increasingly urgent. Existing technologies typically combine multi-source remote sensing imagery with deep learning models. By unifying and segmenting coordinates of multi-temporal remote sensing data, neural networks are used to extract texture and hierarchical image features of land parcels, and historical data is combined with time-series analysis to monitor changes in land use status.

[0003] However, existing methods mostly employ a uniform receptive field when extracting image features, neglecting the differences between the semantics within a plot and the semantics of its boundary neighborhood. They fail to scientifically represent the reliability of plot boundaries by combining boundary continuity and shape regularity, leading to feature confusion when dealing with complex boundaries. Existing models often rely on fixed weights when fusing directional texture and hierarchical distribution features, failing to adjust based on the plot's own boundary reliability, resulting in inaccurate generated structural semantic features. Most existing monitoring methods only output single change detection results, failing to deeply integrate topographic elevation constraints to construct spatiotemporal semantic vectors. They cannot accurately decompose and output multi-dimensional discrimination results such as construction occupation, insufficient reclamation, boundary mismatch, and usage deviation, and lack a closed-loop feedback mechanism to update boundary reliability using boundary mismatch discrimination results. When dealing with large-scale land linkage areas, existing technologies often analyze individual plots in isolation, lacking the ability to analyze adjacent plots with consistent trends as a group. They cannot adjust the group boundaries based on differences in characteristics within the group, nor can they scientifically determine the problem level of plots and output graded processing results, which seriously restricts the level of intelligence and precision in land linkage monitoring. Summary of the Invention

[0004] To address the issues that existing technologies cannot adaptively extract and fuse multi-source features based on the reliability of land parcel boundaries, and lack multi-dimensional discrimination, closed-loop feedback, and cluster analysis that incorporate terrain constraints.

[0005] In the first aspect, the present invention provides the following technical solution: a method for dynamic identification and hierarchical processing of addition and subtraction hook problems based on image semantics, comprising:

[0006] Multi-source data of the land increase / decrease linkage area were acquired and subjected to coordinate unification, temporal alignment and gridding encoding. Remote sensing images were segmented according to the land parcel boundaries, and semantic features of the land parcel interior, boundary neighborhood semantic features and shape features were extracted. Boundary reliability was calculated based on the difference between the semantic features of the land parcel interior and the semantic features of the boundary neighborhood, the boundary continuity of the land parcel and the shape regularity. The semantic representation of the land parcel was constructed by combining the shape features. The semantic representation of the land parcel was input into a dual-path model to extract directional texture features and hierarchical distribution features. The weights of directional texture features and hierarchical distribution features were adjusted according to the boundary reliability and fused into structural semantic features.

[0007] Combining topographic elevation constraints and historical data, temporal change features are extracted from multi-temporal structural semantic features to generate spatiotemporal semantic vectors. Based on the spatiotemporal semantic vectors, the results of construction occupation discrimination, insufficient reclamation discrimination, boundary mismatch discrimination, and use deviation discrimination are output. The boundary reliability is updated according to the boundary mismatch discrimination results to obtain the updated boundary reliability.

[0008] Based on the updated boundary reliability, construction occupation judgment results, insufficient reclamation judgment results, boundary mismatch judgment results, and use deviation judgment results, the problem level of the land parcel is determined. Adjacent target land parcels with consistent change trends are grouped for analysis. The group boundary adjustment results are determined based on the differences in characteristics within the group. The output includes the land parcel problem level and the group boundary adjustment results.

[0009] Optionally, the step of acquiring multi-source data of the addition / reduction hook area and performing coordinate unification, time-relative alignment, and grid encoding includes:

[0010] Extract spatial reference information from multi-source data, and uniformly convert remote sensing images, land parcel boundaries, and terrain elevations from different coordinate systems to the National Geodetic Coordinate System for spatial projection alignment;

[0011] Establish a time series reference axis, set a time matching window, remove incomplete temporal images with cloud and snow coverage exceeding a preset threshold, and interpolate and align the retained multi-temporal remote sensing images and terrain elevations according to the acquisition timestamps.

[0012] A two-dimensional grid matrix is ​​constructed with a preset fixed spatial resolution. The alignment data covering the increase / decrease hook area is divided into independent grid cells of the same size, and a unique and non-repeating spatial index string is assigned to each grid cell to generate a gridded data packet.

[0013] Optionally, the calculation of boundary reliability based on the difference between the semantic features within the land parcel and the semantic features of the boundary neighborhood, the boundary continuity of the land parcel, and the shape regularity includes:

[0014] Extract the feature vector center of the pixel set inside the plot as the semantic feature inside the plot, extract the feature vector center of the pixel set within a preset buffer distance extending outward from the plot boundary as the semantic feature of the boundary neighborhood, calculate the cosine distance between the semantic feature inside the plot and the semantic feature of the boundary neighborhood, and take this distance as the difference degree.

[0015] An edge detection algorithm is used to identify the actual number of connected pixels at the boundary of a land parcel, and the ratio of the actual number of connected pixels to the total number of pixels in the theoretical boundary perimeter is used as the boundary continuity of the land parcel.

[0016] The ratio of the actual area of ​​the plot to the area of ​​its convex hull is used as the shape regularity.

[0017] Preset weights are assigned to the degree of difference, the boundary continuity of the plot, and the shape regularity. The three indicators of each plot are summed by weight to obtain a weighted composite index. The weighted composite index is then mapped to the (0,1) interval by the Sigmoid function to obtain the boundary reliability of the plot.

[0018] Optionally, the step of adjusting the weights of directional texture features and hierarchical distribution features according to boundary reliability and fusing them into structural semantic features includes:

[0019] By using boundary reliability as the attention weight for directional texture features, element-wise multiplication is performed to obtain weighted directional texture features.

[0020] The complementarity coefficient is obtained by subtracting the boundary reliability from the constant 1. The complementarity coefficient is used as the attention weight of the hierarchical distribution feature, and element-wise multiplication is performed to obtain the weighted hierarchical distribution feature.

[0021] Weighted directional texture features and weighted hierarchical distribution features are concatenated along the channel dimension, and then dimensionality reduction and feature mapping are performed through a fully connected layer to generate structural semantic features.

[0022] Optionally, the step of combining terrain elevation constraints and historical data to extract temporal change features from multi-temporal structural semantic features and generate a spatiotemporal semantic vector includes:

[0023] The feature sequence is obtained by arranging the semantic features of the multi-temporal structure in chronological order.

[0024] Calculate the topographic elevation difference of each time phase relative to the initial time phase, use the elevation difference as a spatial constraint term, and concatenate it with the remediation type features extracted from historical data into the corresponding feature sequence node;

[0025] The concatenated feature sequence is input into a long short-term memory network or a gated recurrent unit network. The historical time series state and the current time phase features are filtered and updated through the gating mechanism, and the hidden layer state vector is extracted as a spatiotemporal semantic vector.

[0026] Optionally, the step of outputting construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and use deviation discrimination results based on spatiotemporal semantic vectors, and updating the boundary reliability according to the boundary mismatch discrimination results to obtain the updated boundary reliability, includes:

[0027] The spatiotemporal semantic vector is input into a multi-branch fully connected classifier. The probability distribution vectors of the current plot in the dimensions of construction occupation, insufficient reclamation, boundary mismatch, and use deviation are calculated by the normalized exponential function. The category corresponding to the maximum probability is taken as the construction occupation discrimination result, insufficient reclamation discrimination result, boundary mismatch discrimination result, and use deviation discrimination result, respectively.

[0028] When the boundary mismatch judgment result is that a mismatch exists, calculate the probability offset of the boundary mismatch dimension, multiply the boundary reliability by the penalty factor calculated based on the probability offset, and perform weighted update to obtain the updated boundary reliability.

[0029] When there is no mismatch, the boundary reliability remains unchanged and is used as the updated boundary reliability.

[0030] Optionally, determining the land parcel problem level based on the updated boundary reliability, construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and land use deviation discrimination results includes:

[0031] Construct a level rule matrix containing multiple preset severity levels, and pre-assign severity weight coefficients to the construction occupation judgment results, insufficient reclamation judgment results, boundary mismatch judgment results, and use deviation judgment results;

[0032] The results of the statistical analysis of construction occupation, insufficient reclamation, boundary mismatch, and deviation of use are used to identify abnormal issues, and the severity weight coefficients of the corresponding issues are accumulated.

[0033] The cumulative severity weight coefficients are nonlinearly combined and mapped with the updated boundary reliability to calculate the comprehensive risk index of the land parcel.

[0034] The comprehensive risk index of the land parcel is compared with the threshold range in the level rule matrix to output the corresponding land parcel problem level.

[0035] Optionally, the step of performing cluster analysis on adjacent target plots with consistent changing trends, and determining the cluster boundary adjustment results based on the differences in characteristics within the clusters, includes:

[0036] The adjacency graph set of the target land parcel is retrieved using spatial topological relationships, and the cosine similarity between the target land parcel and each neighboring land parcel in the spatiotemporal semantic vector is calculated.

[0037] Select adjacent land parcels with a cosine similarity greater than a preset merging threshold and the same problem level, and use the connected component algorithm to merge the selected land parcels into a connected cluster.

[0038] Calculate the updated boundary reliability variance of each parcel within the connected cluster. If the variance is greater than a preset difference threshold, then combine the internal characteristics of the connected cluster to smooth the boundary of the parcel with the lowest updated boundary reliability. Based on the corrected parcel topology, regenerate the external contour vector of the connected cluster as the result of the cluster boundary adjustment.

[0039] In a second aspect, the present invention also provides the following technical solution: a dynamic recognition and hierarchical processing system for addition and subtraction hook problems based on image semantics, comprising:

[0040] The extraction module is used to acquire multi-source data of the land parcel increase / decrease linkage area and perform coordinate unification, temporal alignment and grid encoding. It segments the remote sensing image according to the land parcel boundary, extracts semantic features inside the land parcel, semantic features of the boundary neighborhood and shape features. It calculates the boundary reliability based on the difference between the semantic features inside the land parcel and the semantic features of the boundary neighborhood, the boundary continuity of the land parcel and the shape regularity. It constructs the semantic representation of the land parcel by combining the shape features. It inputs the semantic representation of the land parcel into the dual-path model, extracts directional texture features and hierarchical distribution features, and adjusts the weights of directional texture features and hierarchical distribution features according to the boundary reliability and merges them into structural semantic features.

[0041] The update module is used to combine terrain elevation constraints and historical data to extract temporal change features from multi-temporal structural semantic features, generate spatiotemporal semantic vectors, output construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results and use deviation discrimination results based on the spatiotemporal semantic vectors, and update the boundary reliability according to the boundary mismatch discrimination results to obtain the updated boundary reliability.

[0042] The output module is used to determine the problem level of a land parcel based on the updated boundary reliability, construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and use deviation discrimination results. It performs cluster analysis on adjacent target land parcels with consistent change trends, determines the cluster boundary adjustment results based on the differences in characteristics within the cluster, and outputs the processing results containing the land parcel problem level and the cluster boundary adjustment results.

[0043] Preferably, the step of acquiring multi-source data of the addition / reduction hook area and performing coordinate unification, time-relative alignment, and grid encoding includes:

[0044] Extract spatial reference information from multi-source data, and uniformly convert remote sensing images, land parcel boundaries, and terrain elevations from different coordinate systems to the National Geodetic Coordinate System for spatial projection alignment;

[0045] Establish a time series reference axis, set a time matching window, remove incomplete temporal images with cloud and snow coverage exceeding a preset threshold, and interpolate and align the retained multi-temporal remote sensing images and terrain elevations according to the acquisition timestamps.

[0046] A two-dimensional grid matrix is ​​constructed with a preset fixed spatial resolution. The alignment data covering the increase / decrease hook area is divided into independent grid cells of the same size, and a unique and non-repeating spatial index string is assigned to each grid cell to generate a gridded data packet.

[0047] Preferably, the calculation of boundary reliability based on the difference between the semantic features within the land parcel and the semantic features of the boundary neighborhood, the boundary continuity of the land parcel, and the regularity of its shape includes:

[0048] Extract the feature vector center of the pixel set inside the plot as the semantic feature inside the plot, extract the feature vector center of the pixel set within a preset buffer distance extending outward from the plot boundary as the semantic feature of the boundary neighborhood, calculate the cosine distance between the semantic feature inside the plot and the semantic feature of the boundary neighborhood, and take this distance as the difference degree.

[0049] An edge detection algorithm is used to identify the actual number of connected pixels at the boundary of a land parcel, and the ratio of the actual number of connected pixels to the total number of pixels in the theoretical boundary perimeter is used as the boundary continuity of the land parcel.

[0050] The ratio of the actual area of ​​the plot to the area of ​​its convex hull is used as the shape regularity.

[0051] Preset weights are assigned to the degree of difference, the boundary continuity of the plot, and the shape regularity. The three indicators of each plot are summed by weight to obtain a weighted composite index. The weighted composite index is then mapped to the (0,1) interval by the Sigmoid function to obtain the boundary reliability of the plot.

[0052] Preferably, the step of adjusting the weights of directional texture features and hierarchical distribution features according to boundary reliability and fusing them into structural semantic features includes:

[0053] By using boundary reliability as the attention weight for directional texture features, element-wise multiplication is performed to obtain weighted directional texture features.

[0054] The complementarity coefficient is obtained by subtracting the boundary reliability from the constant 1. The complementarity coefficient is used as the attention weight of the hierarchical distribution feature, and element-wise multiplication is performed to obtain the weighted hierarchical distribution feature.

[0055] Weighted directional texture features and weighted hierarchical distribution features are concatenated along the channel dimension, and then dimensionality reduction and feature mapping are performed through a fully connected layer to generate structural semantic features.

[0056] Preferably, the step of combining terrain elevation constraints and historical data to extract temporal change features from multi-temporal structural semantic features and generate a spatiotemporal semantic vector includes:

[0057] The feature sequence is obtained by arranging the semantic features of the multi-temporal structure in chronological order.

[0058] Calculate the topographic elevation difference of each time phase relative to the initial time phase, use the elevation difference as a spatial constraint term, and concatenate it with the remediation type features extracted from historical data into the corresponding feature sequence node;

[0059] The concatenated feature sequence is input into a long short-term memory network or a gated recurrent unit network. The historical time series state and the current time phase features are filtered and updated through the gating mechanism, and the hidden layer state vector is extracted as a spatiotemporal semantic vector.

[0060] Preferably, the step of outputting construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and use deviation discrimination results based on spatiotemporal semantic vectors, and updating the boundary reliability according to the boundary mismatch discrimination results to obtain the updated boundary reliability, includes:

[0061] The spatiotemporal semantic vector is input into a multi-branch fully connected classifier. The probability distribution vectors of the current plot in the dimensions of construction occupation, insufficient reclamation, boundary mismatch, and use deviation are calculated by the normalized exponential function. The category corresponding to the maximum probability is taken as the construction occupation discrimination result, insufficient reclamation discrimination result, boundary mismatch discrimination result, and use deviation discrimination result, respectively.

[0062] When the boundary mismatch judgment result is that a mismatch exists, calculate the probability offset of the boundary mismatch dimension, multiply the boundary reliability by the penalty factor calculated based on the probability offset, and perform weighted update to obtain the updated boundary reliability.

[0063] When there is no mismatch, the boundary reliability remains unchanged and is used as the updated boundary reliability.

[0064] Preferably, determining the land parcel problem level based on the updated boundary reliability, construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and land use deviation discrimination results includes:

[0065] Construct a level rule matrix containing multiple preset severity levels, and pre-assign severity weight coefficients to the construction occupation judgment results, insufficient reclamation judgment results, boundary mismatch judgment results, and use deviation judgment results;

[0066] The results of the statistical analysis of construction occupation, insufficient reclamation, boundary mismatch, and deviation of use are used to identify abnormal issues, and the severity weight coefficients of the corresponding issues are accumulated.

[0067] The cumulative severity weight coefficients are nonlinearly combined and mapped with the updated boundary reliability to calculate the comprehensive risk index of the land parcel.

[0068] The comprehensive risk index of the land parcel is compared with the threshold range in the level rule matrix to output the corresponding land parcel problem level.

[0069] Preferably, the step of performing cluster analysis on adjacent target plots with consistent changing trends, and determining the cluster boundary adjustment results based on the differences in characteristics within the clusters, includes:

[0070] The adjacency graph set of the target land parcel is retrieved using spatial topological relationships, and the cosine similarity between the target land parcel and each neighboring land parcel in the spatiotemporal semantic vector is calculated.

[0071] Select adjacent land parcels with a cosine similarity greater than a preset merging threshold and the same problem level, and use the connected component algorithm to merge the selected land parcels into a connected cluster.

[0072] Calculate the updated boundary reliability variance of each parcel within the connected cluster. If the variance is greater than a preset difference threshold, then combine the internal characteristics of the connected cluster to smooth the boundary of the parcel with the lowest updated boundary reliability. Based on the corrected parcel topology, regenerate the external contour vector of the connected cluster as the result of the cluster boundary adjustment.

[0073] This invention achieves a comprehensive and detailed semantic representation of land parcels by accurately aligning and gridding multi-source data from land consolidation and allocation zones, deeply exploring the semantics within parcels, the semantics of their boundary neighborhoods, and their shape features. Boundary reliability is calculated using boundary continuity and shape regularity, thereby adjusting the fusion weights of directional textures and hierarchical distribution features extracted by the dual-path model, improving the accuracy of extracting complex surface structure features. By combining topographic elevation constraints with historical data to extract multi-temporal variation features, accurate assessment of issues such as construction occupation, insufficient reclamation, boundary mismatch, and use deviation is achieved. Boundary reliability is updated using boundary mismatch judgment results, and cluster analysis and boundary optimization are performed on adjacent parcels, reducing identification errors caused by spatial fragmentation and improving the scientific rigor and overall accuracy of parcel-related problem classification. Attached Figure Description

[0074] Figure 1 This is a flowchart of a dynamic identification and hierarchical processing method for addition and subtraction hook problems based on image semantics;

[0075] Figure 2 A schematic diagram of cloud cover filtering in multi-temporal remote sensing images;

[0076] Figure 3 A schematic diagram showing the classification accuracy and boundary intersection-union ratio of different models. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0078] Firstly, this invention proposes a dynamic identification and hierarchical processing method for addition and subtraction hook problems based on image semantics, such as... Figure 1 As shown, it includes:

[0079] S1: Acquire multi-source data of the land parcel linkage area and perform coordinate unification, temporal alignment and gridding encoding, and divide the remote sensing image according to the land parcel boundary.

[0080] The projection transformation module in the GDAL library was used to convert the UAV orthophotos and vector plot boundaries into the CGCS2000 national geodetic coordinate system. Spline interpolation was used to smooth and align the pixel values ​​of the multi-temporal remote sensing images over time. The GeoHash algorithm was employed to perform multi-level gridded spatial encoding of the study area. Using the mask extraction function in the Rasterio library, the aligned multi-temporal remote sensing images were cropped based on the geometric coordinates of the vector plot boundaries. Individual plot image data were then obtained.

[0081] In one possible embodiment, the step of acquiring multi-source data of the addition / reduction hook area and performing coordinate unification, time-relative alignment, and grid encoding includes:

[0082] Extract spatial reference information from multi-source data, and uniformly convert remote sensing images, land parcel boundaries, and terrain elevations from different coordinate systems to the National Geodetic Coordinate System for spatial projection alignment;

[0083] Establish a time series reference axis, set a time matching window, remove incomplete temporal images with cloud and snow coverage exceeding a preset threshold, and interpolate and align the retained multi-temporal remote sensing images and terrain elevations according to the acquisition timestamps.

[0084] A two-dimensional grid matrix is ​​constructed with a preset fixed spatial resolution. The alignment data covering the increase / decrease hook area is divided into independent grid cells of the same size, and a unique and non-repeating spatial index string is assigned to each grid cell to generate a gridded data packet.

[0085] The multi-source data includes high-resolution optical remote sensing imagery such as 0.8m resolution vector block boundaries in Shapefile format and 30m resolution ASTERGDEM topographic elevation data. EPSG codes from each data source were extracted using a spatial data processing library. For raster data such as remote sensing images and topographic elevations, bilinear interpolation was used for resampling to achieve coordinate transformation. For vector block boundaries, projection transformation was performed directly using coordinate transformation parameters, unifying all spatial data to the National Geodetic Coordinate System (CGCS2000). The natural year was set as the time series reference axis, and the time matching window was set to ±15 days. The FMASK algorithm was used to calculate the cloud and snow cover of each image, and incomplete time-phase images with a cover exceeding a preferred threshold of 15% were removed. For missing time-phase images, a cubic spline interpolation algorithm was used to interpolate and align the time series based on the spectral values ​​and elevation constraints of pixels from preceding and following time phases. Figure 2 This is a time-series diagram illustrating the cloud cover screening process for multi-temporal remote sensing images according to the present invention. The horizontal axis represents the time series in days; the vertical axis represents the cloud cover percentage; solid dots indicate retained images whose cloud cover meets the screening criteria and can be used for subsequent analysis; hollow squares indicate invalid images with excessive cloud cover that are removed; and dashed lines represent the cloud cover screening threshold. Based on a preferred fixed spatial resolution of 2m × 2m, a two-dimensional grid matrix is ​​constructed using a quadtree algorithm with the upper left corner of the study area as the origin. The aligned region is divided into N × M independent grid units, and a unique and non-repeating spatial index string is assigned to each grid according to the rule of hierarchy to row number to column number, such as L3 to R105 to C204. The pixel values ​​and elevation values ​​of each grid are encapsulated and exported as a gridded data package in HDF5 format.

[0086] S2 extracts semantic features within the land parcel, semantic features of the boundary neighborhood, and shape features.

[0087] A ResNet50 convolutional neural network was built based on the PyTorch deep learning framework. The segmented land parcel images were input into this network, and global semantic features within the parcels were extracted using an average pooling layer. Using the buffer analysis function from the Shapely library, a ring-shaped polygon with a width of k pixels (ideally 10) was extended outward from the vector parcel boundary. This ring-shaped polygon was then used to crop the original image, obtaining the boundary neighborhood image, which was then input into the ResNet50 network again to extract boundary neighborhood semantic features. Simultaneously, the contour calculation function from the OpenCV library was used to extract the area, perimeter, and aspect ratio of the minimum bounding rectangle of the land parcel vector polygon. A compactness index was calculated as a shape feature.

[0088] The ResNet50, used as the backbone network for feature extraction, comprises 50 convolutional layers. Its core is the residual block, each consisting of 1×1, 3×3, and 1×1 bottleneck layers and skip connections to address the vanishing gradient problem. The network input is either a land parcel image or a boundary neighborhood image. After convolution, pooling, and four sets of residual modules, a one-dimensional feature vector is output via global average pooling. During training, pre-trained weights on ImageNet are typically used as initialization, followed by fine-tuning based on a land parcel image dataset of the added / removed areas. This involves using a cross-entropy loss function, an optimizer such as AdamW, and a learning rate decay and early stopping strategy. Backpropagation is used to fine-tune only high-level convolutional or fully connected layers, efficiently extracting deep semantic features within and around the land parcels without requiring retraining from scratch.

[0089] S3, calculate the boundary reliability based on the difference between the semantic features inside the plot and the semantic features of the boundary neighborhood, the boundary continuity of the plot, and the shape regularity.

[0090] The spatial distance calculation module in the SciPy library is used to calculate the cosine distance between the semantic features inside the land parcel and the semantic features of its boundary neighborhood, which is used as the dissimilarity index. The Canny edge detection algorithm is used to calculate the pixel gradient of the land parcel boundary region and to statistically analyze the proportion of edge breakpoints, serving as an indicator of the land parcel boundary continuity. The convex hull calculation function in the Shapely library is used to obtain the convex hull area of ​​the land parcel polygon, and the ratio of the actual land parcel area to the convex hull area is calculated as the shape regularity index. The dissimilarity index, boundary continuity index, and shape regularity index are then non-linearly weighted and summed through a fully connected layer. Finally, a sigmoid activation function is applied to output a boundary reliability value ranging from 0 to 1.

[0091] In one possible embodiment, calculating boundary reliability based on the difference between the semantic features within the land parcel and the semantic features of the boundary neighborhood, the boundary continuity of the land parcel, and the shape regularity includes:

[0092] Extract the feature vector center of the pixel set inside the plot as the semantic feature inside the plot, extract the feature vector center of the pixel set within a preset buffer distance extending outward from the plot boundary as the semantic feature of the boundary neighborhood, calculate the cosine distance between the semantic feature inside the plot and the semantic feature of the boundary neighborhood, and take this distance as the difference degree.

[0093] An edge detection algorithm is used to identify the actual number of connected pixels at the boundary of a land parcel, and the ratio of the actual number of connected pixels to the total number of pixels in the theoretical boundary perimeter is used as the boundary continuity of the land parcel.

[0094] The ratio of the actual area of ​​the plot to the area of ​​its convex hull is used as the shape regularity.

[0095] Preset weights are assigned to the degree of difference, the boundary continuity of the plot, and the shape regularity. The three indicators of each plot are summed by weight to obtain a weighted composite index. The weighted composite index is then mapped to the (0,1) interval by the Sigmoid function to obtain the boundary reliability of the plot.

[0096] Global average pooling is used to extract the feature vectors of the pixel set within the plot and the mean value is calculated as the semantic feature of the plot. In the spatial computing environment, a buffer tool is used to extend the buffer outward along the plot vector boundary by a preset buffer distance, preferably within the range of 2 to 5 meters (3 meters in this example). The mean value of the pixel set within this annular buffer zone is extracted as the semantic feature of the boundary neighborhood. The cosine distance between the two feature vectors is calculated and taken as the difference. The larger the cosine distance between the semantic feature inside the plot and the semantic feature of the boundary neighborhood, the more significant the difference between inside and outside the boundary and the clearer the boundary. Therefore, it is directly used as the difference value to participate in the weighted summation, and the boundary obtained by Sigmoid mapping has a higher reliability. Using the Canny edge detection operator with high and low thresholds set to 200 and 100 respectively, the land parcel mask was processed to identify the actual number of connected pixels at the parcel boundary, e.g., 450. The theoretical total number of boundary pixels, e.g., 480, was calculated using the parcel perimeter formula. The ratio of these two values, 0.9375, was taken as the boundary continuity of the parcel. The convex hull of the parcel was calculated using the Shapely library, and its area was obtained. The ratio of the actual parcel area, e.g., 1200 square meters, to the convex hull area, e.g., 1300 square meters, approximately 0.923, was taken as the shape regularity. Based on feature importance, weight parameters were pre-set using the analytic hierarchy process: 0.5 for dissimilarity, 0.3 for boundary continuity, and 0.2 for shape regularity. These three indices were multiplied by their respective weights and summed to obtain a weighted composite index. This weighted composite index was input into the Sigmoid function, mapped to the (0,1) interval, and the boundary reliability of the parcel was output, e.g., 0.815.

[0097] S4, construct a semantic representation of the land parcel by combining shape features.

[0098] The array concatenation function in the NumPy library is called to concatenate the one-dimensional semantic feature vector inside the land parcel with the numerical shape features containing compactness and regularity according to the feature dimension. The numerical differences caused by different physical dimensions are eliminated by the layer normalization algorithm, and the initial land parcel semantic representation tensor that comprehensively represents the visual and spatial geometric information of the land parcel is constructed.

[0099] S5. Input the semantic representation of the land parcel into the dual-path model, extract directional texture features and hierarchical distribution features, adjust the weights of directional texture features and hierarchical distribution features according to the boundary reliability, and fuse them into structural semantic features.

[0100] The initial one-dimensional parcel semantic representation tensor is reconstructed by using a fully connected layer to map the one-dimensional vector into a high-dimensional vector. Then, it is reshaped into a two-dimensional feature map with spatial topology, resulting in a spatialized parcel semantic representation.

[0101] A dual-path deep learning model is constructed, comprising a directional texture extraction branch and a multi-scale feature pyramid network branch. In the first path, a multi-directional and multi-scale Gabor filter bank is applied to convolutionally filter the spatialized parcel semantic representation, generating a directional texture feature matrix. In the second path, a feature pyramid network algorithm is used to perform top-down upsampling and lateral connections on the spatialized parcel semantic representation, generating a tensor representing the spatial hierarchical distribution features. The boundary reliability value is used as the attention allocation coefficient, and the boundary reliability is directly multiplied by the directional texture feature matrix. The value of 1 minus the boundary reliability is then multiplied by the hierarchical distribution feature tensor. The two types of features, after weight adjustment, are superimposed and fused to generate a structural semantic feature vector.

[0102] In one possible embodiment, adjusting the weights of the directional texture features and hierarchical distribution features according to boundary reliability and fusing them into structural semantic features includes:

[0103] By using boundary reliability as the attention weight for directional texture features, element-wise multiplication is performed to obtain weighted directional texture features.

[0104] The complementarity coefficient is obtained by subtracting the boundary reliability from the constant 1. The complementarity coefficient is used as the attention weight of the hierarchical distribution feature, and element-wise multiplication is performed to obtain the weighted hierarchical distribution feature.

[0105] Weighted directional texture features and weighted hierarchical distribution features are concatenated along the channel dimension, and then dimensionality reduction and feature mapping are performed through a fully connected layer to generate structural semantic features.

[0106] Assuming that both the extracted directional texture features and hierarchical distribution features have tensor dimensions of 256 × 1 × 1 times the batch size, the calculated boundary reliability scalar value is 0.75. Using 0.75 as the spatial dimension attention hard weight, an element-wise Hadamard product operation is performed with the directional texture features using a tensor broadcasting mechanism to generate a weighted directional texture feature that retains high-frequency information. A complementarity coefficient of 0.25 is obtained by calculating the difference between the constant 1 and the boundary reliability 0.75. This complementarity coefficient is then used as a soft weight, and an element-wise multiplication operation is performed with the hierarchical distribution features based on the broadcasting mechanism to obtain a weighted hierarchical distribution feature. This automatically adjusts the proportion of macroscopic hierarchical information when the boundary confidence is low. The weighted directional texture features and weighted hierarchical distribution features are concatenated in the channel dimension (axis=1) using the Concat function to obtain a fused feature map with 512 channels. The concatenated features are input into a fully connected network with a Dropout layer configured with ReLU activation and a dropout rate of 0.3. Channel dimensionality reduction and nonlinear mapping are performed, outputting a 128-dimensional structural semantic feature vector. Here, the fully connected network serves as a basic neural network model, taking a 512-dimensional concatenated feature as input. Its structure includes dropout layers, linear transformation layers, and nonlinear activation layers, outputting a 128-dimensional structural semantic feature vector. The computation process of the fully connected network is Y = ReLU(Dropout(X)W + b), where X represents the input feature, W represents the weight matrix, b represents the bias vector, Y represents the output feature, Dropout represents the dropout operation, and ReLU represents the activation function.

[0107] S6 combines terrain elevation constraints with historical data to extract temporal change features from multi-temporal structural semantic features and generate spatiotemporal semantic vectors.

[0108] The Rasterio library is used to read digital elevation model data, and the average elevation and slope of each plot are calculated as topographic elevation constraint variables. Numerical historical record features are extracted by combining an attribute database containing historical land use codes and verification statuses. The structural semantic feature tensors, sorted by timestamps at different time points, are input into a long short-term memory neural network algorithm. The variation patterns of the feature space along the time axis are detected, and a latent state vector of temporal variation features is output. A multilayer perceptron network is used to map the topographic elevation constraint variables, historical record features, and latent state vector of temporal variation features to the same high-dimensional semantic space. A cascaded function is then called to generate a high-dimensional spatiotemporal semantic vector.

[0109] In one possible embodiment, the step of combining terrain elevation constraints and historical data to extract temporal change features from multi-temporal structural semantic features and generate a spatiotemporal semantic vector includes:

[0110] The feature sequence is obtained by arranging the semantic features of the multi-temporal structure in chronological order.

[0111] Calculate the topographic elevation difference of each time phase relative to the initial time phase, use the elevation difference as a spatial constraint term, and concatenate it with the remediation type features extracted from historical data into the corresponding feature sequence node;

[0112] The concatenated feature sequence is input into a long short-term memory network or a gated recurrent unit network. The historical time series state and the current time phase features are filtered and updated through the gating mechanism, and the hidden layer state vector is extracted as a spatiotemporal semantic vector.

[0113] Assuming the multi-temporal data contains four quarterly time nodes, the 128-dimensional structural semantic features generated for each node are arranged according to... to The timestamps are arranged in order to construct an initial sequence matrix with dimensions 4 × 128 times the batch size. The average elevation values ​​for each time phase are extracted using the DEM, and the results are calculated. to Relative to the initial Elevation differences, such as -0.5m and -1.2m, at elevations like 45.2 meters, are used as constraints. These 1D elevation differences, along with 5D historical land reclamation type features generated using one-hot encoding (e.g., 00100 representing reclamation as arable land), are concatenated into the features of the corresponding time nodes along the channel dimension, expanding the features of each node to 134 dimensions, resulting in a multimodal enhanced feature sequence. This feature sequence is input into a Gated Recurrent Unit (GRU) network with a preferred number of hidden layer units of 256. In each time step, the reset and update gates within the GRU calculate Sigmoid activation values ​​ranging from 0 to 1, autonomously selecting the amount of historical information to be forgotten and the amount of current temporal features to be incorporated. After completing the forward propagation calculation for all time steps, the network extracts and outputs the 256-dimensional hidden layer state vector of the last time step, serving as a spatiotemporal semantic vector containing subsidence, change patterns, and global temporal characteristics. Here, the gated recurrent unit network is a neural network for processing sequential data. The input is a multimodal enhanced feature sequence. The structure includes an update gate, a reset gate, and a candidate hidden state calculation module. The output is the hidden state vector of the time step.

[0114] S7. Based on the spatiotemporal semantic vector, output the construction occupation discrimination result, reclamation insufficiency discrimination result, boundary mismatch discrimination result and use deviation discrimination result, and update the boundary reliability according to the boundary mismatch discrimination result to obtain the updated boundary reliability.

[0115] High-dimensional spatiotemporal semantic vectors are simultaneously input into four independent multi-class feedforward neural network heads. The Softmax activation function is used to normalize the probability distribution of the output features of each classifier head. The first classifier head outputs a probability vector indicating newly added construction land, serving as the construction occupation discrimination result. The second classifier head outputs a probability vector indicating vegetation cover is below a threshold, serving as the insufficient reclamation discrimination result. The third classifier head outputs a probability vector indicating inconsistency between the actual land feature outline and the geometric topology of the vector boundary coordinates, serving as the boundary mismatch discrimination result. The fourth classifier head outputs a probability vector indicating discrepancies between the actual land use type and the planned land use code, serving as the use deviation discrimination result. Scalar probability values ​​representing the existence of mismatch categories are extracted from the probability vectors of the boundary mismatch discrimination results. The initial boundary reliability calculated in the preceding steps is used as a priori estimated state through the gain update mechanism of the Kalman filter algorithm. The extracted mismatch probability values ​​are used as the current observed variables for iterative calculation of the state equation. The initial boundary reliability value is smoothly corrected by calculating the prior error covariance and Kalman gain, resulting in a more accurate updated boundary reliability scalar output.

[0116] In one possible embodiment, the step of outputting construction occupancy discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and use deviation discrimination results based on spatiotemporal semantic vectors, and updating the boundary reliability according to the boundary mismatch discrimination results to obtain the updated boundary reliability, includes:

[0117] The spatiotemporal semantic vector is input into a multi-branch fully connected classifier. The probability distribution vectors of the current plot in the dimensions of construction occupation, insufficient reclamation, boundary mismatch, and use deviation are calculated by the normalized exponential function. The category corresponding to the maximum probability is taken as the construction occupation discrimination result, insufficient reclamation discrimination result, boundary mismatch discrimination result, and use deviation discrimination result, respectively.

[0118] When the boundary mismatch judgment result is that a mismatch exists, calculate the probability offset of the boundary mismatch dimension, multiply the boundary reliability by the penalty factor calculated based on the probability offset, and perform weighted update to obtain the updated boundary reliability.

[0119] When there is no mismatch, the boundary reliability remains unchanged and is used as the updated boundary reliability.

[0120] The extracted 256-dimensional spatiotemporal semantic vectors are input in parallel into four independent multi-branch fully connected classifiers. Each branch is processed through a 128-dimensional hidden layer, and the corresponding binary classification probability distribution vector is output through the Softmax normalization exponential function. For example, the probability distributions for the construction occupancy dimension are 0.15 and 0.85. The Argmax function is used to extract the category indexes with probabilities exceeding the threshold of 0.5, generating Boolean-type discrimination results such as "construction occupancy exists," "no insufficient reclamation exists," and "boundary mismatch exists." When a boundary mismatch is detected and a mismatch is determined to exist, and the category probability is, for example, P=0.78, which is greater than 0.5, the difference between the category probability and the judgment threshold of 0.5 is calculated as a probability offset. =0.28. The penalty factor is calculated using the factor formula, with factor coefficients set. If the value is preferably 1.0, then the penalty factor K = 0.72. The boundary reliability output from the previous stage, such as 0.815, is multiplied by the penalty factor 0.72 to perform an explicit weight reduction and decay operation, resulting in an updated boundary reliability of 0.5868. Conversely, if the mismatch probability is less than 0.5, it is considered that the features are perfectly aligned, and the original boundary reliability value is maintained and directly output to the downstream module.

[0121] S8. The problem level of the land parcel is determined based on the updated boundary reliability, construction occupation judgment results, insufficient reclamation judgment results, boundary mismatch judgment results, and use deviation judgment results.

[0122] A land parcel problem level assessment model is constructed using the XGBoost extreme gradient boosting tree algorithm. The updated boundary reliability values ​​and the maximum classification probability extracted from the four discrimination results are concatenated into a one-dimensional assessment feature vector and input into the model. A comprehensive risk assessment score is obtained through node splitting calculations of multiple decision trees within the model. The Jenks natural breakpoint grading algorithm is then used to perform one-dimensional clustering of the comprehensive risk assessment score. The comprehensive risk assessment score is then discretized and mapped to four defined land parcel problem level labels: mild, moderate, severe, and critical.

[0123] In one possible embodiment, determining the land parcel problem level based on the updated boundary reliability, construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and land use deviation discrimination results includes:

[0124] Construct a level rule matrix containing multiple preset severity levels, and pre-assign severity weight coefficients to the construction occupation judgment results, insufficient reclamation judgment results, boundary mismatch judgment results, and use deviation judgment results;

[0125] The results of the statistical analysis of construction occupation, insufficient reclamation, boundary mismatch, and deviation of use are used to identify abnormal issues, and the severity weight coefficients of the corresponding issues are accumulated.

[0126] The cumulative severity weight coefficients are nonlinearly combined and mapped with the updated boundary reliability to calculate the comprehensive risk index of the land parcel.

[0127] The comprehensive risk index of the land parcel is compared with the threshold range in the level rule matrix to output the corresponding land parcel problem level.

[0128] A risk classification matrix is ​​constructed to categorize risks into four preset levels: mild, moderate, severe, and critical. The Delphi method is used to assign severity weight coefficients to anomaly-identified issues, setting the construction occupation coefficient at 0.4, use deviation at 0.3, insufficient reclamation at 0.2, and boundary mismatch at 0.1. Boolean state checks iterate through these four criteria; if a plot is determined to have both construction occupation and boundary mismatch, the weights corresponding to each are summed to obtain a total severity coefficient of 0.5. A nonlinear exponential combination function is used to calculate the comprehensive risk index of the plot, with the function defined as follows: Among them, smooth control hyperparameters The preferred value is 0.5. Substituting the updated boundary reliability of 0.5868 into the calculation yields the following result: A higher severity weighting coefficient indicates a more severe anomaly, hence the risk index is positively correlated with it; conversely, a lower boundary reliability reflects a more irregular or mismatched boundary, hence the risk index is negatively correlated with it, thus accurately characterizing the problem risk. The preset level rule matrix defines 0 to 0.25 as mild, 0.25 to 0.5 as moderate, 0.5 to 0.75 as severe, and 0.75 to 1.0 as critical. Logical branching statements are used to compare these intervals. If the risk index of 0.3725 falls within the moderate threshold range, the problem level of the target plot is output as moderate and written to the attribute database.

[0129] S9. Perform cluster analysis on adjacent target plots with consistent change trends, and determine the cluster boundary adjustment results based on the differences in characteristics within the cluster.

[0130] A spatial weight matrix for land parcels was constructed using the GeoPandas library. The Moran index was calculated to select a set of land parcels with intersecting spatial topologies and belonging to the same classification category. The Moran index is a classic indicator in spatial statistics used to measure the spatial autocorrelation of geographic data. It determines whether the spatial distribution of land parcel attribute values ​​is clustered, dispersed, or random. The DBSCAN density-based spatial clustering algorithm from the Scikit-Learn machine learning library was used to spatially cluster the selected land parcel set, resulting in multiple contiguous land parcel clusters. The area variance and perimeter range of polygons within each cluster were calculated as indicators of feature difference within the cluster. The polygon fusion function from the Shapely library was used to merge the boundaries of adjacent land parcels whose feature difference indicators were below a preset empirical threshold. A morphological opening and closing operation algorithm was used to smooth out internal holes and jagged edges generated after fusion. A vector coordinate set of adjusted cluster boundaries after spatial geometric topology optimization was generated.

[0131] In one possible embodiment, the step of performing cluster analysis on adjacent target plots with consistent changing trends, and determining the cluster boundary adjustment result based on the differences in characteristics within the cluster, includes:

[0132] The adjacency graph set of the target land parcel is retrieved using spatial topological relationships, and the cosine similarity between the target land parcel and each neighboring land parcel in the spatiotemporal semantic vector is calculated.

[0133] Select adjacent land parcels with a cosine similarity greater than a preset merging threshold and the same problem level, and use the connected component algorithm to merge the selected land parcels into a connected cluster.

[0134] Calculate the updated boundary reliability variance of each parcel within the connected cluster. If the variance is greater than a preset difference threshold, then combine the internal characteristics of the connected cluster to smooth the boundary of the parcel with the lowest updated boundary reliability. Based on the corrected parcel topology, regenerate the external contour vector of the connected cluster as the result of the cluster boundary adjustment.

[0135] In a GIS spatial database environment, the Touches or Intersects operators are used to traverse and retrieve all adjacent plots sharing edges or nodes of the target polygonal plot, constructing an undirected adjacency graph set. The 256-dimensional spatiotemporal semantic vectors generated from the target plot and its adjacent plots are extracted, and their inner product is calculated. The ratio of the norm product yields the cosine similarity within the range of -1 to 1. A preset merging threshold of 0.85 is set. Adjacent plots with a cosine similarity greater than 0.85 and assessed as having the same problem level (e.g., all being severe problems) in the upstream module are extracted. Using the Breadth-First Search (BFS) connected component algorithm in graph theory combined with the spatial Dissolve fusion operation, these plots are merged into a highly isomorphic connected cluster after eliminating common boundaries. The updated boundary reliability of each sub-plot within the connected cluster is statistically analyzed, and the variance is calculated. A preset difference threshold of 0.05 is set; if the calculated variance is greater than 0.05, it indicates a sudden change in boundary quality within the cluster. At this point, the boundary of the anomalous sub-plot with the lowest reliability within the cluster (e.g., only 0.21) is located and locked. The Douglas-Peucker algorithm combined with Gaussian filtering is used to resample and thin the node coordinates with a distance tolerance of 1.5 meters, followed by smoothing and transition. After the correction operation is completed, the convex hull or outer envelope polygon extraction algorithm is called to regenerate a closed and smooth outer contour coordinate vector group based on the updated plot node relationship. The vector group is then exported as a group boundary adjustment result in Shapefile or GeoJSON format.

[0136] S10 outputs the processing results, including the issue level of the land parcel and the adjustment results of the cluster boundaries.

[0137] The Fiona spatial data formatting output library is invoked to associate attribute table information with land parcel issue level labels with the original land parcel attribute fields. Simultaneously, the vector coordinate set of the cluster boundary adjustment results after spatial geometric topology optimization is serialized and converted into a unified geometric object model. This data is then packaged and exported as a Shapefile or GeoJSON standard spatial data file format conforming to Geographic Information System standards and saved to disk. This completes the persistent storage of the identification and classification processing results.

[0138] An experimental dataset was constructed based on 15,000 multi-temporal land parcel samples linked to land increase and decrease. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio and uniformly scaled to a resolution of 256×256 pixels. The hardware configuration consisted of two RTX 4090 graphics cards, using PyTorch as the deep learning framework and AdamW as the optimizer. The initial learning rate was set to 0.001 with a cosine annealing decay strategy, the batch size was set to 32, and the total number of iterations was 200.

[0139] To verify the rationality of the core module, four ablation models were set up. The baseline model only used basic remote sensing features as input to the standard GRU network and did not include attention fusion and constraint terms. Variant model A added boundary reliability as a feature fusion attention weight on the basis of the baseline. Variant model B used elevation difference and historical remediation type splicing constraints on the basis of the baseline. The complete model is the complete scheme of this embodiment. The results show that the overall problem classification accuracy of the baseline model is 82.5% and the crossover ratio of the group boundary is 0.76. The classification accuracy of variant model A reaches 87.3% and the crossover ratio increases to 0.84. The classification accuracy of variant model B is 86.9% and the crossover ratio is 0.80. The classification accuracy of the complete model reaches the highest 94.2%, and the crossover ratio of the group boundary is improved to 0.91.

[0140] Based on ablation data comparison, variant model A improved the crossover ratio (CROR) by calculating boundary reliability and adjusting the attention weight of directional texture and hierarchical distribution features, thus filtering noise interference in blurred edge regions. Variant model B utilized elevation differences and historical data as spatial and prior constraints for the temporal network, enhancing its ability to extract subsidence and reclamation change patterns and steadily increasing the discrimination accuracy. The complete model combines the advantages of both and utilizes a mismatch weighting penalty mechanism, achieving complementary advantages between semantic features and temporal patterns. It exhibits the best performance improvement in complex plot classification and cluster contour refinement tasks, such as... Figure 3 As shown in the figure, this is a bar chart comparing the ablation experiment results of the embodiments of the present invention. The figure shows the comprehensive classification accuracy and boundary intersection-union ratio of the baseline model, variant model A with added boundary reliability attention weights, variant model B using terrain and historical prior constraints, and the complete model of the present invention. The comparison demonstrates the performance improvement of the present invention in the tasks of land parcel problem identification and boundary optimization.

[0141] Secondly, the present invention also provides a dynamic recognition and hierarchical processing system for addition and subtraction hook problems based on image semantics, comprising:

[0142] The extraction module is used to acquire multi-source data of the land parcel increase / decrease linkage area and perform coordinate unification, temporal alignment and grid encoding. It segments the remote sensing image according to the land parcel boundary, extracts semantic features inside the land parcel, semantic features of the boundary neighborhood and shape features. It calculates the boundary reliability based on the difference between the semantic features inside the land parcel and the semantic features of the boundary neighborhood, the boundary continuity of the land parcel and the shape regularity. It constructs the semantic representation of the land parcel by combining the shape features. It inputs the semantic representation of the land parcel into the dual-path model, extracts directional texture features and hierarchical distribution features, and adjusts the weights of directional texture features and hierarchical distribution features according to the boundary reliability and merges them into structural semantic features.

[0143] The update module is used to combine terrain elevation constraints and historical data to extract temporal change features from multi-temporal structural semantic features, generate spatiotemporal semantic vectors, output construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results and use deviation discrimination results based on the spatiotemporal semantic vectors, and update the boundary reliability according to the boundary mismatch discrimination results to obtain the updated boundary reliability.

[0144] The output module is used to determine the problem level of a land parcel based on the updated boundary reliability, construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and use deviation discrimination results. It performs cluster analysis on adjacent target land parcels with consistent change trends, determines the cluster boundary adjustment results based on the differences in characteristics within the cluster, and outputs the processing results containing the land parcel problem level and the cluster boundary adjustment results.

[0145] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0146] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for dynamic identification and hierarchical processing of addition and subtraction hook problems based on image semantics, characterized in that, include: Multi-source data of the land increase / decrease linkage area were acquired and subjected to coordinate unification, temporal alignment and gridding encoding. Remote sensing images were segmented according to the land parcel boundaries, and semantic features of the land parcel interior, boundary neighborhood semantic features and shape features were extracted. Boundary reliability was calculated based on the difference between the semantic features of the land parcel interior and the semantic features of the boundary neighborhood, the boundary continuity of the land parcel and the shape regularity. The semantic representation of the land parcel was constructed by combining the shape features. The semantic representation of the land parcel was input into a dual-path model to extract directional texture features and hierarchical distribution features. The weights of directional texture features and hierarchical distribution features were adjusted according to the boundary reliability and fused into structural semantic features. Combining topographic elevation constraints and historical data, temporal change features are extracted from multi-temporal structural semantic features to generate spatiotemporal semantic vectors. Based on the spatiotemporal semantic vectors, the results of construction occupation discrimination, insufficient reclamation discrimination, boundary mismatch discrimination, and use deviation discrimination are output. The boundary reliability is updated according to the boundary mismatch discrimination results to obtain the updated boundary reliability. Based on the updated boundary reliability, construction occupation judgment results, insufficient reclamation judgment results, boundary mismatch judgment results, and use deviation judgment results, the problem level of the land parcel is determined. Adjacent target land parcels with consistent change trends are grouped for analysis. The group boundary adjustment results are determined based on the differences in characteristics within the group. The output includes the land parcel problem level and the group boundary adjustment results.

2. The method according to claim 1, characterized in that, The process of acquiring multi-source data for the increase / decrease hook area and performing coordinate unification, time-relative alignment, and grid encoding includes: Extract spatial reference information from multi-source data, and uniformly convert remote sensing images, land parcel boundaries, and terrain elevations from different coordinate systems to the National Geodetic Coordinate System for spatial projection alignment; Establish a time series reference axis, set a time matching window, remove incomplete temporal images with cloud and snow coverage exceeding a preset threshold, and interpolate and align the retained multi-temporal remote sensing images and terrain elevations according to the acquisition timestamps. A two-dimensional grid matrix is ​​constructed with a preset fixed spatial resolution. The alignment data covering the increase / decrease hook area is divided into independent grid cells of the same size, and a unique and non-repeating spatial index string is assigned to each grid cell to generate a gridded data packet.

3. The method according to claim 2, characterized in that, The calculation of boundary reliability based on the difference between the semantic features within the land parcel and the semantic features of its boundary neighborhood, the boundary continuity of the land parcel, and the regularity of its shape includes: Extract the feature vector center of the pixel set inside the plot as the semantic feature inside the plot, extract the feature vector center of the pixel set within a preset buffer distance extending outward from the plot boundary as the semantic feature of the boundary neighborhood, calculate the cosine distance between the semantic feature inside the plot and the semantic feature of the boundary neighborhood, and take this distance as the difference degree. An edge detection algorithm is used to identify the actual number of connected pixels at the boundary of a land parcel, and the ratio of the actual number of connected pixels to the total number of pixels in the theoretical boundary perimeter is used as the boundary continuity of the land parcel. The ratio of the actual area of ​​the plot to the area of ​​its convex hull is used as the shape regularity. Preset weights are assigned to the degree of difference, the boundary continuity of the plot, and the shape regularity. The three indicators of each plot are summed by weight to obtain a weighted composite index. The weighted composite index is then mapped to the (0,1) interval by the Sigmoid function to obtain the boundary reliability of the plot.

4. The method according to claim 1, characterized in that, The process of adjusting the weights of directional texture features and hierarchical distribution features according to boundary reliability and fusing them into structural semantic features includes: By using boundary reliability as the attention weight for directional texture features, element-wise multiplication is performed to obtain weighted directional texture features. The complementarity coefficient is obtained by subtracting the boundary reliability from the constant 1. The complementarity coefficient is used as the attention weight of the hierarchical distribution feature, and element-wise multiplication is performed to obtain the weighted hierarchical distribution feature. Weighted directional texture features and weighted hierarchical distribution features are concatenated along the channel dimension, and then dimensionality reduction and feature mapping are performed through a fully connected layer to generate structural semantic features.

5. The method according to claim 1, characterized in that, The process of combining terrain elevation constraints and historical data to extract temporal change features from multi-temporal structural semantic features and generate spatiotemporal semantic vectors includes: The feature sequence is obtained by arranging the semantic features of the multi-temporal structure in chronological order. Calculate the topographic elevation difference of each time phase relative to the initial time phase, use the elevation difference as a spatial constraint term, and concatenate it with the remediation type features extracted from historical data into the corresponding feature sequence node; The concatenated feature sequence is input into a long short-term memory network or a gated recurrent unit network. The historical time series state and the current time phase features are filtered and updated through the gating mechanism, and the hidden layer state vector is extracted as a spatiotemporal semantic vector.

6. The method according to claim 1, characterized in that, The method outputs construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and use deviation discrimination results based on spatiotemporal semantic vectors, and updates the boundary reliability based on the boundary mismatch discrimination results to obtain the updated boundary reliability, including: The spatiotemporal semantic vector is input into a multi-branch fully connected classifier. The probability distribution vectors of the current plot in the dimensions of construction occupation, insufficient reclamation, boundary mismatch, and use deviation are calculated by the normalized exponential function. The category corresponding to the maximum probability is taken as the construction occupation discrimination result, insufficient reclamation discrimination result, boundary mismatch discrimination result, and use deviation discrimination result, respectively. When the boundary mismatch judgment result is that a mismatch exists, calculate the probability offset of the boundary mismatch dimension, multiply the boundary reliability by the penalty factor calculated based on the probability offset, and perform weighted update to obtain the updated boundary reliability. When there is no mismatch, the boundary reliability remains unchanged and is used as the updated boundary reliability.

7. The method according to claim 1, characterized in that, The determination of the land parcel problem level based on the updated boundary reliability, construction occupation judgment results, insufficient reclamation judgment results, boundary mismatch judgment results, and land use deviation judgment results includes: Construct a level rule matrix containing multiple preset severity levels, and pre-assign severity weight coefficients to the construction occupation judgment results, insufficient reclamation judgment results, boundary mismatch judgment results, and use deviation judgment results; The results of the statistical analysis of construction occupation, insufficient reclamation, boundary mismatch, and deviation of use are used to identify abnormal issues, and the severity weight coefficients of the corresponding issues are accumulated. The cumulative severity weight coefficients are nonlinearly combined and mapped with the updated boundary reliability to calculate the comprehensive risk index of the land parcel. The comprehensive risk index of the land parcel is compared with the threshold range in the level rule matrix to output the corresponding land parcel problem level.

8. The method according to claim 1, characterized in that, The process of performing cluster analysis on adjacent target plots with consistent changing trends, and determining the cluster boundary adjustment results based on the differences in characteristics within each cluster, includes: The adjacency graph set of the target land parcel is retrieved using spatial topological relationships, and the cosine similarity between the target land parcel and each neighboring land parcel in the spatiotemporal semantic vector is calculated. Select adjacent land parcels with a cosine similarity greater than a preset merging threshold and the same problem level, and use the connected component algorithm to merge the selected land parcels into a connected cluster. Calculate the updated boundary reliability variance of each parcel within the connected cluster. If the variance is greater than a preset difference threshold, then combine the internal characteristics of the connected cluster to smooth the boundary of the parcel with the lowest updated boundary reliability. Based on the corrected parcel topology, regenerate the external contour vector of the connected cluster as the result of the cluster boundary adjustment.

9. A dynamic recognition and hierarchical processing system for addition and subtraction hook problems based on image semantics, characterized in that, include: The extraction module is used to acquire multi-source data of the land parcel increase / decrease linkage area and perform coordinate unification, temporal alignment and grid encoding. It segments the remote sensing image according to the land parcel boundary, extracts semantic features inside the land parcel, semantic features of the boundary neighborhood and shape features. It calculates the boundary reliability based on the difference between the semantic features inside the land parcel and the semantic features of the boundary neighborhood, the boundary continuity of the land parcel and the shape regularity. It constructs the semantic representation of the land parcel by combining the shape features. It inputs the semantic representation of the land parcel into the dual-path model, extracts directional texture features and hierarchical distribution features, and adjusts the weights of directional texture features and hierarchical distribution features according to the boundary reliability and merges them into structural semantic features. The update module is used to combine terrain elevation constraints and historical data to extract temporal change features from multi-temporal structural semantic features, generate spatiotemporal semantic vectors, output construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results and use deviation discrimination results based on the spatiotemporal semantic vectors, and update the boundary reliability according to the boundary mismatch discrimination results to obtain the updated boundary reliability. The output module is used to determine the problem level of a land parcel based on the updated boundary reliability, construction occupation discrimination results, insufficient reclamation discrimination results, boundary mismatch discrimination results, and use deviation discrimination results. It performs cluster analysis on adjacent target land parcels with consistent change trends, determines the cluster boundary adjustment results based on the differences in characteristics within the cluster, and outputs the processing results containing the land parcel problem level and the cluster boundary adjustment results.

10. The system according to claim 9, characterized in that, The process of acquiring multi-source data for the increase / decrease hook area and performing coordinate unification, time-relative alignment, and grid encoding includes: Extract spatial reference information from multi-source data, and uniformly convert remote sensing images, land parcel boundaries, and terrain elevations from different coordinate systems to the National Geodetic Coordinate System for spatial projection alignment; Establish a time series reference axis, set a time matching window, remove incomplete temporal images with cloud and snow coverage exceeding a preset threshold, and interpolate and align the retained multi-temporal remote sensing images and terrain elevations according to the acquisition timestamps. A two-dimensional grid matrix is ​​constructed with a preset fixed spatial resolution. The alignment data covering the increase / decrease hook area is divided into independent grid cells of the same size, and a unique and non-repeating spatial index string is assigned to each grid cell to generate a gridded data packet.

Images