A crop planting map patch intelligent extraction system based on remote sensing information

Abstract

The application relates to the technical field of agricultural remote sensing information processing, and particularly discloses a crop planting map patch intelligent extraction system based on remote sensing information, which realizes automatic correction of abnormal fragments, large patches and topological conflicts by fusing a map patch area, compactness, time sequence vegetation index fluctuation and field ridge slope variation coefficient, constructing a multi-scale feature vector, and combining dynamic threshold adjustment and spatial semantic verification. In complex scenes such as terraces in Yunnan, the system significantly improves map patch boundary precision and topological rationality through vertical field ridge error response suppression and dynamic map reconstruction, and outputs a classification result in a GeoJSON format.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing information processing technology, specifically to an intelligent extraction system for crop planting patches based on remote sensing information. Background Technology

[0002] With the acceleration of agricultural modernization, accurate monitoring of arable land distribution and crop planting patterns has become a core requirement for ensuring food security and optimizing land resource management.

[0003] Traditional spatial feature extraction methods suffer from severe topological distortion in complex agricultural landforms due to a lack of semantic understanding. Specifically, in special scenarios such as terraced fields and irrigation networks, the Canny operator or region growing algorithm relies solely on pixel gradients or spectral similarity for segmentation, failing to distinguish the semantic boundaries between agricultural infrastructure (field ridges, irrigation ditches) and actual planted plots, leading to two types of fatal errors:

[0004] Over-segmentation (e.g., terraced areas): misjudging the continuous strong edges of contour lines as plot boundaries, resulting in individual terraced fields being cut into fragmented patches, disrupting the overall topological structure of farmland;

[0005] Undersegmentation (e.g., irrigation areas): The region growing algorithm spreads rapidly along the irrigation canal, merging cross-canal farmland because the canal water has a similar spectrum to the surrounding crops, generating invalid giant patches covering multiple plots, completely obscuring the independence of the actual planting units. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent extraction system for crop planting patches based on remote sensing information, so as to solve the problems mentioned above.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A smart system for extracting crop planting patches based on remote sensing information includes:

[0009] The topology graph construction module is used to acquire remote sensing image data and generate initial segmented patches, and construct a topology graph structure with patches as nodes and spatial adjacency relationships as edges;

[0010] The dual-channel graph neural network module, connecting to the topology graph construction module, includes:

[0011] Structural analysis sub-units identify anomalous fragments or giant spots based on patch shape and boundary curvature;

[0012] The semantic verification subunit integrates temporal vegetation index and elevation data to verify the phenological consistency of adjacent patches;

[0013] The topology rule constraint module, connected to the dual-channel graph neural network module, is used to generate loss signals corresponding to farmland topology rules and optimize the boundary features of the segmentation model through gradient penalty.

[0014] The dynamic graph reconstruction module, connected to the topology rule constraint module, performs the following operations:

[0015] Merge highly similar adjacent nodes;

[0016] Cut off internal topological conflict nodes;

[0017] Output the corrected vector graphic pattern;

[0018] The iterative optimization module connects to the dynamic graph reconstruction module, and iteratively calls the dual-channel graph neural network module to the dynamic graph reconstruction module until the graph structure is stable, generating a crop planting distribution map.

[0019] As a further aspect of the present invention, the specific steps of the topology graph construction module are as follows:

[0020] Calculate abrupt slope regions using digital elevation model data to generate elevation gradient masks;

[0021] Remove the boundaries of plots with slope changes exceeding the threshold from the initial segmentation patch to eliminate false edges caused by terraced fields;

[0022] Extract the temporal vegetation index curves of adjacent patches and calculate the dynamic temporal regularization distance;

[0023] When the distance is less than the synchronization threshold, establish a spatial adjacency edge;

[0024] When the distance exceeds the asynchronous threshold, the original adjacency relationship is severed.

[0025] Compress the geometric features of the patch into a vector of compactness and aspect ratio;

[0026] By integrating phenological feature vectors and elevation statistical feature vectors, node semantic descriptors are generated.

[0027] Calculate the connection weight of adjacent edges based on the shared boundary length and phenological similarity;

[0028] Apply a penalty weight to edges that cross irrigation canals to suppress incorrect connections across the canals.

[0029] As a further aspect of the present invention, the specific implementation process of the structural analysis subunit is as follows:

[0030] Calculate the curvature sequence of boundary points of the patch and quantify the boundary irregularity using information entropy;

[0031] When the entropy value exceeds the dynamic threshold, it is judged as abnormal fragmentation;

[0032] Generate variable-size structural elements based on the area of ​​the map patch;

[0033] Perform multi-scale etching operations on giant spots to separate the adhered areas;

[0034] Detect the boundaries of nested patches and trigger cutting based on the overlap of bounding boxes and the area of ​​internal voids;

[0035] Fragmented regions are labeled as negative samples and giant spot regions are labeled as positive samples, generating a binary mask that is fed back to the segmentation network.

[0036] As a further aspect of the present invention, the specific implementation process of the semantic verification subunit is as follows:

[0037] Extract the temporal vegetation index within the growth cycle of each patch, and generate a phenological feature vector through one-dimensional convolution.

[0038] Align the elevation zone data with the phenological vectors according to the grid cells, and calculate the elevation variation coefficient to correct the phenological similarity.

[0039] Calculate the dynamic time curvature distance for the phenological feature vectors of adjacent patches;

[0040] When the distance is greater than the asynchronous threshold and the elevation difference is less than the tolerance, it is judged as an incorrect merging across crops;

[0041] Disconnect the edges of phenological conflict and reduce their weight to 10% of their original value.

[0042] As a further aspect of the present invention, the specific implementation process of the topology rule constraint module is as follows:

[0043] Establish a semantic coding library for field ridges and irrigation ditches to vectorize linear structures in high-resolution images;

[0044] When the overlap between the segment boundary and the facility vector line exceeds a threshold, it is marked as an erroneous connection to be penalized;

[0045] Penalty weights are assigned based on the type of incorrect connection: the weight coefficient for misidentification of field ridge boundaries is the reciprocal of the facility width, and the weight coefficient for mismerging of water channels is the square of the canal network density;

[0046] During the backpropagation phase of the segmentation model, the penalty weights are converted into gradient correction amounts for the boundary feature layers.

[0047] Apply facility orientation constraints to the gradient matrix of the convolution kernel to suppress convolution responses perpendicular to the direction of the field ridge.

[0048] As a further aspect of the present invention, the specific steps of the gradient penalty optimization are as follows:

[0049] Topology penalty is applied deep to the encoder layer of the segmentation network, and boundary detail penalty is applied shallow to the decoder layer.

[0050] Calculate the distance from the faulty edge to the nearest farmland facility;

[0051] The gradient penalty intensity is reduced by the inverse of the distance to achieve soft constraints on the facility's influence domain;

[0052] The loss signal output by the topology rule constraint module is weighted and fused with the original loss of the segmentation network;

[0053] The fusion weight decreases exponentially with each training round, starting at 0.8 and decaying to half its original value every 50 rounds.

[0054] As a further aspect of the present invention: the merging of highly similar adjacent nodes specifically includes:

[0055] A three-dimensional similarity vector is constructed based on the geometric feature similarity of adjacent nodes, the correlation coefficient of phenological curves, and the consistency of elevation.

[0056] The shared boundary length is used as a weighting factor to calculate the overall similarity score;

[0057] The merging threshold is dynamically adjusted based on the degree of farmland fragmentation in the region: the threshold decreases by 0.05 for every 10% increase in the fragmentation index;

[0058] The minimum bounding polygon is used to merge nodes, preserving the original image patch attributes;

[0059] Add smoothing constraints at the merge boundary to eliminate jagged edges;

[0060] If irrigation canal vector lines exist between nodes to be merged, the merging operation will be forcibly terminated.

[0061] As a further aspect of the present invention: the cutting of internal topological conflict nodes specifically includes:

[0062] The abnormal probability map output by the dual-channel graph neural network is used to mark the boundary curvature abrupt change area and the phenological anomaly area;

[0063] Create a cutting baseline along the direction of the field ridge or the vertical line of the canal network;

[0064] The cutting angle is corrected based on vector data of farmland facilities to ensure that the path is parallel to agricultural infrastructure.

[0065] Proceeding from the point of conflict along the cutting baseline, stopping when encountering a region of consistent phenology;

[0066] Distribute the original node attributes to the new polygons according to the area ratio;

[0067] Rebuild adjacency relationships and update the topology graph structure.

[0068] As a further aspect of the present invention, the specific implementation steps of the iterative optimization module are as follows:

[0069] Simultaneously monitor the rate of change of the number of patches, the entropy value of the boundary curvature distribution, and the Frobenius norm of the phenological feature similarity matrix;

[0070] The graph structure is considered stable when all three changes are below the dynamic threshold.

[0071] The learning rate is adjusted based on the differences between nodes before and after topology reconstruction.

[0072] For every 10% increase in variance, the learning rate decays to 80% of its original value.

[0073] High-conflict nodes in the historical iterations are marked and prioritized for processing by the dual-channel graph neural network module in the next loop.

[0074] Store the intermediate topology graph structure for each iteration, and roll back to the optimal cached version when convergence fails.

[0075] As a further aspect of the present invention: the generation of the crop planting distribution map specifically includes:

[0076] By tracing the segmentation trajectory of the patch through the graph structure version chain, it is ensured that the final attributes are inherited from the initial classification result;

[0077] Check the geometric non-intersection and attribute mutual exclusion of adjacent patches, and repair overlapping and gap defects;

[0078] Irrigation canals, field ridges and other farmland facilities are overlaid as independent layers to preserve the topological relationships of agricultural infrastructure.

[0079] A confidence heatmap is generated based on the number of iterations and the degree of conflict resolution, and then integrated into the distribution map attribute table.

[0080] The beneficial effects of this invention are:

[0081] (1) This invention constructs an intelligent extraction system for crop planting patches by fusing dual-channel graph neural networks with multi-scale features (geometry, phenology, and elevation), combined with dynamic threshold adjustment and spatial semantic verification. In complex terrain scenarios such as terraced fields in Yunnan, this system significantly improves the accuracy and topological rationality of patch boundaries through vertical field ridge error response suppression, dynamic graph reconstruction (merging or cutting conflict nodes), and cross-ditch edge weight penalty. The final output GeoJSON classification, through multi-source data fusion and adaptive optimization strategies, enhances robustness and generalization ability in complex environments, providing an efficient and accurate solution for agricultural remote sensing.

[0082] (2) This invention achieves adaptive processing of diverse terrains such as hills and terraces by using multi-source data fusion (such as 30-meter resolution digital elevation model, time-series vegetation index, and irrigation canal vector layer) and dynamic iterative optimization strategies (convergence determination based on boundary entropy value and phenological similarity matrix difference). The system's built-in version control, conflict node priority evaluation, and dynamic learning rate decay mechanism ensure robustness and stability under different environmental conditions. For example, in areas with high canal network density, the system solves the problem of incorrect patch merging caused by terrain occlusion or facility interference caused by traditional methods by using cross-canal edge weight penalty and facility semantic constraints (such as matching the direction angle of field ridges). Attached Figure Description

[0083] The invention will now be further described with reference to the accompanying drawings.

[0084] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0085] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0086] Please see Figure 1 As shown, this invention is an intelligent extraction system for crop planting patches based on remote sensing information, comprising:

[0087] The topology graph construction module is used to acquire remote sensing image data and generate initial segmented patches, and construct a topology graph structure with patches as nodes and spatial adjacency relationships as edges;

[0088] The dual-channel graph neural network module, connecting to the topology graph construction module, includes:

[0089] Structural analysis sub-units identify anomalous fragments or giant spots based on patch shape and boundary curvature;

[0090] The semantic verification subunit integrates temporal vegetation index and elevation data to verify the phenological consistency of adjacent patches;

[0091] The topology rule constraint module, connected to the dual-channel graph neural network module, is used to generate loss signals corresponding to farmland topology rules and optimize the boundary features of the segmentation model through gradient penalty.

[0092] The dynamic graph reconstruction module, connected to the topology rule constraint module, performs the following operations:

[0093] Merge highly similar adjacent nodes;

[0094] Cut off internal topological conflict nodes;

[0095] Output the corrected vector graphic pattern;

[0096] The iterative optimization module connects to the dynamic graph reconstruction module, and iteratively calls the dual-channel graph neural network module to the dynamic graph reconstruction module until the graph structure is stable, generating a crop planting distribution map.

[0097] In the topology map construction module, 30-meter resolution digital elevation model data is acquired from a public satellite imagery platform. Projection transformation and resampling are performed on the data to align it spatially with the remote sensing imagery. A 3×3 pixel sliding window is used to calculate the slope change rate for each pixel. Specifically, the sum of the absolute values ​​of the elevation differences between the center pixel and its eight adjacent pixels is calculated. If this value exceeds the threshold corresponding to a 30-degree slope (empirical value is 4.5 meters), it is marked as a slope abrupt change area. A binary mask layer is generated based on the marking results, with a value of 0 for abrupt changes and 1 for non-abrupt changes. The mask is overlaid onto the boundary raster layer of the initial segmented patches, and a logical AND operation is performed: when a boundary pixel is located in an area with a mask value of 0, it is determined to be a false edge caused by terraced fields and is permanently deleted from the boundary set.

[0098] Temporal vegetation index curves of adjacent patches within a complete growth cycle are extracted, with each curve containing no fewer than 12 temporal observations (corresponding to one scene of data per month). A dynamic time warping algorithm is used to calculate the distance between curves: first, the two curves are non-linearly aligned in the time dimension; then, the absolute values ​​of the vegetation index differences at corresponding points after alignment are accumulated; finally, the accumulated value is divided by the maximum time warp path length to obtain the normalized distance. A synchronicity threshold of 0.15 and an asynchronicity threshold of 0.35 are set: if the normalized distance between adjacent patches is less than 0.15, a spatial adjacency edge is forcibly established; if the distance is greater than 0.35, the existing adjacency edge is cut.

[0099] Calculate three types of features for each patch:

[0100] Geometric characteristics: area (unit: hectares), compactness (ratio of the square of the perimeter to 4π times the area), aspect ratio (ratio of the major axis to the minor axis of the smallest circumscribed rectangle).

[0101] Phenological characteristics: mean and variance of temporal vegetation indices, and the timing of peak occurrence during the growing season;

[0102] Elevation characteristics: maximum, minimum, and standard deviation of internal elevation.

[0103] The three types of features are normalized to the [0,1] interval, and then concatenated into a 9-dimensional vector using 3D geometric features, 3D phenological features, and 3D elevation features. Feature compression is performed through a fully connected layer with 5 neurons, outputting a 5-dimensional node semantic descriptor. Adjacent edge weight calculation and cross-canal penalty: The initial weight calculation process for adjacent edges is as follows: the initial weight is obtained by summing the product of 0.7 and the ratio of the shared boundary length to the sum of the perimeters of the two patches, and the product of 0.3 and the cosine similarity of the phenological features.

[0104] Simultaneously, a pre-generated irrigation canal vector layer is loaded: when an adjacent edge intersects with a canal line, a penalty is applied based on the canal network density. The canal network density is defined as the canal line length per square kilometer (unit: km / km²), and the penalty coefficient is (1 - canal network density). The final edge weights are updated by multiplying the original weights by the penalty coefficient.

[0105] Construct a topology graph structure that includes the following elements:

[0106] Node attributes: geographic coordinates, area, 5-dimensional semantic descriptor;

[0107] Edge attributes: adjacency type (natural adjacency or phenological forced connection), weight value, and whether it crosses a channel;

[0108] Output graph structure files in GraphML format.

[0109] In the dual-channel graph neural network module, the curvature sequence of each patch boundary point is calculated: first, the boundary pixel coordinate chain is extracted, and the cosine of the included angle formed by three adjacent points is calculated sequentially as the curvature value, generating a curvature sequence. Boundary irregularity is quantified using information entropy: after normalizing the curvature value to the [0,1] interval, it is divided into 10 equal-width intervals. The frequency of curvature values ​​in each interval is statistically analyzed, and the information entropy value is calculated according to the Shannon entropy formula. A dynamic threshold is set as the sum of the base entropy value of 1.8 and the area adjustment term (the threshold increases by 0.05 for every 1 hectare increase in area). When the entropy value exceeds the dynamic threshold, it is determined to be an abnormal fragment.

[0110] Variable-size structural elements are generated based on the area of ​​the patches: patches with an area less than 0.5 hectares use 3×3 rectangular structural elements, patches with an area of ​​0.5-5 hectares use 5×5 structural elements, and patches with an area greater than 5 hectares use 7×7 structural elements. Multi-scale erosion is performed on giant patches: first, erosion is performed once using the smallest structural element to separate small, adhered areas; then, erosion is performed twice using medium-sized structural elements to separate moderately adhered areas; finally, erosion is performed once using the largest structural element to resolve large-area adhesion. After each erosion, independent regions are marked using eight-neighbor connectivity analysis.

[0111] Detecting nested feature boundaries: Calculate the overlap (overlap area as a percentage of the area of ​​the outer feature) of the smallest bounding rectangle between the outer and inner features. A nested feature is considered valid when the overlap is greater than 70% and the area of ​​the inner feature is greater than 20% of the area of ​​the outer feature. Triggering cutting based on internal void area: If the ratio of the void area to the area of ​​the nested feature exceeds 0.3, a cutting operation is performed along the nesting boundary. The cutting path generation rule is: draw a perpendicular line from the centroid of the nested feature to the nearest boundary point as the cutting baseline.

[0112] Fragmented areas are labeled as negative samples (assigned a value of 0), and giant speckle areas are labeled as positive samples (assigned a value of 1), generating a binary mask. The mask is then converted into a raster layer with the same resolution as the original image through spatial coordinate mapping and overlaid onto the segmentation network output layer. During backpropagation, a 20x weight loss penalty is applied to negative sample areas, and a 10x weight loss penalty is applied to positive sample areas.

[0113] Extract vegetation index curves for at least 12 time phases within the growth cycle of each patch. Generate phenological feature vectors using a three-layer one-dimensional convolution: the first layer uses a convolution kernel with a width of 3 to extract local fluctuation features; the second layer uses a convolution kernel with a width of 5 to extract growth stage features; and the third layer uses global max pooling to compress the temporal dimension. Output a 32-dimensional phenological feature vector. Resample the elevation zonation data to the same spatial resolution as the phenological features (e.g., 10 meters). Align the elevation values ​​with the phenological feature vectors by raster cells and calculate the elevation variation coefficient (the ratio of standard deviation to mean) within each patch. The phenological similarity correction formula is: original similarity × (1 - 0.5 × elevation variation coefficient).

[0114] The dynamic temporal bending distance is calculated for the phenological feature vectors of adjacent patches: an improved algorithm with a step size constraint of 1 is used to accumulate the minimum bending path cost. An asynchronous threshold of 0.4 and an elevation tolerance of 5 meters are set. When the bending distance is greater than 0.4 and the average elevation difference between the two patches is less than 5 meters, it is determined to be an incorrect cross-crop merging.

[0115] Disconnect the physical links of phenological conflict edges: Delete the corresponding edge record in the topological graph data structure. Reduce the weight of the remaining edges to 10% of their original value: if the original weight is W, the updated weight is 0.1W. The updated edge weights participate in subsequent graph neural network message passing calculations.

[0116] In the topology rule constraint module, sub-meter high-resolution remote sensing images are acquired, and a pre-trained U-Net model is used to identify field ridges and irrigation ditches. Field ridge identification is based on their linear shape and topographic features above the field surface, while irrigation ditch identification is based on their regular ditch shape and water spectral characteristics. The identification results are vectorized: field ridges are output as polyline features with a width of 0.3–1.5 meters, and irrigation ditches are output as areal features with a width of 1–5 meters. A semantic encoding library is established to store facility type, geometric centerline coordinates, and width attributes.

[0117] Calculate the spatial overlap between the segment boundary line and the facility vector: Generate a 5-pixel buffer centered on the boundary line, and calculate the proportion of the overlap length between the buffer and the facility vector line to the total length of the boundary line. When the overlap exceeds the threshold (field ridge: 15%; irrigation ditch: 20%), it is marked as an erroneous connection to be penalized. The marking results are stored as triplet data: boundary segment ID, facility type, overlap.

[0118] Assign weight coefficients based on the type of faulty connection:

[0119] Misidentification of field ridge boundaries: The reciprocal of the facility's width is used as the weighting coefficient. For example, a field ridge with a width of 0.5 meters has a weighting coefficient of 2.0.

[0120] Incorrect merging of waterways: The square of the canal network density is used as the weighting coefficient. The canal network density is defined as the length of the canal line within 1 square kilometer (unit: km / km²). For example, when the density is 2.5, the coefficient is 6.25.

[0121] The weight values ​​are written to the fourth field of the triplet data. During the backpropagation phase of the segmentation model, the boundary feature layer (i.e., the convolutional layer of the output boundary probability map) is located. The penalty weights are converted into gradient corrections: for each erroneous connection point, the penalty weight value is added to its gradient matrix. The correction is the product of the original gradient value and 1 plus the weight coefficient. For example, the gradient correction for the misidentified field ridge point (weight 2.0) is 3 times the original value.

[0122] Extract the orientation angle (0-180 degrees) of the centerline of the field ridge. Apply directional constraints to the gradient matrix of the convolution kernels of the boundary feature layers:

[0123] Calculate the angle between the direction of the convolution kernel response and the direction of the field ridge;

[0124] When the included angle approaches 90 degrees (vertical direction), the gradient value is reduced to 30% of its original value;

[0125] When the included angle is less than 30 degrees (parallel direction), the original gradient value is retained.

[0126] Application in Yunnan's terraced fields shows that this constraint reduces erroneous responses perpendicular to the field ridges by 83%.

[0127] Apply differentiated penalties at different levels of the segmentation network:

[0128] Encoder depth (receptive field greater than 100 pixels): For topological errors, the penalty weight is amplified to 1.5 times;

[0129] Decoder shallow layer (receptive field less than 20 pixels): For boundary detail errors, the penalty weight remains the same.

[0130] When calculating gradients, the gradient matrix of deep feature maps is multiplied by a coefficient of 1.5, while the coefficient of shallow feature maps is 1.0.

[0131] Calculate the Euclidean distance from the faulty edge to the nearest farmland facility. Decrease the gradient penalty strength by the inverse of the distance: the penalty weight decreases by 10% for every 1 pixel increase in distance. For example, when the distance is 5 pixels, the weight decreases to 50% of its original value. Set a maximum influence range of 10 pixels; no penalty is applied outside this range.

[0132] In the dynamic graph reconstruction module, three types of features are extracted from adjacent nodes: geometric features include area ratio (area of ​​the larger patch divided by the area of ​​the smaller patch), compactness difference (absolute difference in compactness between the two patches), and aspect ratio cosine similarity; phenological features are the dynamic time-normalized distance of the time-series vegetation index curve; and elevation features are the average elevation difference between the two patches. These three types of features are normalized to the [0,1] interval and then concatenated into a three-dimensional vector. The shared boundary length weighting method is as follows: calculate the proportion of the shared boundary length between the two patches to the perimeter of the smaller patch, and use this as the weighting coefficient multiplied by the modulus of the three-dimensional vector.

[0133] Calculate the regional farmland fragmentation index: Count the number of patches N per square kilometer. The index value is the ratio of N to the baseline value (100 patches / km² in plains areas and 150 patches / km² in hilly areas). The initial merging threshold is set to 0.7. When the fragmentation index increases by 10%, the threshold decreases linearly by 0.05, with a minimum of 0.4. The threshold adjustment calculation process is as follows: First, obtain the difference between the current farmland fragmentation index and the preset baseline index. Divide this difference by 0.1 and round it down. Then multiply it by an adjustment step size of 0.05. Finally, subtract this adjustment amount from the baseline threshold of 0.7.

[0134] An improved Sutherland-Hodgman algorithm is used to calculate the minimum bounding polygon of nodes: first, the convex hulls of the boundaries of two patches are constructed, and then overlapping areas are preserved through line segment clipping. The attribute inheritance rule is: after merging, patches inherit the crop type of the one with the larger area; if the area difference is less than 20%, it is marked as a mixed type. Boundary smoothing is performed using cubic B-spline curve fitting, with the control point spacing set to 5 pixels.

[0135] Load the irrigation canal vector layer and create a spatial index (R-tree). For node pairs to be merged, calculate the number of intersections between the centroids of the two patches and the canal network. Terminate the merge immediately if the number of intersections is greater than or equal to 1. In special cases (canal width less than 1 meter and phenological similarity between the two patches greater than 0.9), merging is allowed, but the merge priority is reduced to the lowest possible level.

[0136] The system receives anomaly probability maps from a dual-channel GNN (channel 1 for geometric anomalies and channel 2 for phenological anomalies). For each patch, the following steps are performed: calculate the mean geometric anomaly probability of the boundary pixels; if it is greater than 0.7, it is marked as a curvature abrupt change region; calculate the maximum phenological anomaly probability of the internal pixels; if it is greater than 0.8, it is marked as a phenological anomaly region. The two types of regions are overlaid to generate a binary conflict heatmap, where conflict regions are assigned a value of 1 and non-conflict regions are assigned a value of 0.

[0137] The vector data of cultivated land facilities is queried to obtain the main direction of the field ridges (average angle of most broken line segments) and the direction of the vertical line of the canal network (canal orientation angle + 90°). The rule for generating the cutting baseline is as follows: in areas of abrupt curvature change, straight lines are preferentially generated along the direction of the field ridges; in areas of phenological anomalies, broken lines are generated along the vertical line of the canal network. When the angle between the facility direction and the major axis of the patch is greater than 45°, the first axis direction of the principal component analysis of the patch is used as the cutting baseline.

[0138] Connected components are extracted from the conflict heatmap as the starting point for cutting. A depth-first strategy is used for propagation: each time, the propagation moves 10 pixels along the baseline, checking the standard deviation of the phenological feature vectors within a 5×5 region surrounding the new location. If the standard deviation is less than 0.1 (consistent region), the propagation stops; otherwise, it continues. The maximum propagation distance is limited to 80% of the minor axis length of the patch.

[0139] Calculate the area ratio of the new patch after cutting (new patch area / original patch area). Attribute allocation rules are as follows: crop type inherits the original patch attributes, area field is adjusted proportionally, and other attributes (such as confidence) are weighted averages of their original values. Adjacency update method: delete all edges of the original node, and perform Delaunay triangulation to reconstruct adjacencies for the new node. The topology graph version number is incremented, and the cutting operation log is recorded.

[0140] The final topology map is converted to GeoJSON format: nodes are converted to polygon features (retaining all attribute fields), and edges are converted to line features (including connection type and weight fields). Douglas-Pock thinning (0.5-meter tolerance) is performed on the geometry to reduce the number of nodes. The output file includes coordinate system information, data version, and conflict resolution records.

[0141] In the iterative optimization module, three parallel monitoring channels are set up to evaluate the stability of the graph structure. The first channel calculates the relative change ratio of the total number of patches between two adjacent iterations. Specifically, it calculates the difference between the current number of patches and the previous iteration, and then divides the difference by the previous iteration. The second channel performs statistical analysis on the curvature sequence of all patch boundaries. First, the curvature values ​​are divided into ten equal-width intervals, the frequency of occurrence of each interval is counted, and then the degree of distribution disorder is calculated according to the information entropy formula. The third channel constructs a similarity matrix of phenological features for all patches, compares the numerical differences between the corresponding positions of the current matrix and the previous iteration matrix, and calculates the square root of the sum of the squares of all differences to obtain the matrix difference measure. When the monitoring values ​​of the three channels are lower than the preset thresholds (3 / 10,000 of a change in number, 0.1% of an entropy difference, and 0.15% of a matrix difference), the system is considered to have reached convergence.

[0142] The learning rate is dynamically adjusted by comparing the changes in node states before and after topology graph reconstruction. First, the percentage of nodes that underwent merging or splitting operations out of the total number of nodes is used as the dissimilarity metric. The learning rate adjustment rule is as follows: using the initial learning rate as a baseline, for every 10 percentage point increase in dissimilarity, the learning rate is adjusted to 80% of the previous value. For example, when the dissimilarity reaches 20%, the learning rate will undergo two decays to return to 64% of the initial value. To prevent training from stalling due to an excessively low learning rate, a minimum learning rate of 10% of the initial value is set.

[0143] A conflict node marking and priority evaluation mechanism was established. Nodes identified as anomalous fragments or giant patches in three consecutive iterations were marked as geometric conflict nodes; nodes with consistently large differences in phenological characteristics from adjacent patches were marked as phenological conflict nodes; and nodes with high overlap with farmland facility boundaries were marked as facility conflict nodes. Different weight scores were assigned to each type of conflict node: geometric conflict was worth three points, phenological conflict two points, and facility conflict one point. At the beginning of each iteration, high-conflict nodes were prioritized for processing according to their total conflict scores, from highest to lowest.

[0144] A comprehensive version control system is designed to manage intermediate results during the iteration process. Each version record contains four core pieces of information: iteration number, timestamp, key convergence metric values, and a list of nodes that changed during that iteration. The topology graph data uses a compressed storage format, including a node attribute table, a connectivity matrix, and a simplified representation of the conflict marker graph. When the system detects that key metrics have continuously deteriorated for three consecutive iterations, a rollback mechanism is automatically triggered to restore the system state to the archive point where the metrics were optimal in a historical version.

[0145] A tree structure is constructed to record the evolution of map patch attributes. Each tree node represents the map patch state in a specific iteration, and tree edges represent the derivation relationships caused by merging or cutting operations. For each map patch in the final output, its origin is traced back along the tree structure to ensure that the classification attributes are inherited from the initial reliable recognition results. For mixed-type map patches formed by merging multiple map patches, the original type of each source map patch and its area proportion in the merged result are recorded in detail.

[0146] Two-level verification is implemented to ensure output quality. The first level of verification detects the spatial relationship between adjacent patch boundaries, using a precise distance calculation method to identify overlapping areas with a spacing of less than one centimeter. These areas undergo a negative buffering process of 0.2 centimeters to eliminate overlap. The second level of verification analyzes the attribute compatibility of adjacent patches. When two non-mixed type patches have different crop classifications, a special transition zone with a width of 50 centimeters is created at their boundary, marked as a type boundary zone.

[0147] Professional methods were employed to integrate farmland infrastructure data. Field ridge data retained its original linear feature form, with the attribute table recording the measured width and orientation angle of each ridge segment. Canal network data was converted from areal features to centerlines with width, and density grading indicators were added. In the results presentation, the facility layer was set to semi-transparent and overlaid on the classification results, maintaining facility visibility without obscuring the underlying map information.

[0148] The reliability index of each map patch is comprehensively evaluated. The iterative stability index is evaluated by statistically analyzing the number of times the map patch undergoes type or boundary changes in the last three iterations; no change is counted as full marks, and 0.1 points are deducted for each change. The conflict resolution index calculates the improvement degree of the initial conflict intensity and the final state of the map patch, converting it into a score from zero to one on a linear scale. The average of the two indices is taken as the comprehensive confidence score, which is visually presented using a gradient color scheme from red to green.

[0149] A standardized output package is constructed. The main map stores crop classification results in a common vector format, including complete attribute fields. Auxiliary layers provide thematic information such as farmland facilities and confidence distribution in separate files. Metadata documents record in detail the key parameters of the data processing process and the operating environment information. All output files are compressed and packaged, and a format specification document is attached.

[0150] The working principle of this invention is to achieve high-precision patch extraction and dynamic optimization through multi-module collaboration. The system takes topological graph construction as its core, combines a dual-channel graph neural network with topological rule constraints, and innovatively introduces multi-scale feature fusion, dynamic threshold adjustment, and spatial semantic verification mechanisms. The specific technical path includes: integrating a 30-meter resolution digital elevation model with remote sensing imagery, generating a binary mask through slope abrupt change detection, and removing false edges of terraced fields; the structural analysis sub-unit identifies abnormal fragments or giant patches based on boundary curvature information entropy and multi-scale erosion operations; the semantic verification sub-unit fuses temporal vegetation index and elevation variation coefficient, and verifies the phenological consistency of adjacent patches through a dynamic time warping algorithm; 3) Dynamic topological reconstruction: combining irrigation canal vector data and farmland facility semantic coding library, establishing a dynamic penalty for spatial adjacent edge weights (weight decay of cross-canal edges, cutting off phenological conflict edges), determining the convergence state through three-channel stability monitoring (patch quantity change, boundary entropy value, phenological similarity matrix difference), and combining conflict node priority evaluation and learning rate dynamic decay mechanism, finally outputting a high-confidence crop planting distribution map in GeoJSON format.

[0151] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A smart system for extracting crop planting patches based on remote sensing information, characterized in that, include: The topology graph construction module is used to acquire remote sensing image data and generate initial segmented patches, and construct a topology graph structure with patches as nodes and spatial adjacency relationships as edges; The dual-channel graph neural network module, connecting to the topology graph construction module, includes: Structural analysis sub-units identify anomalous fragments or giant spots based on patch shape and boundary curvature; The semantic verification subunit integrates temporal vegetation index and elevation data to verify the phenological consistency of adjacent patches; The topology rule constraint module, connected to the dual-channel graph neural network module, is used to generate the loss signal corresponding to the topology rules of cultivated land and optimize the boundary features of the segmentation model through gradient penalty. The dynamic graph reconstruction module, connected to the topology rule constraint module, performs the following operations: Merge highly similar adjacent nodes; Cut off internal topological conflict nodes; Output the corrected vector graphic pattern; The iterative optimization module connects to the dynamic graph reconstruction module, and iteratively calls the dual-channel graph neural network module to the dynamic graph reconstruction module until the graph structure is stable, generating a crop planting distribution map. The specific implementation steps are as follows: Simultaneously monitor the rate of change of the number of patches, the entropy value of the boundary curvature distribution, and the Frobenius norm of the phenological feature similarity matrix; The graph structure is considered stable when all three changes are below the dynamic threshold. The learning rate is adjusted based on the differences between nodes before and after topology reconstruction. For every 10% increase in variance, the learning rate decays to 80% of its original value. High-conflict nodes in the historical iterations are marked and prioritized for processing by the dual-channel graph neural network module in the next loop. Store the intermediate topology graph structure for each iteration, and roll back to the optimal cached version when convergence fails.

2. The intelligent crop planting patch extraction system based on remote sensing information according to claim 1, characterized in that, The specific steps of the topology graph construction module are as follows: Calculate abrupt slope regions using digital elevation model data to generate elevation gradient masks; Remove the boundaries of plots with slope changes exceeding the threshold from the initial segmentation patch to eliminate false edges caused by terraced fields; Extract the temporal vegetation index curves of adjacent patches and calculate the dynamic temporal regularization distance; When the distance is less than the synchronization threshold, establish a spatial adjacency edge; When the distance exceeds the asynchronous threshold, the original adjacency relationship is severed. Compress the geometric features of the patch into a vector of compactness and aspect ratio; By integrating phenological feature vectors and elevation statistical feature vectors, node semantic descriptors are generated. Calculate the connection weight of adjacent edges based on the shared boundary length and phenological similarity; Apply a penalty weight to edges that cross irrigation canals to suppress incorrect connections across the canals.

3. The intelligent crop planting patch extraction system based on remote sensing information according to claim 1, characterized in that, The specific implementation process of the structural analysis subunit is as follows: Calculate the curvature sequence of boundary points of the patch and quantify the boundary irregularity using information entropy; When the entropy value exceeds the dynamic threshold, it is judged as abnormal fragmentation; Generate variable-size structural elements based on the area of ​​the map patch; Perform multi-scale etching operations on giant spots to separate the adhered areas; Detect the boundaries of nested patches and trigger cutting based on the overlap of bounding boxes and the area of ​​internal voids; Fragmented regions are labeled as negative samples and giant spot regions are labeled as positive samples, generating a binary mask that is fed back to the segmentation network.

4. The intelligent crop planting patch extraction system based on remote sensing information according to claim 1, characterized in that, The specific implementation process of the semantic verification subunit is as follows: Extract the temporal vegetation index within the growth cycle of each patch, and generate a phenological feature vector through one-dimensional convolution. Align the elevation zone data with the phenological vectors according to the grid cells, and calculate the elevation variation coefficient to correct the phenological similarity. Calculate the dynamic time curvature distance for the phenological feature vectors of adjacent patches; When the distance is greater than the asynchronous threshold and the elevation difference is less than the tolerance, it is judged as an incorrect merging across crops; Disconnect the edges of phenological conflict and reduce their weight to 10% of their original value.

5. The intelligent extraction system for crop planting patches based on remote sensing information according to claim 1, characterized in that, The specific implementation process of the topology rule constraint module is as follows: Establish a semantic coding library for field ridges and irrigation ditches to vectorize linear structures in high-resolution images; When the overlap between the segment boundary and the facility vector line exceeds a threshold, it is marked as an erroneous connection to be penalized; The penalty weights are assigned according to the type of incorrect connection: the weight coefficient for misidentification of field ridge boundaries is the reciprocal of the facility width, and the weight coefficient for mismerging of water channels is the square of the canal network density. During the backpropagation phase of the segmentation model, the penalty weights are converted into gradient correction amounts for the boundary feature layers. Apply facility orientation constraints to the gradient matrix of the convolution kernel to suppress convolution responses perpendicular to the direction of the field ridge.

6. The intelligent crop planting patch extraction system based on remote sensing information according to claim 1, characterized in that, The specific steps of the gradient penalty optimization are as follows: Topology penalty is applied deep to the encoder layer of the segmentation network, and boundary detail penalty is applied shallow to the decoder layer. Calculate the distance from the faulty edge to the nearest farmland facility; The gradient penalty intensity is reduced by the inverse of the distance to achieve soft constraints on the facility's influence domain; The loss signal output by the topology rule constraint module is weighted and fused with the original loss of the segmentation network; The fusion weight decreases exponentially with each training round, starting at 0.8 and decaying to half its original value every 50 rounds.

7. The intelligent crop planting patch extraction system based on remote sensing information according to claim 1, characterized in that, The merging of highly similar adjacent nodes specifically includes: A three-dimensional similarity vector is constructed based on the geometric feature similarity of adjacent nodes, the correlation coefficient of phenological curves, and the consistency of elevation. The shared boundary length is used as a weighting factor to calculate the overall similarity score; The merging threshold is dynamically adjusted based on the degree of farmland fragmentation in the region: the threshold decreases by 0.05 for every 10% increase in the fragmentation index; The minimum bounding polygon is used to merge nodes, preserving the original image patch attributes; Add smoothing constraints at the merge boundary to eliminate jagged edges; If irrigation canal vector lines exist between nodes to be merged, the merging operation will be forcibly terminated.

8. The intelligent crop planting patch extraction system based on remote sensing information according to claim 1, characterized in that, The process of cutting off internal topological conflict nodes specifically includes: The abnormal probability map output by the dual-channel graph neural network is used to mark the boundary curvature abrupt change area and the phenological anomaly area; Create a cutting baseline along the direction of the field ridge or the vertical line of the canal network; The cutting angle is corrected based on vector data of farmland facilities to ensure that the path is parallel to agricultural infrastructure. Proceeding from the point of conflict along the cutting baseline, stopping when encountering a region of consistent phenology; Distribute the original node attributes to the new polygons according to the area ratio; Rebuild adjacency relationships and update the topology graph structure.

9. The intelligent crop planting patch extraction system based on remote sensing information according to claim 1, characterized in that, The generated crop planting distribution map specifically includes: By tracing the segmentation trajectory of the patch through the graph structure version chain, it is ensured that the final attributes are inherited from the initial classification result; Check the geometric non-intersection and attribute mutual exclusion of adjacent patches, and repair overlapping and gap defects; Irrigation canals and field ridges are overlaid as independent layers to preserve the topological relationships of agricultural infrastructure. A confidence heatmap is generated based on the number of iterations and the degree of conflict resolution, and then integrated into the distribution map attribute table.

Images