Land area correction method and system based on satellite data

Abstract

The invention relates to the technical field of land area correction, and discloses a land area correction method and system based on satellite data, and the method comprises the following steps: inputting a satellite image, extracting multi-level features through a ResNet-101 backbone network, constructing a feature pyramid, and fusing deep semantics and shallow detail features from top to bottom. And outputting multi-scale fusion features, performing channel dimension reduction on the multi-scale fusion features, aggregating multi-scale context information through an ASPP module, integrating the features, then generating a broken plot probability graph, identifying a pixel-level broken region probability, splicing three paths of features, and generating a spatial adaptive weight graph. Through the EC-Net enhancement process, the problems of small plot misrecognition and boundary adhesion under the ground surface fragmentation scene of the land reclamation area are solved, the recognition precision of the small plot is improved, the edge sawtooth effect and interferent misjudgment phenomena in a traditional method are eliminated, and the recognition accuracy of the small plot is improved. The influence of shadow shielding and ore heap interference is overcome by using adaptive feature fusion and mathematical constraint optimization.

Description

TECHNICAL FIELD

[0001] The application relates to the field of land area correction, in particular to a land area correction method and system based on satellite data. BACKGROUND

[0002] With the popularity of high-resolution satellite images, the traditional land area adopts the manual vectorization and threshold segmentation measurement method.

[0003] When solving the high-precision correction of the land area of the reclamation area, due to the surface fragmentation and the interference of the real land block and the interference of the interference material such as the ore pile, the <0.1hm2 land block is misidentified and adhered, and the boundary positioning deviation is large. SUMMARY

[0004] The application provides a land area correction method and system based on satellite data, which solves the technical problem that in the related art, the surface fragmentation and the interference of the real land block and the interference of the interference material such as the ore pile cause the misidentification and adhesion of the <0.1hm2 land block, and the boundary positioning deviation.

[0005] The application provides a land area correction method based on satellite data, which comprises the following steps: S100, inputting satellite images, extracting multi-level features through a ResNet-101 backbone network, constructing a feature pyramid, fusing deep layer semantic features and shallow layer detail features from top to bottom, and outputting multi-scale fusion features with a unified size; S200, performing channel dimension reduction on the multi-scale fusion features, aggregating multi-scale context information through an ASPP module, generating a broken land block probability map after integrating the features, and identifying the pixel-level broken area probability; S300, aligning edge features, semantic features and extended broken features, splicing the three features and generating a spatial adaptive weight map, and outputting a unified feature through weighted fusion according to the weight; S400, converting the fusion features into a foreground probability map and binarizing, dynamically calculating the structure element radius based on the local window average land block area, performing adaptive closing operation to fill holes and connect fragments, and outputting a smooth boundary vector; S500, discretely sampling the boundary and calculating the curvature feature, searching for adjacent land blocks, determining the merging condition through curvature similarity, performing Bézier curve refitting and area conservation correction on the merged land blocks, and outputting the topologically optimized boundary; S600, identifying a micro land block, searching for adjacent large land blocks of the same type, calculating the correction area based on the distance weight, maintaining the total area conservation through reverse compensation of the large land block, and outputting the final area result with a confidence interval.

[0006] Further, in S100, the specific steps are as follows: S110, backbone network level feature extraction: input satellite image, extract four levels of feature maps through ResNet-101 backbone network, including shallow detail feature, middle detail feature, middle-deep feature and deep semantic feature; S120, feature pyramid construction: in the top-down path, starting from the deep semantic feature, through convolution dimension reduction to 256 channels, step-by-step upsampling and fusion of shallow features, in the bottom-up path, add lateral connection to enhance details; S130, multi-scale feature fusion: sample each layer feature in the feature pyramid to a unified size, concatenate and fuse along the channel dimension, output multi-scale fusion features.

[0007] Further, in S200, the specific steps are as follows: S210, feature dimension reduction and enhancement: input the multi-scale fusion features output by S130, through convolution compression channel number and enhance feature representation; S220, multi-scale context aggregation through spatial pyramid pooling module: use four parallel hollow convolution branches and global average pooling branch, upsample the outputs of all branches to the same spatial size as the input, then concatenate by channel; S230, feature integration: through convolution integrates the features output by the spatial pyramid pooling module; S240, broken probability prediction: generate pixel-level broken patch probability map.

[0008] Further, the calculation formula of the pixel-level broken patch probability map in S240 is as follows: ; Where, is a 1x1 point convolution, is a Sigmid activation function, is the integrated feature, is the broken patch probability map; ; Where, is a batch normalization layer, is an activation function, is the feature output by the spatial pyramid pooling module, is the integrated feature; ; ; ; ; ; ; wherein, is a 3x3 dilated convolution with dilation rate k, is a global average pooling operation, is a bilinear interpolation up-sampling to , is a dilated convolution output feature map with dilation rate k, is a global context feature, wherein is a dilation rate, .

[0009] Further, in S300, the specific steps are as follows: S310, input feature alignment: obtaining three input features, including edge features, semantic features and broken features, and aligning the three input features; S320, adaptive weight generation: splicing the aligned three input features, generating spatial adaptive weights through two convolution layers, and outputting a three-channel weight map; S330, three-flow feature weighted fusion: element-wise weighting of three features according to the weight map.

[0010] Further, in S400, the specific steps are as follows: S410, foreground probability map generation: inputting the fusion feature of S330, and generating a foreground probability map through a 1x1 convolution; S420, dynamic structure element calculation: binaryzation processing of the foreground probability map, traversal of the image with a sliding window, calculation of the average patch area and structure element radius in the window; S430, adaptive closing operation: performing adaptive closing operation, including dilation operation and erosion operation; S440, morphological optimization: filling holes and removing isolated points by morphological method; S450, boundary smoothing output: converting the optimized binary map into a vector boundary, and finally outputting a smooth boundary.

[0011] Further, the calculation formula of the adaptive closing operation is as follows: Dilation operation: ; Erosion operation: ; wherein, is a circular structure element with radius , is a dilation operator, is an erosion operator, for position-adaptive structuring element radius, for binary image after dilation operation, for binary image after closing operation; morphological formula: ; where the calculation formula of morphological optimization is as follows: hole filling: ; isolated point removal: ; wherein, is a morphological reconstruction operation, is a 3x3 square structuring element, is a complement operator, is a seed image of the reconstruction operation, is a binary image after hole filling, is a binary image after cleaning.

[0012] Further, in S500, the specific steps are as follows: S510, boundary discrete point sampling: input the smoothed boundary vector of S450, sample discrete points along the boundary at equal intervals, and output a discrete point set; S520, curvature calculation and characterization: calculate the local curvature using the three-point method, encode the curvature feature, and generate a curvature feature vector; wherein the calculation formula of the three-point method is as follows: ; wherein, is the local curvature of the ith point, is the vector cross product, is the vector norm, is the coordinates of the three consecutive sampling points; ; ; wherein, is the sampling interval, is the length of the line segment, is the two-dimensional coordinates of the ith sampling point, is the coordinates of the kth boundary vertex, is the Euclidean distance, is the number of sampling points; S530, adjacent land parcel matching: construct a KD-Tree spatial index of boundary points, search for adjacent point pairs and filter same-land parcel point pairs; S540, curvature similarity measure: obtain the merging condition by calculating the curvature feature similarity of the point pairs; S550, topological constraint merging: merging adjacent plots, re-fitting by Bézier curve, and optimizing control points; S560, area conservation correction: calculating the area change rate before and after merging, when the area change rate before and after merging is greater than 0.05, performing area correction; S570, output merging boundary: generating the vector boundary after merging.

[0013] Further, in S600, the specific steps are as follows: S610, micro plot identification: input the merged vector boundary of S570, calculate the area of each plot and screen the micro plot; S620, search for adjacent same type plots: calculate the plot center, and perform same type matching and distance weight calculation; S630, weighted area correction: calculating the weighted average, adjusting the boundary constraint; S640, area conservation compensation: calculating the total area change, adjusting the area of large plots; S650, uncertainty quantization output: outputting the final area with uncertainty according to the area confidence interval.

[0014] The application also proposes a land area correction system based on satellite data for executing the steps in the aforementioned land area correction method based on satellite data, comprising: The feature extraction module receives the preprocessed satellite image, extracts multi-level features through the ResNet-101 backbone network, constructs a feature pyramid, fuses different scale features from top to bottom, and generates multi-scale fusion features; The probability prediction module performs dimensionality reduction processing on the multi-scale features, captures multi-scale context information using a hollow spatial pyramid pooling, integrates the features through a convolution layer, and outputs a pixel-level broken plot probability map to identify the micro plot area; The feature three-flow fusion module aligns the three inputs of edge detection flow, semantic segmentation flow and broken feature flow, splices the features and generates a spatial adaptive weight map, dynamically fuses the three features according to the weight, and outputs a unified enhanced feature; The adaptive morphological module converts the fused features into a foreground probability map and binarizes it, dynamically calculates the structure element size according to the average area of the plot in the local window, performs adaptive closing operation, fills holes and connects fragments, and outputs a smooth boundary vector; The plot merging module discretely samples the boundary and calculates the curvature feature, searches for adjacent plot boundary points, determines the merging condition through curvature similarity, re-fits the Bézier curve for plots meeting the condition, optimizes the topological connection and corrects the area conservation; Area correction module: identify small plots and search for adjacent large plots of the same type, calculate the correction area based on distance weight, maintain total area conservation through large plot reverse compensation, generate the final area report with confidence interval.

[0015] The beneficial effects of the present application are: The present application solves the problems of small plot misidentification and boundary adhesion in the land reclamation area surface fragmentation scene through EC-Net enhanced process, completely eliminates the edge sawtooth effect and interference error phenomenon in the traditional method, realizes sub-pixel level boundary positioning accuracy by using adaptive feature fusion and mathematical constraint optimization, and effectively overcomes the influence of shadow shielding and ore pile interference. The whole process of the present application greatly improves the processing efficiency while ensuring the accuracy of area calculation, provides an automatic solution with high reliability and strong anti-interference for land reclamation supervision, and significantly improves the scientific nature and decision-making reliability of land management. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of a land area correction method based on satellite data of the present application; Figure 2 is a structural block diagram of a land area correction system based on satellite data of the present application.

[0017] In the figure: 101, feature extraction module; 102, probability prediction module; 103, feature three-flow fusion module; 104, adaptive morphological module; 105, plot merging module; 106, area correction module. DETAILED DESCRIPTION

[0018] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the present description. Each of the various examples can omit, substitute or add various procedures or components as appropriate or desired. In addition, features described in relation to some examples can also be combined in other examples.

[0019] As shown in Figure 1 A land area correction method based on satellite data, comprising the following steps: S100, multi-scale feature pyramid extraction: input satellite image, extract multi-level features through ResNet-101 backbone network, construct feature pyramid, fuse deep semantic and shallow detail features from top to bottom, and output multi-scale fusion features with uniform size; In one embodiment of the present application, the following steps are specifically included: S110, backbone network level feature extraction: input pre-processed satellite image wherein, is the image height, is the image width, is the number of multi-spectral bands; Four levels of feature maps are extracted by ResNet-101 backbone network: , , and ; wherein, is the shallow layer detail feature, , is the middle layer detail feature, , is the middle-deep layer feature, , is the deep layer semantic feature, ; Formula: ; wherein, is the 101-layer residual network, is the input image; S120, feature pyramid (FPN) construction: Top-down path: Starting from the deep layer semantic feature , dimension reduction is performed to 256 channels by convolution: ; Step-by-step upsampling and fusion of shallow layer features: ; ; ; Bottom-up path: Add transverse connection to enhance details: ; ; ; ; wherein, is the bilinear interpolation upsampling, is the convolution kernel, is the layer feature map, i.e. ​This is the second layer feature map. This is the feature map of layer 3. This is the feature map of layer 4. This is the feature map of layer 5. For the first The enhanced feature map after layer enhancement, i.e. This is the feature map after the second layer enhancement. This is the enhanced feature map of layer 3. This is the enhanced feature map of layer 4. This is the enhanced feature map of layer 5; S130, Multi-scale Feature Fusion: Integrating features from each layer of the Feature Pyramid Network (FPN) Upsampling to uniform size ; Blending along the channel dimension: ; in, For multi-scale fusion features, , For feature splicing operations, For upsampling operation, The magnification factor is 1, 2, 4, or 8. S200, Fragmented Plot Probability Prediction: Channel dimensionality reduction is performed on multi-scale fused features, multi-scale context information is aggregated through the ASPP module, and after feature integration, a fragmented plot probability map is generated, identifying the probability of pixel-level fragmented areas. In one embodiment of the present invention, the following steps are specifically included: S210, Feature Dimensionality Reduction and Enhancement: Multi-scale fusion features input to S130 and output from S130 ; pass Convolution compresses the number of channels and enhances feature representation: ; in, It uses a 3×3 convolution kernel and has 512 output channels (stride=1, padding=1). This is a batch normalization layer used to standardize the feature distribution. The activation function for the rectified linear unit is... , The input is a multi-scale fusion feature. The feature map after dimensionality reduction. ; S220 performs multi-scale context aggregation through the ASPP (Spatial Pyramid Pooling) module: Parallel dilated convolution: utilizing the dilation rate of Dilated convolution and global average pooling branch: ; ; ; ; ; Feature concatenation: ; wherein, is a 3x3 dilated convolution with dilation rate k, i.e. is a 3x3 dilated convolution with dilation rate 1, is a 3x3 dilated convolution with dilation rate 6, is a 3x3 dilated convolution with dilation rate 12, is a 3x3 dilated convolution with dilation rate 18, is a global average pooling operation, is bilinear interpolation up-sampling to , is a convolution output feature map with dilation rate k, i.e. is a convolution output feature map with dilation rate 1, is a convolution output feature map with dilation rate 6, is a convolution output feature map with dilation rate 12, is a convolution output feature map with dilation rate 18, is a dilation rate, , is a global context feature, and each convolution branch outputs 128 channels, is a feature output by the ASPP module, ; S230, feature integration: integrating the features output by the ASPP module by convolution: ; wherein, is a 1x1 point convolution, outputting 256 channels, is a batch normalization layer, is an activation function, is a feature output by the ASPP module, is the integrated feature, ; S240, crack probability prediction: generating a pixel-level crack patch probability map; ; wherein, 1x1 point convolution, output channel number 1 (no activation function), Sigmid activation function, , Integrated features, Fragmented block probability map, (value range [0, 1]); S300, edge-semantic-fragment feature three-flow fusion: align the edge features, semantic features and extended fragment features, splice three features and generate a spatial adaptive weight map, and output unified features by weighted fusion according to the weight; In an embodiment of the present application, the following steps are specifically included: S310, input feature alignment: Obtain three input features: Edge features (from the edge detection branch of the general process, size ); Semantic features (from the semantic segmentation branch of the general process, size ); Fragmented features (extend the size of in S240 to 256 channels through 1x1 convolution); ; Wherein is a convolution expansion operation; Feature alignment: ; Wherein, is an input feature, which can be , or , 1x1 point convolution operation, rectified linear unit activation function, feature alignment function; Ensure that the spatial size of the three features is consistent ; S320, adaptive weight generation: Splice features: ; Wherein, is a feature splicing operation, is a spliced feature map, , is the number of channels after splicing; Weight prediction: Generate spatial adaptive weights by two layers of convolution: ; wherein is a weight map, ; Output a three-channel weight map: ; wherein, is an Sftmax function (channel-wise normalization), is a 3x3 convolution, output channel number 512, is a 1x1 convolution, output channel number 3, is a spatial position coordinate, is the weight map of the c-th channel, is a channel index; S330, three-stream feature weighted fusion: Element-wise weighting of three features according to the weight map: ; wherein, is a Hadamard product (element-wise multiplication), is the c-th channel weight map, is the c-th feature map, defined as follows: ; Output: ; wherein, is the spatial size of the feature map, is the output feature channel number, is the final fused feature, which integrates edge, semantic and broken information; S400, adaptive morphological closing operation: convert the fused feature into a foreground probability map and binarize it, dynamically calculate the structure element radius based on the local window average patch area, perform adaptive closing operation to fill holes and connect fragments, and output a smooth boundary vector; In an embodiment of the present application, the following steps are specifically included: S410, foreground probability map generation: Input the fused feature of S330 Generate a foreground probability map through a 1x1 convolution: ; wherein, is a Sigmid activation function that maps the output to the [0,1] interval, is a 1x1 convolution operation with an output channel number of 1, is the input fused feature, is the foreground probability map, which represents the probability of each pixel belonging to the foreground, ; S420, dynamic structuring element calculation: Binaryzation process: ; where is the binaryzated image, is the binaryzation threshold; Local window analysis: traverse the image with a sliding window (size , e.g. 200x200 pixels); Calculate the average patch area within the window: ; where is the number of connected components within the window, is the area of the i-th connected component, is the size of the sliding window, is the average patch area within the window; Structuring element radius calculation: ; where is the minimum patch threshold, , is an empirical coefficient, is a zero-preventing constant, is the window center coordinate, is the position-adaptive structuring element radius; S430, adaptive closing operation: Dilation operation: ; Erosion operation: ; where is the circular structuring element with radius , is the dilation operator, is the erosion operator, is the position-adaptive structuring element radius, is the binary image after dilation operation, is the binary image after closing operation; Morphological formula: ; ; where is the circular structuring element, is the current processed pixel position, to shift the relative position within the structure element; S440, morphological optimization: filling holes: ; removing isolated points: ; wherein, is a morphological reconstruction operation for filling internal holes while keeping the shape of the boundary unchanged, is a 3x3 square structure element, is a complement operator, is a seed image for the reconstruction operation, is the binary image after filling holes, is the cleaned binary image; S450, boundary smoothing output: convert the optimized binary image to a vector boundary: ; output the smoothed boundary: ; ; wherein, is a binary image to vector operation, is the coordinate of the i-th boundary vertex, wherein and represent the horizontal and vertical coordinates, respectively, is the total number of boundary vertices, is an intermediate vectorization result, storing temporary data structures of the boundary contour, is the final smoothed boundary point set, storing all boundary vertices in order, is the vertex index; S500, curvature-constrained land merging: discretely sampling the boundary and calculating the curvature feature, searching for adjacent lands, determining the merging condition through curvature similarity, performing Bézier curve refitting and area conservation correction on the merged lands, and outputting the topologically optimized boundary; In one embodiment of the present application, the following steps are specifically included: S510, boundary discrete point sampling: input the smoothed boundary vector of S450: ; wherein, is the coordinate of the i-th boundary vertex, is the total number of boundary vertices; sample discrete points along the boundary at equal intervals: ; ; where, is the sampling interval, m, is the segment length, is the 2D coordinate of the i-th sampling point, is the number of sampling points; Output discrete point set: ; where, is the total number of sampling points, is the discrete point set, is the 2D coordinate of the m-th sampling point; S520, curvature computation and characterization: Local curvature computation (3-point method): ; where, is the local curvature of the i-th point, is the vector cross product, is the vector norm, is the coordinates of the consecutive three sampling points; Curvature feature encoding: ; where, is the 5D curvature feature vector, is the curvature value of the j-th point; S530, neighboring patch matching: Construct boundary point KD-Tree spatial index; Search neighboring point pairs: ; where, m is the search radius, m, is the neighboring point set of the i-th point, is the Euclidean distance, is the coordinates of the two boundary points; Filter same-patch point pairs (by patch ID label); S540, curvature similarity measure: Compute point pair curvature feature similarity: ; where, is the similarity score, is the Gaussian kernel bandwidth parameter, is the 5D curvature feature of the i-th point, is the 5D curvature feature of the j-th point, is the square of the Euclidean distance; Merge condition: ; wherein, is , is the area upper threshold; S550, topological constraint merging: Merge adjacent patches: ; wherein, is the boundary of the two patches to be merged, is the boundary after merging; Bézier curve refitting: ; wherein, is the parameter, , is the number of control points, is the control point coordinates, is the binomial coefficient, is the point coordinate on the parametric curve; Control point optimization: ; wherein, is the second derivative of the curve, is the regularization coefficient, is the discontinuity penalty term, is the optimization of the control point coordinates, is the parameter integral, calculating the properties of the entire curve; S560, area conservation correction: Calculate the area change rate before and after merging: ; wherein, is the patch area after merging, respectively represent the original area of the two patches to be merged, is the relative area change rate; If , area correction is performed: ; wherein, is the patch center distance, is the distance decay coefficient, m, is the corrected area, is the distance decay factor, is the corrected patch area; S570, output merging boundary: Generate the merged vector boundary: ; wherein, is the i-th Bézier curve, is the number of boundary segments; S600, area-weighted correction: identify small patches, search for adjacent large patches of the same type, calculate the correction area based on distance weight, maintain total area conservation by large patch reverse compensation, output the final area result with confidence interval; In an embodiment of the present application, the following steps are specifically included: S610, small patch identification: input the merged vector boundary of S570 ; Calculate the area of each patch: ; wherein, is the area of the t-th patch, is the coordinate of the i-th vertex, is the coordinate of the i+1-th vertex, is the total number of patch vertices; Screen small patches: ; wherein, is the index set of small patches, is the area threshold, is the set symbol, is the patch index; S620, search for adjacent patches of the same type: Patch center calculation: ; wherein, is the center coordinate of the t-th patch, is the patch vertex coordinate, is the total number of vertices; Same type matching: Based on semantic features (S300) extract patch category label , search for large patches of the same category ( ); Distance weight calculation: ; wherein, is the weight of the l-th adjacent patch, is the center coordinate of the original area of the two patches to be merged, is the distance attenuation coefficient, Euclidean distance; S630, Weighted Area Correction: Weighted Average Calculation: ; where, is the corrected area estimate of the tth parcel, is the set of neighboring parcels of the tth parcel, is the weight of the lth neighboring parcel, is the area of the neighboring parcel; Boundary Restriction Adjustment: ; where, is the final corrected area, is the correction range limit; S640, Area Conservation Compensation: Total Change Calculation: ; where, is the total change of the micro-parcel area, is the set of micro-parcels, is the area of the tth parcel; Large Parcel Area Adjustment: ; where, is the adjusted large parcel area, is the original large parcel area, is the total change of the micro-parcel area, is the micro-parcel area; S650, Uncertainty Quantification Output: Area Confidence Interval: ; where, is the area uncertainty, is the number of neighboring parcels, is the average area of the neighboring parcels, is the coefficient of 95% confidence level, is the operator not belonging to; Final Output: ; where, is the final area with uncertainty, is the corrected area, is the area uncertainty; As Figure 2As shown, in one embodiment of the present application, a land area correction system based on satellite data is also proposed, which includes the following modules: Feature extraction module 101: receives pre-processed satellite images, extracts multi-level features (shallow details + deep semantics) through ResNet-101 backbone network, constructs feature pyramid (FPN), and fuses features of different scales from top to bottom to generate multi-scale fusion features of uniform size; Probability prediction module 102: performs dimensionality reduction processing on multi-scale features, captures multi-scale context information using Atrous Spatial Pyramid Pooling (ASPP), integrates features through convolution layers, and outputs pixel-level broken patch probability map to identify small patch regions; Feature three-flow fusion module 103: aligns the three inputs of edge detection flow, semantic segmentation flow and broken feature flow, splices features and generates a spatial adaptive weight map, dynamically fuses three features according to the weight, and outputs a unified enhanced feature; Adaptive morphological module 104: converts the fused features into a foreground probability map and binarizes it. According to the average area of the patch in the local window, the structural element size is dynamically calculated, and the adaptive closing operation (dilation + erosion) is performed to fill the holes and connect the fragments, and the output is a smooth boundary vector; Patch merging module 105: discretely samples the boundary and calculates the curvature feature, searches for adjacent patch boundary points, determines the merging condition through curvature similarity, and optimizes the topological connection and corrects the area conservation by re-fitting the Bézier curve for the patches that meet the conditions; Area correction module 106: identifies small patches and searches for adjacent large patches of the same type, calculates the correction area based on distance weight, maintains total area conservation through large patch reverse compensation, generates the final area report with confidence interval, and quantifies uncertainty.

[0020] Based on the above method and system, the following application example is proposed. In the North China mining area, land reclamation projects are completed, and the cultivated land area after reclamation needs to be accurately calculated; There are three challenges in the region: Surface fragmentation - a large number of gravel piles, irrigation ditches divide farmland (patch <0.1 hectares); Fuzzy boundary - the boundary between the reclamation area and the unrepaired mine pit is distorted; Mixed land types - new reclaimed farmland and residual ore piles have similar spectral characteristics; EC-Net enhanced process application: The input data includes high-resolution satellite images (0.5 meters), pre-reclamation topographic elevation maps, and land type label databases (cultivated land / ore / bare soil); The processing process is as follows: Feature pyramid extraction: the image is automatically identified by deep network to identify the micro-terrain features such as ridges and gravel piles; Fragmented land perception: 37 gravel pile interference areas and 15 irrigation channel segmented farmland fragments are marked out; Three-flow feature fusion: the broken feature weight is strengthened in the ore pile area, and the edge feature is strengthened in the farmland area; Adaptive closing operation: increase the structure element in the dense gravel pile area and fill the gap between the ridges; small structure element is used for fine smoothing of the mine area boundary; Curvature constraint merging: three broken farmlands (divided by ditches) with continuous curvature are merged into a complete land; Area weighted correction: the area of the small gravel pile is corrected by referring to the adjacent ore pile; the isolated farmland patch is calibrated by referring to the adjacent large land; Output results: Vector boundary map: generate continuous farmland boundaries, automatically separate ore piles and farmland; Area report: Core farmland area: 126.5 hectares (12% higher integrity than traditional methods); Ore area: 38.2 hectares (60% reduction in misjudgment rate); Engineering application: the pass rate of reclamation acceptance is improved; Through the system and correction method in the application, the two pain points of "broken land block missing" and "ore farmland confusion" in the reclamation area are successfully solved, and the "map-number-real land" full-chain accurate matching of the mine reclamation land is realized for the first time, which provides accurate data support for subsequent ecological restoration of the mine area.

[0021] The embodiments of the application are described above, but the application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative and not limiting. Those skilled in the art can make many forms under the inspiration of the application, which all belong to the protection of the application.

Claims

1. A land area correction method based on satellite data, characterized in that, The method comprises the following steps: S100, inputting a satellite image, extracting multi-level features through a ResNet-101 backbone network, constructing a feature pyramid, fusing deep layer semantic features and shallow layer detail features from top to bottom, and outputting multi-scale fusion features with a unified size; S200, performing channel dimension reduction on the multi-scale fusion features, aggregating multi-scale context information through an ASPP module, and generating a broken land block probability map after integrating the features to identify the probability of a pixel-level broken area; S300, aligning edge features, semantic features and extended broken features, splicing the three features and generating a spatial adaptive weight map, and outputting a unified feature by weighted fusion according to the weight; S400, converting the fusion features into a foreground probability map and binarizing the foreground probability map, dynamically calculating a structure element radius based on a local window average land block area, performing an adaptive closing operation to fill holes and connect fragments, and outputting a smooth boundary vector; S500, discretely sampling the boundary and calculating curvature features, searching for adjacent land blocks, determining a merging condition through curvature similarity, performing Bézier curve refitting and area conservation correction on the merged land blocks, and outputting a topologically optimized boundary; S600, identifying a small land block, searching for adjacent large land blocks of the same type, calculating a correction area based on a distance weight, and maintaining total area conservation through reverse compensation of the large land blocks, and outputting a final area result with a confidence interval.

2. The method for land area revision based on satellite data according to claim 1, wherein, In S100, the specific steps are as follows: S110, backbone network level feature extraction: inputting a satellite image, extracting four levels of feature maps through a ResNet-101 backbone network, including shallow layer detail features, middle layer detail features, middle-deep layer features and deep layer semantic features; S120, feature pyramid construction: in the top-down path, starting from the deep semantic feature, through convolution dimension reduction to 256 channels, step-by-step up-sampling and fusion of shallow features, in the bottom-up path, add lateral connection to enhance details; S130, multi-scale feature fusion: sampling each layer feature in the feature pyramid to a unified size, splicing and fusing along the channel dimension, and outputting multi-scale fusion features.

3. The method for land area revision based on satellite data according to claim 2, wherein, In S200, the specific steps are as follows: S210, feature dimension reduction and enhancement: input the multi-scale fusion features output by S130, pass through convolution to compress the number of channels and enhance the feature representation; S220, multi-scale context aggregation through a spatial pyramid pooling module: using four parallel hollow convolution branches and a global average pooling branch, upsampling the outputs of all branches to the same spatial size as the input, and then splicing by channel; S230, feature integration: integrate the features output by the convolutional integration spatial pyramid pooling module; the features output by the convolutional integration spatial pyramid pooling module; S240, broken probability prediction: generating a pixel-level broken land block probability map.

4. The method for land area revision based on satellite data according to claim 3, wherein, The calculation formula of the pixel-level broken land block probability map in S240 is as follows: ； wherein, is a 1 x 1 point convolution, is a Sigmid activation function, is the integrated feature, is a probability map of fragmented patches; ； wherein, is a batch normalization layer, is an activation function, is a feature output by the spatial pyramid pooling module, is the integrated feature; ； ； ； ； ； ； wherein, is a 3x3 dilated convolution with dilation rate k, is a global average pooling operation, is a bilinear interpolation up-sampling to , is a convolution output feature map with dilation rate k, is a global context feature, wherein is a dilation rate, .

5. The method for land area revision based on satellite data according to claim 4, wherein, In S300, the specific steps are as follows: S310, input feature alignment: obtaining three input features, including edge features, semantic features and broken features, and aligning the three input features; S320, adaptive weight generation: splicing the aligned three input features, generating spatial adaptive weights through two convolution layers, and outputting a three-channel weight map; S330, three-flow feature weighted fusion: element-wise weighting of the three features according to the weight map.

6. The method for land area revision based on satellite data according to claim 5, wherein, In S400, the specific steps are as follows: S410, foreground probability map generation: inputting the fusion features of S330, generating a foreground probability map through a 1x1 convolution; S420, dynamic structure element calculation: binarizing the foreground probability map, traversing the image with a sliding window, calculating the average land block area and the structure element radius in the window; S430, adaptive closing operation: performing adaptive closing operation including dilation operation and erosion operation; S440, morphological optimization: filling holes and removing isolated points by using morphological method; S450, boundary smoothing output: converting the optimized binary image into a vector boundary, and finally outputting a smooth boundary.

7. The method for land area revision based on satellite data according to claim 6, wherein, The calculation formula of the adaptive closing operation is as follows: Dilation operation: ； Erosion operation: ； in, For radius Circular structural elements, For expansion operator, For erosion operators, The radius of the position-adaptive structural element. This is the binary image after the dilation operation. The binary image after the closing operation; Morphological formula: ； The calculation formula of the morphological optimization is as follows: Hole filling: ； Isolated point removal: ； wherein, is a morphological reconstruction operation, is a 3x3 square structuring element, is a complement operator, is a seed image for the reconstruction operation, is the binary image after hole filling, is the binary image after cleaning.

8. The method for land area revision based on satellite data according to claim 7, wherein, In S500, the specific steps are as follows: S510, boundary discrete point sampling: inputting the smooth boundary vector of S450, sampling discrete points along the boundary at equal intervals, and outputting a discrete point set; S520, curvature calculation and characterization: calculating local curvature by using three-point method, encoding curvature features, and generating a curvature feature vector; The calculation formula of the three-point method is as follows: ； wherein is the local curvature for the ith point, is the vector cross product, is the vector norm, are the coordinates of the three consecutive sampling points; ； ； wherein, is a sampling pitch, is a line segment length, is a two-dimensional coordinate of the i-th sampling point, is a k-th boundary vertex coordinate, is an Euclidean distance, is a number of sampling points; S530, adjacent land block matching: constructing a KD-Tree spatial index of boundary points, searching for adjacent point pairs and filtering same land block point pairs; S540, curvature similarity measurement: obtaining a merging condition by calculating the similarity of point pair curvature features; S550, topological constraint merging: merging adjacent land blocks, re-fitting by Bézier curve, and optimizing control points; S560, area conservation correction: calculating the area change rate before and after merging, and when the area change rate before and after merging is greater than 0.05, performing area correction; S570, output merged boundary: generating a merged vector boundary.

9. The method for land area revision based on satellite data according to claim 8, wherein, In S600, the specific steps are as follows: S610, micro land block identification: inputting the merged vector boundary of S570, calculating the area of each land block and screening micro land blocks; S620, adjacent same type land block search: calculating land block centers, and performing same type matching and distance weight calculation; S630, weighted area correction: calculating weighted average, and adjusting boundary constraints; S640, area conservation compensation: calculating the total area change, and adjusting the area of large land blocks; S650, uncertainty quantization output: outputting the final area with uncertainty according to the area confidence interval.

10. A satellite data based land area revision system, characterized by, A method for performing the steps in a satellite data-based land area correction method as claimed in any one of claims 1-9, comprising: a feature extraction module: receiving a pre-processed satellite image, extracting multi-level features through a ResNet-101 backbone network, constructing a feature pyramid, fusing different scale features from top to bottom, and generating multi-scale fusion features; a probability prediction module: performing dimensionality reduction processing on the multi-scale features, capturing multi-scale context information by using a dilated spatial pyramid pooling, integrating features by using a convolutional layer, and outputting a pixel-level broken land block probability map to identify a micro land block region; a feature three-flow fusion module: aligning three inputs of edge detection flow, semantic segmentation flow and broken feature flow, splicing features and generating a spatial adaptive weight map, dynamically fusing three features according to the weight, and outputting unified enhanced features; an adaptive morphological module: converting the fused features into a foreground probability map and binarizing, dynamically calculating a structure element size according to an average land block area in a local window, performing adaptive closing operation, filling holes and connecting fragments, and outputting a smooth boundary vector; Patch merging module: Discretely sampling the boundary and calculating the curvature feature, searching for adjacent boundary points, determining the merging condition through curvature similarity, re-fitting the Bézier curve for the patches meeting the condition, optimizing the topological connection and correcting the area conservation; Area correction module: Identifying small patches and searching for adjacent large patches of the same type, calculating the correction area based on distance weight, maintaining the total area conservation through reverse compensation of large patches, and generating the final area report with confidence interval.

Images