Natural resource right confirmation method and system based on multi-source heterogeneous data

Abstract

The present application relates to the technical field of natural resource right confirmation, and particularly relates to a natural resource right confirmation method and system based on multi-source heterogeneous data. The method comprises the following steps: acquiring time-series remote sensing images, digital elevation model data and right ownership map polygon vector data, and stacking the rasterized and encoded data with the images and elevation data to construct a multi-source feature tensor; inputting a multi-task deep learning model to output a pixel-by-pixel change probability map and a land class semantic map in parallel, and enabling a dropout layer in the inference stage to perform multiple forward propagations to predict a variance and generate an uncertainty heat map; taking the boundaries of the stock map polygons as initial contours, and using an active contour model to extract updated boundaries by adjusting an energy function of the heat map; calculating a comprehensive confidence index by fusing average uncertainty values and land class transfer logic conflict degrees, and determining the index according to a threshold value as confirmation, manual review or field verification. The scheme of the present application improves the boundary geometric precision, business logic consistency and result disposal reliability.

Description

Technical Field

[0001] This invention relates to the field of natural resource ownership confirmation technology, specifically to a method and system for natural resource ownership confirmation based on multi-source heterogeneous data. Background Technology

[0002] In the process of natural resource ownership registration, ownership patches are used to represent the spatial boundaries and attribute information of natural resources. The accuracy of their boundaries, the consistency of their attributes, and the timeliness of their updates directly affect the accuracy of subsequent patch verification, change identification, and boundary update results. With the continuous advancement of urban construction, ecological restoration, agricultural management methods, and natural succession, land cover types, spatial boundary morphology, and ownership utilization status may all undergo dynamic changes. Existing ownership patches need to be verified and updated in a timely manner based on the actual land surface conditions. Current update operations typically utilize multi-period remote sensing imagery, digital elevation models, and historical vector data to identify, interpret, and correct boundaries of areas of land surface change. However, faced with natural resource spatiotemporal data that covers a wide area, has multiple data sources, and exhibits significant spatial scale differences, traditional operational methods still rely heavily on manual visual interpretation, indoor comparison, and field verification. This not only consumes considerable time and manpower but also makes it difficult to fully unify the interpretation scales and boundary correction standards among different operators, hindering the adaptation to the needs of routine, high-frequency, and refined ownership registration and update.

[0003] In recent years, technologies such as machine vision, remote sensing intelligent interpretation, and deep learning have been gradually introduced into map patch update operations, providing automated support for change detection and land use identification. However, in natural resource ownership confirmation scenarios, factors such as topographic relief, landform differences, overlapping land use types, occlusion, and complex historical ownership boundaries still affect the model extraction results. Existing methods often focus on image feature recognition, making it difficult to fully integrate elevation information, temporal change characteristics, and the prior boundary morphology of existing ownership patches. This leads to automatically extracted update boundaries being prone to local shifts, distortions, or pseudo-change expansions in complex areas. Furthermore, existing map patch update strategies typically lack quantitative assessments of the reliability of model prediction results, making it difficult to determine which update results can be directly confirmed, which require manual review, and which require field verification. It also struggles to combine land cover type transfer patterns to conduct business logic checks on abnormal changes. Therefore, the workload of internal review and field surveys remains significant, hindering the automation and refinement of the natural resource ownership confirmation map patch update process.

[0004] Chinese patent document CN113255832B discloses a method for identifying long-tailed distributions using a dual-branch, multi-center approach. This method employs a deep convolutional neural network as the backbone for feature extraction, with two functionally distinct branches at the network's backend. These branches are responsible for feature representation learning and classification decision optimization, respectively. The multi-center category representation approach mitigates the identification bias caused by the severely uneven distribution of training samples across different categories, thereby improving classification accuracy and generalization ability when the number of categories is sparse. While this method improves discrimination performance under class imbalance conditions through its dual-branch network structure, it processes only single-form image samples, outputting discrimination results limited to the category level. It lacks the ability to uniformly organize and jointly express multi-source, multi-modal observation information, and it struggles to provide continuous discrimination results corresponding to spatial locations and their spatial distribution of reliability. Therefore, its applicability is limited in scenarios involving the identification of changes in surface elements and the generation of spatial results that can be directly applied to operational tasks.

[0005] Chinese patent application CN121598799A discloses a method and system for predicting the on-orbit behavior of multiple targets in space based on large-scale model reasoning. It embeds prior rule constraints into the reasoning process, outputs prediction results in the form of probability distributions for the behavior categories of multiple observed targets, and adds confidence evaluations to the prediction conclusions to improve the interpretability and reliability of multi-target behavior judgment. However, this application focuses on behavior inference of target-level temporal state sequences. Its confidence evaluation remains at the overall target level, and the rule constraints are mainly used to limit the reasonable value range of the reasoning process. It lacks the ability to express confidence levels per location in image space, making it difficult to feed the confidence level back to the extraction and adjustment of surface element boundaries. Its applicability is still limited in situations where geometric contour accuracy, attribute transformation rationality, and the results need to be graded accordingly. Summary of the Invention

[0006] To address the problems in existing technologies, such as excessive reliance on manual interpretation for map patch updates, insufficient stability of boundary extraction, lack of credibility assessment for model prediction results, and large workload for review and field verification, this invention proposes a natural resource ownership confirmation method and system based on multi-source heterogeneous data.

[0007] Firstly, this invention proposes a method for confirming natural resource ownership based on multi-source heterogeneous data, comprising the following steps: acquiring time-series remote sensing images, digital elevation model data, and existing ownership map vector data of the area to be updated; rasterizing the vector data into ownership and land type identification layers; performing category encoding on the ownership and land type identification layers; stacking the layers with the images and elevation data to construct a multi-source feature tensor; inputting the multi-source feature tensor into a multi-task deep learning model to output a pixel-by-pixel change probability map and a pixel-by-pixel current land type semantic map in parallel; and, during the model inference stage, enabling a dropout layer to perform multiple forward propagation calculations to generate and normalize the prediction variance to obtain a pixel-by-pixel uncertainty heatmap; and based on the pixel-by-pixel change probability map, using the boundaries of existing ownership maps... An active contour model is used to extract the initial update boundary for the initial contour. The energy function of the active contour model is controlled by the pixel-by-pixel uncertainty heatmap and includes an external energy term and a priori shape-fidelity energy term. The weights of the external energy term and the priori shape-fidelity energy term are determined based on local uncertainty. A comprehensive credibility index is calculated for the patches of the initial update boundary. The comprehensive credibility index integrates the average uncertainty value of the boundary and non-overlapping areas and the logical conflict degree of land use transfer. The logical conflict degree of land use transfer is calculated by comparing the land use type pairs before and after the transformation with a preset land use type transfer rule matrix. The update result is determined as confirmed, pending manual review, or pending field verification based on the relationship between the comprehensive credibility index and the first and second preset thresholds.

[0008] By stacking temporal remote sensing imagery, digital elevation model data, and rasterized, categorically encoded ownership and land use identifier layers into a multi-source feature tensor along the channel dimension, the information loss of single spectral information in shadowed and topographically undulating areas is compensated for, enhancing the ability to distinguish between similar objects with different spectra and similar objects with different spectra. Through a multi-task deep learning model that outputs pixel-by-pixel change probability maps and pixel-by-pixel current land use semantic maps in parallel, the change location and the changed land use attributes are obtained synchronously in a single inference, avoiding inconsistencies caused by separate processing of change detection and land use interpretation. By enabling a dropout layer during the inference phase to perform multiple forward propagations and generating pixel-by-pixel features using prediction variance normalization, a multi-source feature tensor is generated. The deterministic heatmap makes the confidence level of the model available in a spatially distributed form. By adjusting the weights of the external energy term and the prior shape-fidelity energy term in the energy function of the active contour model, the contour evolves along the edge of change in the predicted reliable area and adheres to the original boundary of the existing ownership patches in the ambiguous area, reducing the disorderly expansion of pseudo-change boundaries. Furthermore, the comprehensive confidence index, which combines the average uncertainty value of the fused boundary and non-overlapping areas and the degree of logical conflict of land use transfer, is used to classify and confirm the data, requiring manual review or field verification. This directs human and field efforts towards patches that are indeed in doubt, improving the geometric accuracy, business logic consistency, and reliability of the results processing of ownership update.

[0009] Furthermore, constructing the multi-source feature tensor includes: resampling the temporal remote sensing image to a uniform spatial resolution using a cubic convolution interpolation algorithm; resampling the digital elevation model data to the spatial resolution using a bilinear interpolation algorithm; extracting the resampled spectral information matrix and elevation information matrix; converting the ownership and land use identification layer into a category coding matrix; and performing a concatenation and splicing operation on the spectral information matrix, the elevation information matrix, and the category coding matrix in the channel dimension to output the multi-source feature tensor.

[0010] By resampling temporal remote sensing imagery using cubic convolution interpolation and digital elevation model data using bilinear interpolation to achieve a unified spatial resolution, the spectral information matrix is ​​more detailed and the elevation information matrix is ​​smoother at undulation transitions, thus solving the inherent resolution heterogeneity problem of multi-source input data. Furthermore, the spectral information matrix, elevation information matrix, and category coding matrix obtained from the ownership and land use identification layers are concatenated and stitched together along the channel dimension, ensuring strict alignment of each channel in the same pixel row and column structure. This avoids spatial misalignment caused by resolution inconsistencies and provides a directly usable unified input for subsequent pixel-by-pixel inference.

[0011] Furthermore, the ownership and land category identification layer includes an ownership identification raster layer and a land category identification raster layer, both of which have the same projection system, spatial extent, resolution, and pixel row and column size as the time-series remote sensing image; the category coding process involves converting the land category identification into a one-hot coding matrix and the ownership identification into an embedded coding matrix to generate an ownership and land category coding matrix; the continuous spectral channel and elevation channel in the multi-source feature tensor are standardized, while the category coding channel retains the encoded feature form.

[0012] By ensuring that the ownership identifier raster layer and the land use identifier raster layer have the same projection system, spatial extent, resolution, and pixel row and column size as the time-series remote sensing image, it is guaranteed that historical ownership attributes and image observations correspond one-to-one at the same pixel location, facilitating subsequent reading of historical land use types by coordinate index. By converting land use identifiers into one-hot coding matrices and ownership identifiers into embedded coding matrices, it is prevented that category codes are treated as continuous values ​​in calculations, thus avoiding false ordering relationships. By standardizing only the continuous spectral and elevation channels, the category coding channels retain their encoded feature forms, preventing standardization from destroying the meaning of category code values ​​and enabling continuous features and category features to coexist stably in the same tensor.

[0013] Further, outputting the pixel-by-pixel change probability map and the pixel-by-pixel current land class semantic map includes: inputting the multi-source feature tensor into the backbone encoding network of the multi-task deep learning model to extract features and generate a multi-scale feature map; inputting the multi-scale feature map into the first branch decoder, and outputting the probability value of land cover change for each pixel through upsampling and convolution operations to form the pixel-by-pixel change probability map; inputting the multi-scale feature map into a second branch decoder independent of the first branch decoder, performing feature mapping and class probability calculation through the Softmax activation function, and performing maximum probability class index extraction on the class probability distribution of each pixel to output the pixel-by-pixel current land class semantic map.

[0014] By extracting multi-scale feature maps uniformly through a backbone coding network and feeding them into two independent branch decoders, change detection and land use semantic segmentation share the same underlying feature representation, reducing the processing burden caused by repeated feature extraction, while ensuring that the two outputs are naturally aligned in spatial location. The first branch decoder outputs the probability value of land cover changes pixel by pixel through upsampling and convolution, providing continuous driving information for subsequent boundary extraction. The second branch decoder obtains the category probability distribution through Softmax mapping and performs maximum probability category index extraction, directly providing the current land cover category index that can be used for land use transfer comparison, so that the change location and the transformed land cover are obtained simultaneously in one forward inference, which facilitates subsequent judgment on whether the land cover transfer conforms to the rules.

[0015] Furthermore, the backbone coding network adopts a residual network structure, which downsamples through multiple layers of residual blocks, expanding the number of feature channels while compressing the spatial resolution, and outputting multi-scale feature maps with multiple resolutions; the first branch decoder adopts a skip connection structure, which fuses the coding layer features of the corresponding scale during the upsampling process, and then reduces the number of channels to 1 through convolution, and outputs probability values ​​in the range of 0 to 1 through the Sigmoid activation function; the second branch decoder has the same structure as the first branch decoder but does not share parameters, and its output channel number is mapped to the total number of target land classes.

[0016] By employing a residual network structure in the backbone coding network and downsampling step-by-step with multiple residual blocks, the number of feature channels is expanded while compressing spatial resolution, enabling the representation of land cover morphology and texture at different scales and improving the feature degradation problem in deep network training. The first branch decoder uses skip connections to fuse coding layer features of the corresponding scale during step-by-step upsampling, restoring the spatial details lost during downsampling and making the output probability map more closely match the actual contour at the edge of the patch. Dimensionality reduction through terminal convolution and outputting probability values ​​in the 0 to 1 interval using sigmoid ensures that the degree of change is expressed as a continuous quantity rather than a direct binary determination, preserving boundary transition information. The second branch decoder has the same structure but does not share parameters, and the number of output channels is mapped to the total number of target land cover types, allowing the two tasks to maintain independent discrimination capabilities while sharing features, avoiding mutual interference between tasks.

[0017] Further, generating the pixel-by-pixel uncertainty heatmap includes: maintaining the activation state of the random dropout layer in the model decoding network during the model inference phase, and setting the node dropout probability to 0.15 to 0.30; continuously performing independent forward propagation inference a preset number of times on the multi-source feature tensor of the same input to obtain a preset number of prediction probability base map sets; calculating the discrete distribution sample variance of the corresponding value of each pixel in the prediction probability base map set for each pixel; normalizing the sample variance of all pixels and mapping it to the interval between 0 and 1, and arranging them according to spatial coordinates to generate the pixel-by-pixel uncertainty heatmap.

[0018] Further, the gradient field matrix of the two-dimensional image gradient of the pixel-by-pixel change probability map is calculated, and the gradient magnitude of the gradient field matrix is ​​extracted and normalized; the coordinates of the current contour node are obtained during the active contour evolution process, and the normalized uncertainty value of the corresponding coordinates is extracted from the pixel-by-pixel uncertainty heatmap; the difference between the constant 1 and the normalized uncertainty value is calculated as a reduction factor, and the reduction factor is multiplied by the negative of the normalized gradient magnitude of the corresponding coordinate in the gradient field matrix to obtain the external energy term.

[0019] Further, the Euclidean distance from the current contour node to the boundary of the existing ownership patch is calculated and normalized. The normalized Euclidean distance is multiplied by the normalized uncertainty value to obtain the prior shape-fidelity energy term. The external energy term is added to the prior shape-fidelity energy term to obtain the local total energy value controlling the contour evolution. The contour node position is updated by iteratively minimizing the total energy function. When the displacement of the contour node coordinates in multiple consecutive iterations is less than a preset pixel threshold, the contour is determined to converge. The converged contour line is used as the initial update boundary.

[0020] Further, within the patch range of the initial update boundary, pixels with probability values ​​greater than a preset threshold in the pixel-by-pixel change probability map are selected to form a set of changed pixels. For each pixel in the set of changed pixels, the historical land cover category index before the transformation is extracted from the ownership and land cover identification layer, and the current land cover category index after the transformation is extracted from the pixel-by-pixel current land cover semantic map. Using the historical land cover category index as the row coordinate and the current land cover category index as the column coordinate, the logical conflict penalty constant in the interval of 0 to 1 is extracted and the average value is calculated as the logical conflict degree of the land cover transfer.

[0021] By selecting pixels with high change probability within the initial updated boundary patch range using a preset threshold to form a changed pixel set, the interference of low-probability pixels on the statistical results is eliminated, ensuring that land use transfer judgments are based on areas that have actually changed. By extracting the land use category index before and after the transformation of the same pixel from the ownership and land use identifier layer and the pixel-by-pixel current land use semantic map, and using it as row and column coordinates to address and extract the logical conflict penalty constant in the interval of 0 to 1 in the land use type transfer rule matrix, the reasonableness of the conversion between different land use categories is measured on a unified scale, identifying transfer situations that do not conform to common sense. By averaging all penalty constants in the changed pixel set to obtain the land use transfer logical conflict degree, the misjudgment of individual pixels will not affect the overall conclusion, providing a stable business logic evaluation basis for the comprehensive credibility index.

[0022] In a second aspect, the present invention provides a natural resource rights confirmation system based on multi-source heterogeneous data, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the natural resource rights confirmation method based on multi-source heterogeneous data of the present invention is implemented.

[0023] The technical effects of this invention are as follows: It uniformly resamples and fuses temporal remote sensing imagery, digital elevation model data, and rasterized coded ownership and land use identification layers along the channel dimension, compensating for the information loss in shadowed and undulating areas in a single spectral dimension, and enhancing the ability to distinguish between different spectra of the same object and different objects with the same spectrum. Based on this, a shared coding network and a dual-branch decoder simultaneously output pixel-by-pixel change probability maps and current land use semantic maps, ensuring a strict spatial correspondence between the changed location and the transformed land use, avoiding inconsistencies caused by step-by-step processing. During the inference phase, a dropout layer is enabled, and predictions are made through multiple forward propagations. Variance normalization generates an uncertainty heatmap, allowing the model confidence level to participate in subsequent processing in a spatially distributed form. Based on this, the influence strength of the external energy term and the prior shape-preserving energy term is adjusted, so that the contours conform to the evolution of change edges in the credible region and adhere to the existing ownership boundaries in the ambiguous region, reducing the disorderly expansion of pseudo-change boundaries. Finally, the comprehensive credibility index of the average uncertainty value of the fusion boundary and non-overlapping region and the degree of logical conflict of land category transfer is used to classify and determine the criteria, so that confirmation, internal review and external verification each have their own basis, thereby improving geometric accuracy, consistency of business transfer logic and reliability of updated results processing. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart illustrating the natural resource ownership confirmation method based on multi-source heterogeneous data according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the predicted change probability curves of three types of pixels under multiple forward propagations in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the distribution of the comprehensive credibility index of each updated polygon object in an embodiment of the present invention; Figure 4 This is a schematic bar chart illustrating the performance comparison of three experimental schemes in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the structural block diagram of a natural resource rights confirmation system based on multi-source heterogeneous data according to an embodiment of the present invention. Detailed Implementation

[0025] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0026] Before further explaining the embodiments of this application, the nouns and terms involved in the embodiments of this application are explained, and the nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0027] 1. A patch refers to a spatial area in a map or remote sensing image that has a clear boundary, the same attributes, or the same land type, such as a piece of farmland, a piece of forest land, a body of water, or a property area.

[0028] 2. Existing ownership map patches refer to spatial map patch data that has been registered or has historical records, representing the scope of ownership of a certain natural resource.

[0029] In the first embodiment, the present invention proposes a method for natural resource ownership confirmation based on multi-source heterogeneous data, such as... Figure 1 As shown, it includes the following steps: S1. Acquire time-series remote sensing images, digital elevation model data, and existing ownership vector data of the area to be updated. Rasterize the vector data into ownership and land type identification layers, and perform category encoding on the ownership and land type identification layers. Stack these layers with the images and elevation data to construct a multi-source feature tensor. Input the multi-source feature tensor into a multi-task deep learning model to output a pixel-by-pixel change probability map and a pixel-by-pixel current land type semantic map in parallel. During the model inference stage, enable the dropout layer to perform multiple forward propagation calculations to generate and normalize the prediction variance to obtain a pixel-by-pixel uncertainty heatmap.

[0030] The system reads previous and current Sentinel multispectral remote sensing images and SRTM digital elevation model data of the area to be updated, and uses bilinear interpolation to resample and unify the spatial resolution. It then reads the shapefile vector data of existing ownership patches, extracting ownership identifier and land use code fields. Using the `features.rasterize` function from the Rasterio library, combined with the affine transformation matrix of the remote sensing images, the vector polygons are converted into ownership identifier raster layers and land use code raster layers. One-hot encoding or embedding encoding is performed on the ownership identifier raster layers and land use code layers to generate ownership and land use category coding matrices. The system concatenates previous spectral bands, current spectral bands, elevation values, and ownership and land use category coding matrices along the channel dimension to construct a multi-source feature tensor with a shape equal to the number of channels multiplied by the height multiplied by the width. The continuous spectral and elevation channels are standardized using the Z-score algorithm, while the category coding channels retain their encoded feature form. This tensor is then input into the deep learning framework.

[0031] A multi-task convolutional neural network model based on the UNet architecture is constructed, comprising a shared feature extraction encoder and two independent decoders, corresponding to the change detection task and the semantic segmentation task, respectively. The pre-trained model weights are loaded, and the Dropout layer is forcibly enabled by traversing the model modules, using the Monte Carlo Dropout algorithm to perform 30 independent forward propagations on the input multi-source feature tensor. The probability values ​​of the Sigmoid activation output from the change detection decoder are calculated after each forward propagation, and the per-pixel sample variance matrix of these 30 outputs along the inference count dimension is calculated. The values ​​of the variance matrix are linearly mapped to the interval between 0 and 1 to generate a per-pixel uncertainty heatmap; the mean of the 30 change prediction probabilities is taken as the per-pixel change probability map, and the mean of the Softmax class probability distributions of the 30 semantic segmentation branch outputs is calculated. The maximum probability class index is extracted for each pixel to obtain the per-pixel current land class semantic map.

[0032] In an optional embodiment, the vector data is rasterized into a land ownership and land type identifier layer, and after performing category encoding on the land ownership and land type identifier layer, it is stacked with imagery and elevation data to construct a multi-source feature tensor, including: A cubic convolution interpolation algorithm is used to resample temporal remote sensing images to a uniform spatial resolution; The digital elevation model data is resampled to spatial resolution using a bilinear interpolation algorithm; Extract the resampled spectral information matrix and elevation information matrix, convert the ownership and land category identification layers into category coding matrices, perform a cascaded stitching operation in the channel dimension, and output a multi-source feature tensor.

[0033] To address the inherent resolution heterogeneity issue of multi-source input data, a baseline spatial resolution parameter is selected, for example, set to 1.0 meter or 0.5 meter, with the target resolution set to 1.0 meter. and For dual-temporal four-band remote sensing images, namely red-green-blue and near-infrared time-series remote sensing images, such as GF2 images with an original resolution of 2 meters, a cubic convolution interpolation algorithm is used to construct a cubic polynomial weighted average using the gray values ​​of 16 adjacent pixels, and the remote sensing image is resampled to generate a spectral information matrix. For digital elevation models, such as ALOS PALSAR data with an original resolution of 12.5 meters, a bilinear interpolation algorithm is used to calculate the distance weights of 4 neighboring pixels for smoothing, and the images are resampled to the same 1.0-meter resolution to output a single-channel elevation information matrix.

[0034] The ownership boundaries and land use attribute data in the existing ownership patch vector data are rasterized to generate ownership and land use identifier raster layers with the same projection system, spatial extent, resolution, and pixel row and column size as the remote sensing image. For example, a raster layer with a size of [missing information] is generated. The raster slice data is then processed. Subsequently, the ownership identifier raster layer and the land category identifier raster layer are subjected to category encoding processing; wherein, preferably, the land category identifier is converted into a OneHot encoding matrix and the ownership identifier is converted into an embedded encoding matrix, or an equivalent category encoding method is used for conversion, thereby generating the ownership and land category encoding matrix.

[0035] The dual-temporal spectral information matrix, elevation information matrix, and ownership and land use category coding matrix are concatenated and stitched together along the channel dimension to generate a multi-source feature tensor. Specifically, the dual-temporal spectral information matrix contains 4 spectral bands each from the previous and current periods, for a total of 8 channels; the elevation information matrix contains 1 channel; and the ownership and land use category coding matrix contains several category coding channels. After merging the above matrices along the channel dimension, a shape of BatchSize× is obtained. A multi-source feature tensor of 1024×1024, where This is the sum of the number of spectral channels, the number of elevation channels, and the number of category coding channels. For example, when the category coding matrix contains 9 channels, =8+1+9=18.

[0036] In an optional embodiment, the multi-source feature tensor is input into a multi-task deep learning model, which outputs a pixel-wise change probability map and a pixel-wise current land class semantic map in parallel, including: Multi-source feature tensors are input into the backbone coding network of a multi-task deep learning model to extract features and generate multi-scale feature maps. The multi-scale feature map is input into the first branch decoder, and the probability value of each pixel changing ground features is output through upsampling and convolution operations to form a pixel-wise change probability map. The multi-scale feature map is input into the second branch decoder, which is independent of the first branch. The feature mapping and class probability calculation are performed through the Softmax activation function. The maximum probability class index extraction is performed on the class probability distribution of each pixel, and the output is a pixel-by-pixel semantic map of the current land class with land cover class index.

[0037] The multi-task deep learning model includes a backbone encoder network, a first-branch decoder, and a second-branch decoder. The backbone encoder network preferably employs a residual network structure to extract multi-scale spatial features from multi-source feature tensors. The first-branch decoder performs a change detection task, outputting a pixel-by-pixel change probability map. The second-branch decoder performs a land use semantic segmentation task, outputting a pixel-by-pixel semantic map of the current land use.

[0038] Specifically, the multi-source feature tensor is input into the backbone coding network, such as a ResNet50 network. The backbone coding network downsamples through multiple layers of residual blocks, compressing the spatial resolution while expanding the number of feature channels, and outputs the corresponding original... Figure 1Multiscale feature maps with resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32.

[0039] Multi-scale feature maps are input into two decoding branches for parallel processing. The first branch decoder uses a UNet-style skip connection structure to upsample the multi-scale feature maps output by the backbone coding network at each level. During the upsampling process, features from the corresponding coding layers are fused to recover spatial detail information, ultimately restoring the feature maps to their final size. The original spatial resolution is then used; subsequently, variation features are extracted through convolutional and batch normalization layers, and the end is processed using... Convolution reduces the number of channels to 1 and outputs a pixel-wise change probability map in the range of 0 to 1 through the Sigmoid activation function, where the pixel value represents the probability of a ground feature change occurring at that pixel.

[0040] The second branch decoder has the same structure as the first branch decoder, but their parameters are not shared. The second branch decoder upsamples and decodes the same multi-scale feature maps, restoring them to their original resolution. Convolution maps the number of output channels to the total number of target land types. ,For example This corresponds to eight land cover types, including paddy fields, woodlands, and built-up land. Then, a Softmax operation is performed on each pixel along the channel dimension to obtain the normalized probability distribution of the pixel's classification for each land cover type. Next, the maximum probability class index is extracted from the probability distribution to obtain the corresponding current land cover type index, ultimately generating a two-dimensional, single-channel, pixel-by-pixel semantic map of the current land cover type. Thus, the multi-task deep learning model can simultaneously output patch change information and current land cover attribute information during a single forward inference process. The predicted change probability curves of the three types of pixels under multiple forward propagations are shown below. Figure 2 As shown.

[0041] In an optional embodiment, during the model inference phase, a dropout layer is enabled to perform multiple forward propagation calculations to generate and normalize the prediction variance, resulting in a pixel-by-pixel uncertainty heatmap, including: Maintain the activation state of the randomly dropped layer in the model decoding network during the model inference phase; Perform independent forward propagation inference a preset number of times on the multi-source feature tensor of the same input to obtain a preset number of prediction probability base map sets; For each pixel, calculate the discrete distribution sample variance of the corresponding value in the predicted probability base map set; The sample variance of all pixels is normalized and mapped to the interval between 0 and 1, and then arranged according to spatial coordinates to generate a pixel-by-pixel uncertainty heatmap.

[0042] The Monte Carlo dropout method is used to represent the spatial uncertainty of the model output. During the model inference phase, the random dropout layer within the model decoding network remains active, and the node dropout probability, i.e., ..., is... The parameter should be set between 0.15 and 0.30, for example, set to... For an input spatial dimension of... Multi-source feature tensors, executed continuously a preset number of times For example, let's say Independent forward propagation inference operations.

[0043] Because the Dropout layer randomly cuts off a portion of the neurons' responses with probability each time, this perturbation will cause differences in the results of 30 forward inferences. 30 samples of the same size were collected. The independent prediction probability matrices are stacked according to the inference count dimension to construct a prediction probability base map set. After obtaining the base map set, for any single pixel point in the two-dimensional space... , Extract the 30 different output probability values ​​along the inference count dimension, such as 0.71, 0.75, 0.69, etc. Calculate the sample variance of this discrete distribution of values ​​using the unbiased variance formula, which is the sum of the squares of the differences between the values ​​and the arithmetic mean divided by... The original variance matrix is ​​obtained. A max-min normalization algorithm is used to globally calculate the range of the original variance matrix, linearly mapping each variance value in the matrix to the interval between 0 and 1. Based on the horizontal and vertical geospatial coordinate structure of the input slice, a pixel-by-pixel uncertainty heatmap representing confidence fluctuations is output. Areas with values ​​closer to 1 are considered "hot," indicating lower model confidence in predicting that area.

[0044] S2, based on the pixel-wise change probability map, uses the boundary of the existing ownership patch as the initial contour and adopts an active contour model to extract the initial update boundary. The energy function of the active contour model is controlled by the pixel-wise uncertainty heatmap, which includes an external energy term and a prior shape-fidelity energy term. The weights of the external energy term and the prior shape-fidelity energy term are determined based on local uncertainty.

[0045] The Sobel gradient field is calculated based on the pixel-wise change probability map and used as the image driving force basis for the active contour model. At the same time, the boundary vector node coordinate sequence of the existing ownership patches is extracted and used as the initial contour point set of the active contour model.

[0046] The active contour model is implemented using a parametric active contour algorithm. During energy function construction, for each contour point on the current contour, the normalized uncertainty value of its pixel position in the pixel-by-pixel uncertainty heatmap is read. The difference between the constant 1 and the normalized uncertainty value is used as the adjustment coefficient of the external energy term and multiplied by the basic external weight coefficient to regulate the attraction of the image gradient on contour evolution. Simultaneously, the Euclidean distance from the current contour point to the boundary of the initial existing ownership patch is calculated, and this distance is used as the basis for the prior shape-fidelity energy term, with the product of the normalized uncertainty value and the basic shape-fidelity weight coefficient as its weight. Subsequently, the contour point positions are updated by iteratively minimizing the total energy function of the active contour model. When the displacement of the contour point coordinates in multiple consecutive iterations is less than a preset pixel threshold, contour convergence is determined, and the converged contour line is used as the initial update boundary.

[0047] In an optional embodiment, the energy function of the active contour model is modulated by a pixel-wise uncertainty heatmap, comprising an external energy term and a prior shape-fidelity energy term. The weights of the external energy term and the prior shape-fidelity energy term are determined based on local uncertainty, including: The gradient field matrix is ​​generated by calculating the gradient of the two-dimensional image from the pixel-by-pixel change probability map, and the gradient magnitude of the gradient field matrix is ​​extracted and normalized. During the active contour evolution process, the coordinates of the current contour node are obtained, and the normalized uncertainty value of the corresponding coordinates is extracted from the pixel-by-pixel uncertainty heatmap. The difference between constant 1 and the normalized uncertainty value is calculated as a reduction factor. The coefficient is multiplied by the negative of the normalized gradient magnitude of the corresponding coordinate in the gradient field matrix to obtain the dimensionless external energy term. Calculate the Euclidean distance from the current contour node to the boundary of the existing ownership patch and normalize it. Multiply the normalized Euclidean distance by the normalized uncertainty value to obtain the dimensionless prior shape-fidelity energy term. The external energy term is added to the prior shape-fidelity energy term to obtain the local total energy value controlling the contour evolution.

[0048] For the output of the first branch A pixel-wise probability map of size variation, using a size of The Sobel convolution operator is used for the image respectively. direction and The first-order partial derivatives in space are obtained by calculating the first-order partial derivatives in the direction, and the arithmetic square root of the sum of the squares of the two is calculated to obtain the absolute gradient magnitude of each pixel. Then, the gradient is normalized to generate a gradient field matrix in the interval between 0 and 1.

[0049] When initiating polygon evolution iterations of the active contour model, for example, limiting the maximum iterations to 500, the control nodes that constitute the initial closed contour of the polygon are tracked. When processing reaches a certain current contour node... And obtain the node's coordinates in the image coordinate system. When using pixel coordinates, the normalized uncertainty value of the corresponding coordinate is extracted from the pixel-by-pixel uncertainty heatmap using that coordinate as an index. ,For example =0.8. The difference between the constant 1 and the uncertainty value, i.e., 1 - 0.8 = 0.2, is used as the reduction factor. The normalized gradient field matrix is ​​extracted from this. gradient magnitude at ,For example =0.9, multiply the reduction factor by the negative of the gradient magnitude, i.e. The external energy term of the node is obtained. For example, 0.2 × (-0.9) = -0.18. This negative value design causes the contour to converge toward the gradient extremum region, that is, the change boundary region, but in the high uncertainty region, that is, when the reduction coefficient is small, the effect of external image information will be weakened.

[0050] To introduce shape constraints in areas of image blur, the current node is calculated. The straight-line Euclidean distance to the original boundary of the nearest existing ownership patch is normalized to obtain a distance parameter between 0 and 1. ,For example =0.6. The normalized distance parameter is then compared with the extracted normalized uncertainty value. That is, multiplying by 0.8 to find the prior shape-fidelity energy term. That is, 0.6 × 0.8 = 0.48. When the node is located in the high uncertainty region, that is... Approaching 1 and moving away from the original boundary, i.e. When the value is large, the penalty term increases. In each evolutionary microstep, the external energy term (-0.18) is added to the prior shape-fidelity energy term (0.48) to obtain the local total energy value (0.30), which is used as the evaluation value when the node moves to the candidate position in its neighborhood. This mechanism tends to rely more on the evolution of image change edges when the model prediction is reliable, while it tends to adhere to the original boundary when the model uncertainty is high, thereby reducing the spread of spurious changes.

[0051] S3. Calculate the comprehensive credibility index for the map patches of the initial update boundary. The comprehensive credibility index integrates the average uncertainty value of the boundary and non-overlapping areas and the logical conflict degree of land use transfer. The logical conflict degree of land use transfer is calculated by comparing the land use type pairs before and after the transformation with the preset land use type transfer rule matrix. Based on the relationship between the comprehensive credibility index and the first and second preset thresholds, the update result is determined as confirmed, pending manual review, or pending field verification.

[0052] Based on the updated initial boundary patches and existing ownership patches, corresponding polygons are constructed, and spatial difference analysis is performed on them to extract the non-overlapping regions between the updated patches and existing ownership patches. Subsequently, the pixel values ​​in the pixel-by-pixel uncertainty heatmap are statistically analyzed for pixels within the updated boundary buffer and pixels within the non-overlapping regions, and the average uncertainty of the boundary and the average uncertainty of the non-overlapping regions are calculated.

[0053] Simultaneously, a land cover type transfer rule matrix is ​​pre-constructed, where the matrix rows represent the historical land cover type before the transformation, the matrix columns represent the current land cover type after the transformation, and the matrix elements are logical conflict penalty constants in the range of 0 to 1. For pixels with a change probability greater than a preset threshold within the initial update boundary patch area, their historical land cover index and current land cover index are read respectively, and the corresponding logical conflict penalty constant is found in the land cover type transfer rule matrix based on the historical land cover index and the current land cover index; the logical conflict penalty constants corresponding to all changed pixels are averaged to obtain the land cover transfer logical conflict degree.

[0054] Furthermore, the average uncertainty of the boundary, the average uncertainty of the non-overlapping area, and the logical conflict degree of land use transfer are weighted and summed according to preset weights to obtain the total error penalty value. The difference between the constant 1 and the total error penalty value is subjected to amplitude limiting to obtain the comprehensive credibility index. The first preset threshold is set to 0.85, and the second preset threshold is set to 0.60. When the comprehensive credibility index is greater than or equal to 0.85, the updated map piece is marked as confirmed and entered into the database. When the comprehensive credibility index is greater than or equal to 0.60 and less than 0.85, the updated map piece is marked as awaiting manual review. When the comprehensive credibility index is less than 0.60, a field verification task package is generated, and the updated map piece is marked as awaiting field verification.

[0055] In an optional embodiment, the logical conflict degree of land use type transfer is calculated by comparing the land use type pairs before and after the transformation with a preset land use type transfer rule matrix, including: Within the patch area where the initial update boundary is generated, pixels with probability values ​​greater than a preset threshold in the pixel-by-pixel change probability map are selected to form a set of changed pixels; For each pixel in the changed pixel set, extract the historical land feature category index before the pixel transformation from the ownership and land type identification layer, and extract the current land feature category index after the transformation from the pixel-by-pixel current land type semantic map; Using the historical land cover category index as the row coordinate and the current land cover category index as the column coordinate, the system addresses the data in the preset land cover type transfer rule matrix and extracts the corresponding logical conflict penalty constant in the interval between 0 and 1. Summing up the logical conflict penalty constants of all pixels in the changed pixel set and calculating the average value, the average value is used as the logical conflict degree of land category transfer.

[0056] For the closed update patch output by the active contour model, within the polygonal region of the patch, using a binarization threshold of 0.5, active pixels with values ​​greater than 0.5 in the pixel-by-pixel change probability map are filtered and retained. These active pixels are then grouped into a set of changing pixels of size N, for example, N=450. For any specified pixel in this set... The corresponding land use code is read from the input historical land use identifier raster layer, or the historical land use category index of the pixel is restored according to the coding mapping relationship of the category coding matrix. Let the extracted... This indicates shrubland; the value of the corresponding coordinate in the pixel-by-pixel semantic map of the current land cover type is indexed to obtain the current land cover type index. Let the extracted... , indicating bare land.

[0057] The pre-defined land cover type transfer rule matrix is ​​invoked. This matrix is ​​a two-dimensional lookup table, where each element represents the constraint level for conversion between specific land cover types. For example, grassland to dry land is set to a low value of 0.1, and forest to bare land is set to a high value of 0.9. This is achieved by using the acquired historical index. Mapped to row coordinate dimension, i.e., row 5, the current index... The mapping is done along the column coordinate dimension, specifically the 3rd column. Addressing within this two-dimensional array, the logical conflict penalty constant within the range of 0 to 1 is extracted, assuming a penalty value P = 0.90. For all N changing pixels within the patch set, the corresponding penalty constant is extracted. These penalty constants are then summed to obtain the total penalty, which is divided by the total number of valid pixels N in the set to obtain the arithmetic mean. Assuming the total penalty for 450 pixels in this patch is 315, the arithmetic mean is 315 / 450 = 0.70. This arithmetic mean is then set as the land use transfer logical conflict degree.

[0058] In an optional embodiment, a comprehensive credibility index is calculated for the patches of the initial updated boundary. The comprehensive credibility index integrates the average uncertainty value of the boundary and non-overlapping areas and the logical conflict degree of land use transfer, including: Extract all pixels on the initial update boundary and calculate the average value of the pixels in the pixel-by-pixel uncertainty heatmap as the first evaluation index; Extract all pixels within the non-overlapping area of ​​the existing ownership patches and the updated generated patches, and calculate the average value of the pixels in the pixel-by-pixel uncertainty heatmap as the second evaluation index. The degree of logical conflict in land category transfer is used as the third evaluation indicator; The first evaluation indicator, the second evaluation indicator, and the third evaluation indicator are assigned preset weight coefficients and then weighted and summed to obtain the total error penalty value. The difference between the constant 1 and the total error penalty is subjected to amplitude limiting and mapped to the standard real number range of 0 to 1 to obtain the comprehensive credibility index.

[0059] Multidimensional evaluation metrics are extracted for a single updated patch. First, a set of boundary pixel coordinates is extracted along the closed boundary of the updated patch, for example, the boundary contains 200 pixel coordinates. Then, the uncertainty values ​​corresponding to the above 200 boundary pixels are read from the pixel-by-pixel uncertainty heatmap, and their arithmetic mean is calculated, for example, the average value is 0.12. This average value is used as the first evaluation metric. .

[0060] Obtain all pixels within the non-overlapping region between the existing ownership patch region and the generated updated patch region, for example, containing 600 pixels. Read the uncertainty values ​​corresponding to these 600 coordinates from the uncertainty heatmap and calculate their average value, for example, a mean of 0.18, as the second evaluation metric. The degree of logical conflict in land category transfer is used as the third evaluation indicator. The assumed value is 0.45. Preset weighting coefficients are assigned to each of the three evaluation indicators. For example, the weighting coefficient for the first indicator is... The weighting coefficient for the second indicator is set to 0.25. The weighting coefficient for the third indicator is set to 0.35. Set it to 0.40 to ensure the sum of all weights. =1.0. The total error penalty is obtained by performing a weighted calculation. Substituting the values ​​into the example, we get 0.25×0.12+0.35×0.18+0.40×0.45=0.273. Subtracting this total error penalty from the scalar constant 1 yields the residual, i.e., 1-0.273=0.727. Applying a limiting function to set constraints, with 0 as the lower bound and 1 as the upper bound for truncation correction, the output value of 0.727 is a floating-point number used as the comprehensive credibility index for this individual patch object. The closer it is to 1, the higher the spatial accuracy and business logic compliance of the extraction.

[0061] In an optional embodiment, the update result is determined as confirmed, pending manual review, or pending field verification based on the relationship between the comprehensive credibility index and the first and second preset thresholds, including: Set the first preset threshold to be greater than the second preset threshold; If the overall credibility index is greater than or equal to the first preset threshold, the update result is determined to meet the automatic confirmation condition, a confirmation command is triggered, and the updated patch is entered into the database. If the overall credibility index is greater than or equal to the second preset threshold and less than the first preset threshold, it is determined to be subject to manual review. The map patch is marked and assigned to the review queue of the internal review system to await manual interaction confirmation. If the overall credibility index is less than the second preset threshold, it is determined that it needs to be verified in the field. The field verification task package containing the coordinates of the map patch and the abnormal features is automatically generated and sent to the field survey terminal to perform on-site survey and closed-loop confirmation.

[0062] Set the first preset threshold to 0.85; set the second preset threshold to 0.60. Receive the calculated comprehensive credibility index. .like For example, calculated If the update result meets the automatic confirmation conditions, a data insertion or update command is triggered, and the geometric and attribute data of the updated patch are synchronously written into the spatial database for confirmation. ,For example If the map patch is marked as requiring manual review, it will be added to the review queue for manual confirmation. ,For example If the updated polygon is deemed to have a high need for verification, a field verification task package is generated, including the polygon's spatial extent, boundary coordinate nodes, type of ambiguity, comprehensive credibility index, regional uncertainty index, and land use transfer logic conflict degree. The polygon is then marked as requiring field verification. On-site personnel perform fixed-point verification, on-site photography for evidence collection, and re-survey according to the field verification task package, and feed the verification results back to the system to complete the closed-loop verification process. The distribution of the comprehensive credibility index for each updated polygon object is as follows: Figure 3 As shown.

[0063] The experiment was conducted based on dual-temporal high-resolution remote sensing imagery covering 2000 square kilometers of landform and corresponding existing ownership vector datasets. The hardware environment employed multiple high-performance GPUs for parallel processing. The evaluation criteria included the intersection-over-union ratio (IoU) and boundary F1 score. The experiment defined a baseline group using a conventional multi-task segmentation network that only inputs spectral information; an ablation group using a network that stitches together multi-source features such as elevation and ownership but outputs a binarized version; and a complete group incorporating a scheme that integrates multi-source feature fusion to calculate an uncertainty heatmap and controls the contour evolution process.

[0064] Experimental data were obtained through independent forward inference comparisons on a unified test set. The cross-union ratio of the variation patches in the baseline group was [value missing]. The boundary F1 score is Incorporating multi-source ablation groups will increase the crossover ratio to [missing information]. The boundary F1 score rose to The complete set measured with the mechanism of this invention achieved a crossover ratio (CRO). The boundary F1 score reached The proportion of map patches whose overall credibility index reached the confirmation threshold increased. .like Figure 4 As shown.

[0065] The fusion and cascading of multi-source features compensates for the lack of underlying information in shadowed and undulating terrain areas by a single spectral dimension, enhancing the network's ability to distinguish features of the same object with different spectra and the same spectrum with different objects, thus improving segmentation accuracy. After utilizing a contour model driven by uncertainty heatmap regulation, the model more closely adheres to change edges in areas with high prediction confidence, while in ambiguous areas with low confidence, it adheres to the original ownership boundaries due to shape-fidelity energy penalties. This mechanism reduces the disorderly expansion and spread of spurious change boundaries, improving the geometric accuracy and consistency of updated patches in terms of business transfer logic.

[0066] On the other hand, this invention also provides a natural resource rights confirmation system based on multi-source heterogeneous data. For example... Figure 5 The natural resource ownership confirmation system based on multi-source heterogeneous data shown includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the natural resource ownership confirmation method based on multi-source heterogeneous data according to the first aspect of the present invention.

[0067] The natural resource ownership confirmation system based on multi-source heterogeneous data also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

Claims

1. A method for confirming natural resource ownership based on multi-source heterogeneous data, characterized in that, Includes the following steps: Acquire time-series remote sensing images, digital elevation model data, and existing ownership patch vector data of the area to be updated. Rasterize the vector data into ownership and land use identification layers. After classifying the ownership and land use identification layers, stack them with the images and elevation data to construct a multi-source feature tensor. The multi-source feature tensor is input into a multi-task deep learning model to output a pixel-by-pixel change probability map and a pixel-by-pixel current land category semantic map in parallel. During the model inference stage, a dropout layer is enabled to perform multiple forward propagation calculations to generate and normalize the prediction variance to obtain a pixel-by-pixel uncertainty heatmap. Based on the pixel-wise change probability map, the initial update boundary is extracted using an active contour model with the boundary of the existing ownership patch as the initial contour. The energy function of the active contour model is controlled by the pixel-wise uncertainty heatmap and includes an external energy term and a prior shape-fidelity energy term. The weights of the external energy term and the prior shape-fidelity energy term are determined based on local uncertainty. A comprehensive credibility index is calculated for the patches of the initial updated boundary. The comprehensive credibility index integrates the average uncertainty value of the boundary and non-overlapping areas and the logical conflict degree of land use transfer. The logical conflict degree of land use transfer is obtained by comparing the land use type pairs before and after the transformation with the preset land use type transfer rule matrix. Based on the relationship between the comprehensive credibility index and the first and second preset thresholds, the update result is determined as confirmed, pending manual review, or pending field verification.

2. The method for confirming natural resource rights based on multi-source heterogeneous data according to claim 1, characterized in that, Constructing the multi-source feature tensor includes: The temporal remote sensing images are resampled to a uniform spatial resolution using a cubic convolution interpolation algorithm. The digital elevation model data is resampled to the spatial resolution using a bilinear interpolation algorithm; Extract the resampled spectral information matrix and elevation information matrix, and convert the ownership and land category identification layer into a category coding matrix. Perform a concatenation and splicing operation on the spectral information matrix, the elevation information matrix and the category coding matrix in the channel dimension to output the multi-source feature tensor.

3. The method for confirming natural resource rights based on multi-source heterogeneous data according to claim 2, characterized in that, The ownership and land category identification layer includes an ownership identification raster layer and a land category identification raster layer, both of which have the same projection system, spatial extent, resolution, and pixel row and column size as the time-series remote sensing image. The category coding process involves converting land category identifiers into one-hot coding matrices and ownership identifiers into embedded coding matrices to generate ownership and land category coding matrices. The continuous spectral and elevation channels in the multi-source feature tensor are standardized, while the category-coded channels retain the encoded feature form.

4. The method for confirming natural resource rights based on multi-source heterogeneous data according to claim 1, characterized in that, Outputting the pixel-by-pixel change probability map and the pixel-by-pixel current land class semantic map includes: The multi-source feature tensor is input into the backbone coding network of the multi-task deep learning model to extract features and generate multi-scale feature maps. The multi-scale feature map is input into the first branch decoder, and the probability value of each pixel undergoing ground feature change is output through upsampling and convolution operations to form the pixel-by-pixel change probability map. The multi-scale feature map is input into a second branch decoder that is independent of the first branch decoder. The feature map and class probability are calculated by using the Softmax activation function. The maximum probability class index is extracted for the class probability distribution of each pixel, and the pixel-by-pixel current land class semantic map is output.

5. The method for confirming natural resource rights based on multi-source heterogeneous data according to claim 4, characterized in that, The backbone coding network adopts a residual network structure, which downsamples through multiple layers of residual blocks, compressing the spatial resolution while expanding the number of feature channels, and outputting multi-scale feature maps with multiple resolutions. The first branch decoder adopts a skip connection structure. During the upsampling process, it fuses the coding layer features of the corresponding scale and then reduces the number of channels to 1 through convolution. It then outputs the probability value in the range of 0 to 1 through the Sigmoid activation function. The second branch decoder has the same structure as the first branch decoder but does not share parameters, and its output channel number is mapped to the total number of target land types.

6. The method for confirming natural resource rights based on multi-source heterogeneous data according to claim 1, characterized in that, Generating the pixel-by-pixel uncertainty heatmap includes: maintaining the activation state of the random dropout layer in the model decoding network during the model inference phase, and setting the node dropout probability to 0.15 to 0.30; Perform independent forward propagation inference a preset number of times on the multi-source feature tensor of the same input to obtain a preset number of prediction probability base map sets; For each pixel, calculate the discrete distribution sample variance of the corresponding value of that pixel in the predicted probability base map set; The sample variance of all pixels is normalized and mapped to the interval between 0 and 1, and then arranged according to spatial coordinates to generate the pixel-by-pixel uncertainty heatmap.

7. The method for confirming natural resource rights based on multi-source heterogeneous data according to claim 1, characterized in that, The gradient field matrix is ​​generated by calculating the two-dimensional image gradient of the pixel-by-pixel change probability map, and the gradient magnitude of the gradient field matrix is ​​extracted and normalized. During the active contour evolution process, the coordinates of the current contour node are obtained, and the normalized uncertainty value of the corresponding coordinates is extracted from the pixel-by-pixel uncertainty heatmap. The difference between the constant 1 and the normalized uncertainty value is calculated as a reduction factor. The reduction factor is then multiplied by the negative of the normalized gradient magnitude of the corresponding coordinate in the gradient field matrix to obtain the external energy term.

8. The method for confirming natural resource rights based on multi-source heterogeneous data according to claim 7, characterized in that, Calculate the Euclidean distance from the current contour node to the boundary of the existing ownership patch and normalize it. Multiply the normalized Euclidean distance by the normalized uncertainty value to obtain the prior shape fidelity energy term. The external energy term is added to the prior shape-fidelity energy term to obtain the local total energy value controlling the contour evolution; The contour node positions are updated by iteratively minimizing the total energy function. When the displacement of the contour node coordinates in multiple consecutive iterations is less than a preset pixel threshold, the contour is determined to converge, and the converged contour line is used as the initial update boundary.

9. The method for confirming natural resource rights based on multi-source heterogeneous data according to claim 1, characterized in that, Within the patch area of ​​the initial update boundary, select pixels with probability values ​​greater than a preset threshold in the pixel-by-pixel change probability map to form a set of changed pixels; For each pixel in the changed pixel set, extract the historical land cover category index before the transformation from the ownership and land cover identification layer, and extract the current land cover category index after the transformation from the pixel-by-pixel current land cover semantic map; Using the historical land cover category index as the row coordinate and the current land cover category index as the column coordinate, the logical conflict penalty constant in the interval from 0 to 1 is extracted and the average value is taken as the logical conflict degree of the land cover transfer.

10. A natural resource ownership confirmation system based on multi-source heterogeneous data, characterized in that, It includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the natural resource rights confirmation method based on multi-source heterogeneous data as described in any one of claims 1 to 9 is implemented.

Images