AI-based territorial change investigation proof system and method

Abstract

The invention discloses an AI-based territorial change investigation proof system and method, and belongs to the technical field of territorial investigation. The method specifically comprises the following steps: S1, multi-source data acquisition: synchronously acquiring a satellite remote sensing image, an unmanned aerial vehicle aerial image, GIS basic geographic data and mobile terminal field acquisition data of a land investigation area; s2, historical data preprocessing: taking land investigation historical reference data, and sequentially performing denoising, coordinate unification, format standardization, data enhancement and sample labeling processing to generate a training data set and a reference comparison data set; by unifying a coordinate system, a data format and de-noising enhancement processing, the problem of inconsistency of multi-source data is solved, manual operation errors are reduced, and data consistency of large-scale investigation is guaranteed. A hybrid model of CNN, Transform and U-Net is fused, an attention mechanism and a feature pyramid structure are matched, the change area recognition precision and the boundary segmentation effect are improved, and the manual intervention cost is greatly reduced through the automatic process.

Description

Technical Field

[0001] This invention relates to the field of land survey technology, and in particular to an AI-based evidence collection system and method for land change surveys. Background Technology

[0002] Land use change surveys are a fundamental and legally mandated task in natural resource management. Their core objective is to dynamically grasp real-time changes in land use types, scope, and ownership through routine and precise change monitoring and evidence collection, providing crucial data support for land approval, law enforcement supervision, adjustments to territorial spatial planning, and the implementation of farmland protection responsibilities. According to requirements, land use change surveys must achieve "accurate land type, accurate boundaries, precise area, and adequate evidence." The evidence collection stage, as a key basis for verifying the results of the change survey, directly determines the overall quality and speed of progress of the land use survey work through its efficiency and accuracy.

[0003] With the advancement of the modernization of my country's land space governance system and governance capabilities, the scope of land change surveys has expanded from routine annual surveys to special surveys (such as the conversion of arable land to non-grain crops, illegal expansion of construction land, and changes within ecological protection red lines). The frequency of surveys has increased from once a year to quarterly monitoring and even monthly verification, placing unprecedented demands on the automation, accuracy, and standardization of evidence collection. Current technologies suffer from inconsistent multi-source data formats, chaotic coordinate systems, and reliance on manual preprocessing, resulting in low efficiency and large errors, making it difficult to meet the data consistency requirements of large-scale surveys. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide an evidence-gathering system and method for land change surveys based on AI; which can solve the problem of insufficient standardization in data collection and processing in existing technologies.

[0005] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, an AI-based method for providing evidence in land change surveys, specifically including the following steps:

[0006] S1. Multi-source data acquisition: Simultaneously acquire satellite remote sensing images, drone aerial images, GIS basic geographic data, and mobile terminal field data of the land survey area;

[0007] S2. Historical data preprocessing: retrieve historical benchmark data from the land survey, and sequentially perform noise reduction, coordinate unification, format standardization, data augmentation, and sample labeling to generate training datasets and benchmark comparison datasets;

[0008] S3. Deep Feature Extraction: Based on the preprocessed data and multi-source collected data, a CNN convolutional neural network is used to extract primary features. Based on the primary features, a CBAM attention mechanism is introduced to obtain intermediate features. Based on the intermediate features, transposed convolution is used for upsampling. An FPN feature pyramid structure is introduced, and high-level features are obtained through convolution.

[0009] S4. Hybrid Model Training: Using high-level features as input, construct a Transformer-U-Net hybrid deep learning recognition model, and complete the training through an adaptive optimization strategy until the model evaluation metrics reach the preset standards.

[0010] S5. Intelligent identification of land use change survey: After preprocessing and extracting deep features from the multi-source data of the area to be investigated, the data is input into the trained model to automatically identify the area, scope, area and type of land use change.

[0011] S6. Automatic generation of supporting evidence: Based on the identification results, a structured and standardized supporting evidence package is automatically generated in accordance with the "Technical Regulations for Land Change Survey".

[0012] S7. Blockchain-based evidence storage and AI-based dual-level review: The evidence package is uploaded to the consortium blockchain for encrypted storage and a timestamp is generated. The completeness, compliance, and authenticity of the materials are verified through AI and automated review, as well as secondary manual review, and the final review result is output.

[0013] Furthermore, in step S1, the data collected on-site by the mobile terminal includes high-definition photos with a resolution of 50 megapixels or higher, 4K resolution videos, centimeter-level GPS positioning information, electronic signature confirmation information from investigators, and on-site environmental parameters.

[0014] Furthermore, in step S2, the preprocessing specifically includes:

[0015] Denoising: A combined Gaussian filtering and termination foil denoising algorithm is used to remove salt-and-pepper noise and Gaussian noise from the image;

[0016] Coordinate unification: All data are uniformly converted to the 2000 National Geodetic Coordinate System using a seven-parameter coordinate transformation model;

[0017] Format standardization: Convert data of different formats into GeoTIFF image format and JSON attribute format;

[0018] Data augmentation: Random flipping, rotation, brightness and contrast adjustment, and Gaussian blur are used to augment the training samples;

[0019] Sample standard: Label 15 types of land use change scenarios according to the "Classification of Current Land Use" GB / T21010-2017, with labeling accuracy ≤0.1 meters.

[0020] Furthermore, step S3 specifically includes the following steps:

[0021] S31. A 6-layer CNN convolutional neural network structure is adopted. The input is standardized 3-channel image data, and the output is a basic feature map containing texture features, spectral features and spatial location features. Among them, texture features include the granularity and texture density of land cover; spectral features include the band reflectance differences of different land types; and spatial location features include the relative coordinate relationship of land types.

[0022] S32. Introducing a one-dimensional mechanism of CBAM, which is divided into channel attention and spatial attention. Channel attention is achieved by fusing global average reddening and global maximum reddening to assign 1.2-2.5 times weight to change-related feature channels, such as the built-up land spectral channel and the cultivated land texture channel, to suppress interference from irrelevant features. Spatial attention is achieved by convolution to capture spatial correlation information and combine it with temporal change features to finally obtain intermediate features.

[0023] S33. Four layers of transposed convolution are used to upsample intermediate features. An FPN feature pyramid structure is introduced to fuse intermediate features at different scales, mine features in multi-scale changing regions, and perform feature dimensionality reduction and fusion through convolution kernels to finally obtain high-level features.

[0024] Furthermore, in step S4, a Transformer encoder with a multi-head attention mechanism and a feedforward neural network in each layer is used as the encoder, and a U-Net decoder is used. The high-level features output by the Transformer encoder are concatenated with the low-level features extracted by the CNN through skip connections to make up for the Transformer's shortcomings in capturing local features and improve the accuracy of boundary segmentation of change regions. A sigmoid activation function is used to output a probability map of change regions, and a softmax activation function is used to output the classification result of change type. The training set, validation set and test set are divided in a 7:2:1 ratio. A weighted fusion of the Dice loss function and the cross-entropy loss function is used as the loss function, and an evaluation index including accuracy, recall and precision is used.

[0025] Furthermore, in step S5, the multi-source data of the area to be investigated are preprocessed and deep feature extraction is performed sequentially to obtain high-level features. The deep features are input into the trained hybrid model, and the probability of the changed area and the classification results of the change type are input. Morphological opening operation is used to remove small noise points. The boundary of the changed area is optimized by polygon fitting. The changed area is calculated, and structured data containing the boundary coordinates of the changed area, area, change type, probability value of model recognition, and time node prediction is generated. It supports export in both JSON and XML formats.

[0026] Furthermore, in step S6, based on the structured results of model recognition, a preset template library and a legal provisions database are invoked to automatically associate relevant information and generate various documents; the template library includes preset change comparison chart templates and change detail report templates; the legal provisions database includes the "Land Management Law," the "Land and Space Planning Law," and the "Technical Regulations for Land Change Survey," and automatically generates a list of legal bases through keyword matching; the QR code of the electronic evidence report adopts dynamic generation technology, containing the unique identifier of the evidence material, the evidence storage location, and query permissions. After scanning, it can jump to the blockchain evidence storage query page. The data traceability file records the time, operator, equipment information, and parameter settings of each step of the process to ensure full traceability.

[0027] Furthermore, in step S7, the blockchain evidence storage adopts a main chain + side chain architecture. The main chain stores core evidence materials and evidence storage information, while the side chain stores full-process log data. The PBFT consensus mechanism is used, and the data is encrypted and transmitted and stored using the SM2 asymmetric encryption algorithm. The data hash value is generated using the SM3 algorithm to ensure that the data is tamper-proof. The consortium chain nodes are led by the natural resources authorities and include third-party investigation agencies, supervision units, data storage centers, and notary offices. Node access requires identity authentication and permission review, and different nodes are assigned different read and write permissions. An AI review model verifies the data according to three dimensions: completeness, compliance, and authenticity. Materials that pass the review are directly archived, while materials that fail the review or are suspected of being abnormal are pushed to manual review.

[0028] According to another aspect of the present invention, an AI-based evidence collection system for land change surveys is provided. This system is used to implement the above-described AI-based evidence collection method for land change surveys, including: a multi-source data acquisition module, a historical data preprocessing module, a deep feature extraction module, a hybrid model training module, a change intelligent recognition module, an automatic evidence material generation module, and a blockchain evidence storage and AI dual-level review module.

[0029] Multi-source data acquisition module: used to acquire satellite remote sensing images, UAV aerial images and GIS basic geographic data, and to collect high-definition photos, videos, location information, electronic signatures and environmental parameters on site through mobile terminals, supporting offline acquisition and data synchronization after network recovery;

[0030] Historical data preprocessing module: used to perform noise reduction, coordinate unification and format standardization on historical benchmark data, expand training samples through data augmentation, complete sample labeling according to land change type, and generate training dataset and benchmark comparison dataset;

[0031] The deep feature extraction module is used to extract primary features based on preprocessed data and multi-source collected data using a CNN convolutional neural network. Based on the primary features, a CBAM attention mechanism is introduced to obtain intermediate features. Based on the intermediate features, transposed convolution is used for upsampling. An FPN feature pyramid structure is introduced, and high-level features are obtained through convolution.

[0032] Hybrid Model Training Module: Using high-level features as input, a Transformer-U-Net hybrid deep learning recognition model is constructed, and training is completed through an adaptive optimization strategy until the model evaluation metrics reach the preset standards;

[0033] The intelligent identification module for land use changes is used to preprocess multi-source data of the area to be investigated, extract deep features, and then input the data into the trained model to automatically identify the areas, scope, area, and type of land use changes.

[0034] Automatic evidence generation module and blockchain storage: Used to automatically generate structured and standardized evidence packages according to the identification results and the "Technical Regulations for Land Change Survey";

[0035] The AI-powered two-level review module is used to upload the evidence package to the consortium blockchain for encrypted storage and to generate a timestamp. It verifies the completeness, compliance, and authenticity of the materials through AI and automated review, as well as secondary manual review, and outputs the final review result.

[0036] Beneficial Effects: By unifying the coordinate system, data format, and denoising enhancement processing, the inconsistency of multi-source data is resolved, reducing human error and ensuring data consistency in large-scale surveys. A hybrid model integrating CNN, Transformer, and U-Net, combined with an attention mechanism and feature pyramid structure, improves the accuracy of change area identification and boundary segmentation, while automated processes significantly reduce the cost of manual intervention. Structured materials are automatically generated according to technical procedures, linked to legal text databases and dynamic QR codes, ensuring the compliance and traceability of evidence content and shortening the material preparation cycle. Blockchain consortium chain evidence storage and SM2 / SM3 encryption technology are used, with core data and full-process logs stored separately to ensure data authenticity and security. AI-automated review combined with secondary manual review verifies completeness, compliance, and authenticity, reducing the risk of errors and omissions and meeting the statutory requirements for natural resource management. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the method. Detailed Implementation

[0038] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] Example 1

[0040] Special Implementation of Survey on Changes in Cultivated Land from Grain to Farmland

[0041] I. Implementation Steps

[0042] S1, Multi-source data acquisition

[0043] Satellite remote sensing imagery: 10-meter resolution multispectral imagery from the Gaofen-6 satellite, covering 12 key grain-producing counties in the province.

[0044] Drone aerial photography: A 1:2000 scale aerial photography was conducted on three farmland protection red line areas to obtain 4K resolution orthophotos.

[0045] GIS basic data: retrieve the farmland protection red line vector data and land use status database of the province in the 2000 National Geodetic Coordinate System.

[0046] Mobile data collection: Investigators used tablets equipped with 50-megapixel cameras to collect high-definition photos (no less than 3 photos per area) and 4K videos (duration ≥ 30 seconds) of suspected changed areas. They also recorded centimeter-level GPS location information, electronic signatures, and on-site parameters such as soil moisture and crop type. Two remote areas used offline data collection mode, and the data was automatically synchronized after the network was restored.

[0047] S2, Historical Data Preprocessing

[0048] We retrieved the farmland baseline data for the first quarter of 2024 and used a combined Gaussian filtering and termination foil denoising algorithm to remove salt-and-pepper noise from the images.

[0049] Using a seven-parameter coordinate transformation model, all data were uniformly converted to the 2000 National Geodetic Coordinate System.

[0050] The data format is standardized to GeoTIFF image format and JSON attribute format.

[0051] Data augmentation: Randomly flip the training samples horizontally, rotate them by 15°-30°, adjust the brightness and contrast by ±15%, and apply a slight Gaussian blur to expand the sample size by 3 times.

[0052] Sample labeling: According to the "Classification of Current Land Use" GB / T21010-2017, 15 types of change scenarios such as cultivated land, forest land, and orchard are labeled, and the labeling accuracy is controlled within 0.08 meters.

[0053] S3, Deep Feature Extraction

[0054] A 6-layer CNN convolutional neural network was used to extract farmland texture (granularity, density), spectral (vegetation band reflectance) and spatial location features from the input standardized 3-channel image.

[0055] The CBAM attention mechanism is introduced to assign a weight of 1.8 times to the cultivated land texture channel, suppressing interference from non-cultivated land features, and capturing the association information between cultivated land and surrounding land types through spatial convolution.

[0056] A four-layer transposed convolutional upsampling method is adopted, combined with the FPN feature pyramid structure, to fuse features at scales of 256×256 and 512×512. Feature dimensionality reduction and fusion are completed through 3×3 convolutional kernels to output high-level feature maps.

[0057] S4, Hybrid Model Training

[0058] Construct a Transformer-U-Net hybrid model: a 6-layer Transformer encoder (including a multi-head attention mechanism) + a U-Net decoder, which splices high-level and low-level features through skip connections.

[0059] Dataset partitioning: The dataset is divided into a training set (12,000 samples), a validation set (3,000 samples), and a test set (1,500 samples) in a 7:2:1 ratio.

[0060] Loss function: The Dice loss function and the cross-entropy loss function are weighted and fused (weight ratio 1:1.5). Training is stopped when the accuracy, recall and precision reach 95% and 94% respectively.

[0061] S5, Change Investigation Intelligent Recognition

[0062] After preprocessing and deep feature extraction, the data to be investigated is input into the trained model, which outputs a probability map of the changed area and classification results.

[0063] Morphological opening operations are used to remove noise points smaller than 10 square meters, and polygon fitting is used to optimize the boundary of the changed region to calculate the changed area.

[0064] Generate structured data (JSON format): including the boundary coordinates of 8 changed areas, total area of ​​5.2 hectares, change type (3.7 hectares of farmland converted to forest land and 1.5 hectares of farmland converted to orchard land), model recognition probability value (0.96-0.99), and change time prediction (within the last 3 months).

[0065] S6. Automatic generation of supporting documents

[0066] The system automatically generates the "Investigation and Evidence Report on the Change of Farmland to Non-Grain Use" by calling the change comparison map template (overlay of historical farmland status and current changed image) and the change details report template.

[0067] The legal provisions database is matched with Article 37 of the Land Administration Law and Section 6.2.3 of the Technical Regulations for Land Change Survey to generate a list of legal bases.

[0068] A QR code is dynamically generated, containing a unique identifier for the evidence and a blockchain storage address. Scanning the code will redirect you to the evidence query page. The data traceability file records the processing time (accurate to the minute), operator, equipment number, and parameter settings for each step.

[0069] S7, Blockchain Evidence Storage and AI-based Dual-Level Audit

[0070] The evidence package is uploaded to the consortium blockchain. The consortium blockchain nodes include the Provincial Department of Natural Resources, third-party investigation agencies, and provincial notary offices. Node access is verified and read / write permissions are assigned.

[0071] It adopts the PBFT consensus mechanism, transmits data through SM2 asymmetric encryption, generates data hash values ​​using the SM3 algorithm, stores the core evidence report on the main chain, and stores the full process log on the side chain.

[0072] The AI ​​review model verifies the completeness of the materials (complete photos, videos, and location information), compliance (matching legal basis), and authenticity (consistency between images and location coordinates). Once the review is passed, the materials are directly archived, and no manual review is required if there are no abnormalities.

[0073] II. Implementation Results

[0074] The accuracy rate of area change identification reached 95.3%, with a boundary error of ≤0.5 meters, meeting the accuracy requirements of provincial verification.

[0075] The time for generating evidence materials has been reduced from 3 days in the traditional manual process to 4 hours, increasing work efficiency by 87%.

[0076] The blockchain-based evidence data is traceable and verified through the provincial-level regulatory platform, and all supporting materials pass the verification process in one go.

[0077] Example 2

[0078] Special Investigation into Illegal Expansion and Alteration of Construction Land

[0079] I. Implementation Steps

[0080] S1, Multi-source data acquisition

[0081] Satellite remote sensing imagery: 5-meter resolution imagery from Sentinel-2 satellite, covering 800 square kilometers of the urban planning area.

[0082] Drone aerial photography: Conduct 1:1000 scale aerial photography of areas with high incidence of violations, such as industrial parks and urban-rural fringe areas, to obtain 4K resolution oblique photography images.

[0083] GIS basic data: retrieve vector data of urban master plan, construction land approval red line data, and topographic data of the 2000 national geodetic coordinate system.

[0084] Mobile data collection: Using a professional surveying tablet, high-definition photos (including four directions: front, back, left, and right) and 4K videos (including panoramic shots of the surrounding environment) of suspected violation areas are collected, and centimeter-level GPS positioning, electronic signatures of investigators, and environmental parameters such as terrain slope and distance to surrounding roads are recorded.

[0085] S2, Historical Data Preprocessing

[0086] The benchmark data for construction land in the previous month of 2024 was retrieved, and the noise in the aerial images was processed using a combined Gaussian filtering and termination aluminum foil denoising algorithm.

[0087] A unified coordinate system is achieved through a seven-parameter coordinate transformation model, with the format standardized to GeoTIFF and JSON formats.

[0088] Data augmentation: Randomly flip the training samples horizontally / vertically, rotate them from 15° to 45°, and adjust the brightness by ±20%, increasing the sample size by 2.5 times.

[0089] Sample labeling: According to the "Classification of Current Land Use" GB / T21010-2017, 15 categories of labels such as construction land, cultivated land, and unused land are labeled, with a labeling accuracy of 0.09 meters.

[0090] S3, Deep Feature Extraction

[0091] A 6-layer CNN convolutional neural network was used to extract the texture (building density), spectrum (difference in reflectance between concrete and soil), and spatial location features of the construction land.

[0092] The CBAM attention mechanism assigns 2.2 times the weight to the spectral channels of built-up land, enhancing the identification of built-up land features, and spatial attention captures the correlation information of the contiguous distribution of buildings.

[0093] Four layers of transposed convolution upsampling, combined with the FPN feature pyramid structure, fuse multi-scale features from 128×128 to 512×512, and complete feature fusion through 5×5 convolution kernels to output high-level features.

[0094] S4, Hybrid Model Training

[0095] We construct a hybrid model consisting of an 8-layer Transformer encoder and a U-Net decoder, using skip connections to compensate for insufficient local feature capture.

[0096] Dataset partitioning: The training set (9,000 samples), validation set (2,500 samples), and test set (1,200 samples) are allocated in a 7:2:1 ratio.

[0097] The training was terminated when the Dice and cross-entropy weighted loss functions (weight ratio 1:2) were used and the accuracy, recall, and precision were achieved.

[0098] S5, Change Investigation Intelligent Recognition

[0099] After preprocessing and feature extraction, the data to be investigated is input into the model, and morphological opening operations are used to remove noise points smaller than 15 square meters.

[0100] Polygon fitting optimization of the boundaries of illegal construction land, calculation of the changed area, and generation of XML-formatted structured data: containing 5 illegal areas with a total area of ​​3.8 hectares, all of which are construction land built without prior approval, with model identification probability values ​​of 0.97-0.99.

[0101] S6. Automatic generation of supporting documents

[0102] Use the change comparison map template (comparison between urban planning map and current image) and the violation details report template to generate the "Evidence Report on Illegal Expansion of Construction Land".

[0103] Match Article 45 of the Territorial Spatial Planning Law and Article 53 of the Land Administration Law to generate a list of legal bases.

[0104] The dynamic QR code contains a unique identifier for evidence collection and an address for evidence storage. The traceability file records the parameters of the data collection device, the version of the processing algorithm, and the information of the operator.

[0105] S7, Blockchain Evidence Storage and AI-based Dual-Level Audit

[0106] The evidence materials package is uploaded to the consortium blockchain (Municipal Natural Resources Bureau, Supervision Unit, Data Storage Center), using SM2 encryption and PBFT consensus mechanism.

[0107] AI review found a deviation of ±0.3 meters between the image and the location in one area. The image was then sent to a human reviewer, who confirmed that the deviation was due to terrain undulations. After the human reviewer added annotations, the image was approved.

[0108] The core evidence and review results are stored on the consortium blockchain and timestamps are generated for law enforcement agencies to query.

[0109] II. Implementation Results

[0110] The accuracy rate of identifying illegally constructed land was 96.1%, and the accuracy of boundary verification met the requirements of law enforcement.

[0111] The efficiency of generating evidence materials is 80% higher than that of traditional manual methods, and law enforcement agencies can complete on-site verification and handling within 3 days based on the evidence data.

[0112] Blockchain-stored data provides tamper-proof evidence for subsequent administrative penalties, eliminating the possibility of administrative review disputes.

[0113] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An AI-based method for providing evidence in land use change surveys, characterized in that, Specifically, the following steps are included: S1. Multi-source data acquisition: Simultaneously acquire satellite remote sensing images, drone aerial images, GIS basic geographic data, and mobile terminal field data of the land survey area; S2. Historical data preprocessing: retrieve historical benchmark data from the land survey, and sequentially perform noise reduction, coordinate unification, format standardization, data augmentation, and sample labeling to generate training datasets and benchmark comparison datasets; S3. Deep Feature Extraction: Based on the preprocessed data and multi-source collected data, a CNN convolutional neural network is used to extract primary features. Based on the primary features, a CBAM attention mechanism is introduced to obtain intermediate features. Based on the intermediate features, transposed convolution is used for upsampling. An FPN feature pyramid structure is introduced, and high-level features are obtained through convolution. S4. Hybrid Model Training: Using high-level features as input, construct a Transformer-U-Net hybrid deep learning recognition model, and complete the training through an adaptive optimization strategy until the model evaluation metrics reach the preset standards. S5. Intelligent identification of land use change survey: After preprocessing and extracting deep features from the multi-source data of the area to be investigated, the data is input into the trained model to automatically identify the area, scope, area and type of land use change. S6. Automatic generation of evidence materials: Based on the recognition results, a structured and standardized evidence material package is automatically generated; S7. Blockchain-based evidence storage and AI-based dual-level review: The evidence package is uploaded to the consortium blockchain for encrypted storage and a timestamp is generated. The completeness, compliance, and authenticity of the materials are verified through AI and automated review, as well as secondary manual review, and the final review result is output.

2. The method for providing evidence in an AI-based land change survey according to claim 1, characterized in that: In step S1, the data collected on-site by the mobile terminal includes high-definition photos with a resolution of 50 megapixels or higher, 4K resolution videos, centimeter-level GPS positioning information, electronic signature confirmation information of investigators, and on-site environmental parameters.

3. The method for providing evidence in an AI-based land change survey according to claim 1, characterized in that: In step S2, the preprocessing specifically includes: Denoising: A combined Gaussian filtering and termination foil denoising algorithm is used to remove salt-and-pepper noise and Gaussian noise from the image; Coordinate unification: All data are uniformly converted to the 2000 National Geodetic Coordinate System using a seven-parameter coordinate transformation model; Format standardization: Convert data of different formats into GeoTIFF image format and JSON attribute format; Data augmentation: Random flipping, rotation, brightness and contrast adjustment, and Gaussian blur are used to augment the training samples; Sample standard: Label 15 types of land change scenarios according to the standard, with a labeling accuracy of ≤0.1 meters.

4. The method for providing evidence in an AI-based land change survey according to claim 1, characterized in that: Step S3 specifically includes the following steps: S31. A 6-layer CNN convolutional neural network structure is adopted. The input is standardized 3-channel image data, and the output is a basic feature map containing texture features, spectral features and spatial location features. Among them, texture features include the granularity and texture density of land cover; spectral features include the band reflectance differences of different land types; and spatial location features include the relative coordinate relationship of land types. S32. Introducing a one-dimensional mechanism of CBAM, which is divided into channel attention and spatial attention. Channel attention is achieved by fusing global average reddening and global maximum reddening to assign 1.2-2.5 times weight to change-related feature channels, such as the built-up land spectral channel and the cultivated land texture channel, to suppress interference from irrelevant features. Spatial attention is achieved by convolution to capture spatial correlation information and combine it with temporal change features to finally obtain intermediate features. S33. Four layers of transposed convolution are used to upsample intermediate features. An FPN feature pyramid structure is introduced to fuse intermediate features at different scales, mine features in multi-scale changing regions, and perform feature dimensionality reduction and fusion through convolution kernels to finally obtain high-level features.

5. The method for providing evidence in an AI-based land change survey according to claim 1, characterized in that: In step S4, a Transformer encoder with a multi-head attention mechanism and a feedforward neural network in each layer is used as the encoder, and a U-Net decoder is used. The high-level features output by the Transformer encoder are concatenated with the low-level features extracted by the CNN through skip connections. The sigmoid activation function is used to output the probability map of the change region, and the softmax activation function is used to output the classification result of the change type. The training set, validation set and test set are divided in a 7:2:1 ratio. The Dice loss function and the cross-entropy loss function are weighted and fused as the loss function. The evaluation index includes accuracy, recall and precision.

6. The method for providing evidence in an AI-based land change survey according to claim 1, characterized in that: In step S5, the multi-source data of the area to be investigated are preprocessed and deep feature extraction is performed sequentially to obtain high-level features. The deep features are input into the trained hybrid model, and the probability of the changed area and the classification results of the change type are input. Morphological opening operation is used to remove small noise points. The boundary of the changed area is optimized by polygon fitting, the area of ​​change is calculated, and structured data containing the boundary coordinates of the changed area, area, change type, probability value of model recognition, and time node prediction are generated.

7. The method for providing evidence in an AI-based land change survey according to claim 1, characterized in that: In step S6, based on the structured results of model recognition, a preset template library and a legal provisions database are called to automatically associate relevant information and generate various types of documents; The template library includes preset change comparison chart templates and change detail report templates; it automatically generates a list of legal bases through keyword matching; the QR code for the electronic evidence report uses dynamic generation technology and includes the unique identifier of the evidence material, the evidence storage location, and query permissions. After scanning, it can jump to the blockchain evidence storage query page. The data traceability file records the time, operator, equipment information, and parameter settings of each step of the process.

8. The method for providing evidence in an AI-based land change survey according to claim 1, characterized in that: In step S7, the blockchain evidence storage adopts a main chain + side chain architecture. The main chain stores core evidence materials and evidence storage information, while the side chain stores full-process log data. The PBFT consensus mechanism is used, and the data is encrypted and transmitted and stored using the SM2 asymmetric encryption algorithm. The data hash value is generated using the SM3 algorithm. The consortium chain nodes are led by the natural resources authorities and include third-party investigation agencies, supervision units, data storage centers, and notary offices. Node access requires identity authentication and permission review, and different nodes are assigned different read and write permissions. The AI ​​review model verifies the materials according to three dimensions: completeness, compliance, and authenticity. Materials that pass the review are directly archived, while materials that fail the review or are suspected of being abnormal are pushed to manual review.

9. An AI-based evidence collection system for land change surveys, characterized in that: The system is used to implement the evidence collection method for land change survey based on AI as described in any one of claims 1-8, including: a multi-source data acquisition module, a historical data preprocessing module, a deep feature extraction module, a hybrid model training module, a change intelligent recognition module, an automatic evidence material generation module, and a blockchain evidence storage and AI dual-level review module. Multi-source data acquisition module: used to acquire satellite remote sensing images, UAV aerial images and GIS basic geographic data, and to collect high-definition photos, videos, location information, electronic signatures and environmental parameters on site through mobile terminals, supporting offline acquisition and data synchronization after network recovery; Historical data preprocessing module: used to perform noise reduction, coordinate unification and format standardization on historical benchmark data, expand training samples through data augmentation, complete sample labeling according to land change type, and generate training dataset and benchmark comparison dataset; The deep feature extraction module is used to extract primary features based on preprocessed data and multi-source collected data using a CNN convolutional neural network. Based on the primary features, a CBAM attention mechanism is introduced to obtain intermediate features. Based on the intermediate features, transposed convolution is used for upsampling. An FPN feature pyramid structure is introduced, and high-level features are obtained through convolution. Hybrid Model Training Module: Using high-level features as input, a Transformer-U-Net hybrid deep learning recognition model is constructed, and training is completed through an adaptive optimization strategy until the model evaluation metrics reach the preset standards; The intelligent identification module for land use changes is used to preprocess multi-source data of the area to be investigated, extract deep features, and then input the data into the trained model to automatically identify the areas, scope, area, and type of land use changes. Automatic evidence generation module and blockchain storage: used to automatically generate structured and standardized evidence packages based on the identification results; The AI-powered two-level review module is used to upload the evidence package to the consortium blockchain for encrypted storage and to generate a timestamp. It verifies the completeness, compliance, and authenticity of the materials through AI and automated review, as well as secondary manual review, and outputs the final review result.

Images