A land change survey verification method based on a computer vision model

Abstract

The application discloses a land change survey verification method based on a computer vision model, relates to the technical field of land change verification, and solves the problems of high artificial verification cost, low efficiency, low artificial intelligence verification land class identification accuracy, and inaccurate land class segmentation and area calculation in the prior art. The method comprises the following steps: land class identification is performed by using a plurality of AI land class identification models, and weighted voting is performed to improve the accuracy of the identification result; the area of the land class is more accurately calculated by using a world coordinate system; the boundary region with differences is re-demarcated; the misidentified region with extremely small area is filtered out in combination with the minimum threshold of the land class; a verification report is generated according to the historical land class information of the verification region, in combination with the area, land class and score of each identification region, the reliability and accuracy of land class discrimination are greatly improved, and meanwhile, the high automation and intelligence of land change survey are realized.

Description

Technical Field

[0001] This application relates to the field of land change verification technology, and in particular to a land change survey and verification method based on a computer vision model. Background Technology

[0002] With the increasing demand for land and resources management, traditional land change survey methods, such as manual verification, ground surveys, and traditional image analysis, have revealed problems such as low efficiency, high cost, and large errors. Existing image processing technologies mainly rely on manual intervention, cannot achieve full automation, and suffer from problems such as inaccurate land use identification and large spatial data transformation errors.

[0003] Currently, although some AI-based image recognition technologies are gradually being applied to geographic data processing, a large number of experienced map interpreters are still needed to perform visual interpretation and analysis of land feature classification. Furthermore, there are still problems such as low accuracy in land category identification and inaccurate land category segmentation and area calculation. Summary of the Invention

[0004] The purpose of this application is to overcome the problems of high cost and low efficiency of manual verification in the prior art, as well as the low accuracy of land category identification, inaccurate land category segmentation and area calculation in artificial intelligence verification, and to provide a land change survey and verification method based on computer vision model.

[0005] Firstly, a method for verifying land use changes based on computer vision models is provided, including:

[0006] Obtain historical land classification information, images to be processed, and corresponding shooting parameter information for the verification area;

[0007] The image to be processed is used to identify land use categories using a pre-trained AI land use identification model set to obtain multiple identification results, wherein the identification results include the outline of at least one identification region, as well as the land use category and score corresponding to the identification region;

[0008] The land category and score for each identified area are obtained by weighting multiple identification results;

[0009] The composition method is used to decompose the divergent boundary area into at least two sub-regions. Based on the historical land use information of each sub-region and the land use information of adjacent regions, the divergent identification area is redefined.

[0010] Based on the shooting parameter information of the image to be processed, its pixel coordinates are transformed to the world coordinate system, and the area of ​​each recognition region is calculated in the world coordinate system;

[0011] Determine whether each identified area is smaller than the preset threshold corresponding to the land type. If the determination result is yes, merge the identified area into a suitable adjacent area and update the identification result.

[0012] The identification results are compared with historical land use information to obtain information on land use changes, and the confidence score of land use change is calculated based on the preset land use conversion probability matrix and area change trend.

[0013] Based on the historical land use information of the verification area, a verification report is generated by combining the area, land use type and score of each identified area. The verification report includes the location of the changed area, the type of change, the land use type, confidence score and area of ​​each identified area.

[0014] Among the possible implementations are:

[0015] Establish a land use verification rule base, which includes the characteristics of various land features, spectral properties, texture features, and land use conversion constraints;

[0016] The land category verification rule base is used as the basis for judgment in the AI ​​land category identification model group.

[0017] In some possible implementations, the AI ​​land cover identification model group includes multiple AI land cover identification models with different model structures or different training methods, wherein the training methods of the AI ​​land cover identification models include:

[0018] Choose the basic framework of the model;

[0019] Acquire land images containing different land cover types;

[0020] The land images are labeled with land cover types to obtain a labeled dataset, and the labeled dataset is reviewed and corrected.

[0021] The labeled dataset is divided into training, validation and test sets according to a preset ratio. The selected model is trained using a preset training method to obtain a trained AI land use identification model.

[0022] Some possible implementations also include image enhancement of the acquired land image to achieve image augmentation, the image enhancement including: geometric transformation, illumination adjustment, adding noise and adding color jitter.

[0023] In some possible implementations, the training method for the AI ​​land category identification model further includes:

[0024] The trained AI land use identification model was validated using a validation set;

[0025] The accuracy of various AI land type recognition models in identifying different land types was statistically analyzed.

[0026] The weight of each AI land use identification model for different land use types is determined based on the accuracy of each model in identifying different land use types.

[0027] In some possible implementations, the training method for the AI ​​land category identification model further includes:

[0028] Optimize the labeled dataset based on the test results;

[0029] The trained AI land cover identification model was optimized using the optimized labeled dataset.

[0030] Some possible implementations also include: simplifying the boundaries of the identified regions in the recognition results.

[0031] In some possible implementations, based on the image capture parameters, the pixel coordinates of the image to be processed are transformed to the world coordinate system, and the area of ​​each recognition region is calculated in the world coordinate system, including:

[0032] Acquire the shooting parameter information of the image to be processed, including focal length, pixels, camera position and shooting posture information;

[0033] Calculate the transformation parameters from pixel coordinates to the world coordinate system based on the shooting parameter information;

[0034] The pixel coordinates of the image to be processed are transformed into world coordinates according to the transformation parameters.

[0035] The area of ​​each recognition region in the image to be processed is calculated based on world coordinates.

[0036] Some possible implementations also include: verifying the recognition result based on the area features of the recognition region, specifically including:

[0037] Establish a statistical feature database of area for each land type, which includes the typical area range, shape complexity index, and area ratio of different land types with surrounding land types.

[0038] Features are extracted from the area of ​​each identified region, including: area size, perimeter-to-area ratio, and area ratio with surrounding land types;

[0039] The features are matched with an area statistical feature database to obtain an area rationality score, which serves as the basis for land category verification.

[0040] Specifically, in the land use classification process, the area rationality score is calculated by analyzing the area ratio between the target land use type and surrounding land use types, and matching its features with a pre-constructed feature library of typical land use type combinations. The specific method is as follows: First, the number of pixels of each classification result in the current image is counted and converted into a normalized area ratio to form an area feature vector; then, the cosine similarity of this vector with the reference vector in the feature library is calculated to measure the degree of matching of their distribution patterns; several reference items with the highest matching degree are selected and weighted to obtain the overall matching score; finally, the matching score is linearly mapped to the [0,1] interval to generate the area rationality score, which serves as an auxiliary indicator for land use classification.

[0041] For example, in a certain image segmentation result, the area proportions of buildings, roads, farmland, trees, construction sites, water bodies, and hydraulic structures are 35%, 20%, 10%, 15%, 10%, and 10%, respectively. After constructing the corresponding area feature vectors, they are compared with the reference vectors of "typical urban residential areas" and "typical farmland areas" in the feature library. It is found that the cosine similarity with "urban residential areas" is 0.98, while the similarity with "farmland areas" is only 0.65. Taking the first optimal match and combining it with the matching degree range [0.5, 1.0] set in the feature library, the final area rationality score is calculated to be 0.96, indicating that the land type combination structure of this area is highly consistent with that of urban residential areas and has high spatial rationality.

[0042] In some possible implementations, an area adjustment factor is introduced during the weighted calculation of multiple recognition results. The specific calculation formula is as follows:

[0043] S(c) = ∑(w_i × v_i(c)) × A_f(c)

[0044] Where S(c) is the score of land category c, w_i is the weight of the i-th AI land category identification model, v_i(c) is the voting result of the i-th AI land category identification model for land category c, and A_f(c) is the area adjustment factor for land category c. A_f(c) is calculated based on the reasonableness of the area of ​​land category c. When the area of ​​the identified land category c falls within the reasonable area range of land category c, A_f(c) takes a larger value; otherwise, the weight is reduced (i.e., the value of A_f(c) is reduced). Through the weighted voting mechanism and the area adjustment factor, the reliability and accuracy of land category identification are greatly improved.

[0045] This application has the following beneficial effects:

[0046] 1. This application achieves high-precision identification of land types in images captured by handheld devices and aerial images by combining the identification results of multiple models with accurate land type model data through the joint operation of multiple models. This overcomes the problems of low labor cost and low efficiency in traditional manual identification methods. At the same time, the weighted voting mechanism greatly improves the reliability and accuracy of land type identification.

[0047] 2. This application breaks through the limitations of traditional two-dimensional pixel area calculation and innovatively converts two-dimensional images into real-world coordinates through spatial mapping, which significantly improves the accuracy of area calculation and is particularly suitable for complex terrain areas such as mountains and hills;

[0048] 3. This application constructs a complete technical chain from identification to measurement and verification. Through steps such as area feature verification, boundary optimization and area threshold filtering, it realizes a high degree of automation and intelligence in land change surveys. Furthermore, this application is applicable to images and aerial image data captured by various handheld devices and can adaptively process them according to different terrain features and land type distribution characteristics, thus having broad application prospects. Attached Figure Description

[0049] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0050] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of a land change survey and verification method based on a computer vision model, according to an embodiment of this application. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] like Figure 1 As shown in the embodiments of this application, a land change survey verification method based on a computer vision model includes:

[0055] S100. Obtain historical land use information of the verification area (including land use information obtained from third parties and from historical identification results), the image to be processed, and the corresponding shooting parameter information.

[0056] If it is necessary to verify the land use change data of a certain area, it is necessary to use handheld shooting equipment or drone aerial photography equipment to take images of the area as images to be processed, and to collect the shooting parameter information at the time of shooting. The shooting parameter information includes focal length, pixels, camera position and shooting posture information. The purpose of obtaining the shooting parameter information is to prepare for the subsequent coordinate system transformation. In addition, in order to identify land use changes, it is also necessary to obtain the historical land use information of the area, so as to determine the change status by comparing the current land use information with the historical land use information. The historical land use information can be the historical model recognition results or the land use data of the area to be verified provided by a reliable third party.

[0057] S200. Use a pre-trained AI land type recognition model group to perform land type recognition on the image to be processed to obtain multiple recognition results, wherein the recognition results include the outline of at least one recognition area, as well as the land type and score corresponding to the recognition area.

[0058] Specifically, the pre-trained AI land use identification model group includes multiple AI land use identification models with different model structures or different training methods. Among them, the model structure can be an improved version of U-Net (which introduces an attention mechanism on the basis of U-Net), PSPNet, and DeepLabV3+, etc., and the training methods include supervised learning, transfer learning, and reinforcement learning, etc. The specific training process is as follows:

[0059] First, you need to select the basic framework of the model, for example, choose PSPNet as the model framework structure.

[0060] To prepare the dataset for multi-model training, a large number of land images are needed, such as evidence images obtained from the land cloud platform. It is essential to ensure that the samples cover different land use categories, different seasons, and different geographical regions. The selected images must have clear land feature characteristics and high resolution to guarantee the quality of subsequent annotation.

[0061] Next, the acquired images are divided into land use categories according to land features: water area and water conservancy facilities, roads, cultivated land (including paddy fields, dry land, and irrigated land), forest land, garden land, construction sites, buildings and their ancillary land, etc. The labelme tool is used to annotate the segmented images to obtain an annotated dataset. It should be noted that a unified annotation standard should be followed during the annotation process.

[0062] To ensure annotation quality, a quality inspection mechanism is established, employing random sampling or full review. Any annotation errors are recorded and returned for correction, ensuring that the sample covers different land cover categories, seasons, and geographical regions. Selected imagery must possess clear ground feature characteristics and high resolution to guarantee the annotation quality within the dataset.

[0063] The labeled dataset constructed in the above steps is divided into training set, validation set and test set according to the ratio, for example: the ratio of training set: validation set: test set is 70%:15%:15%, and it is ensured that the distribution of various land features is balanced in each set.

[0064] In one optional embodiment, to expand the dataset, the acquired land images are augmented in the following way to enhance the model's adaptability to different shooting conditions:

[0065] Geometric transformations (rotation, flipping, scaling, etc.);

[0066] Lighting adjustment (brightness, contrast changes, etc.);

[0067] Add noise, color jitter, etc.

[0068] During the model training phase, the Cross-Entropy Loss function is used as the main optimization objective; Dice Loss or IoU Loss is introduced to improve the segmentation effect of small-class land features; pre-trained weights (e.g., ImageNet) are used for transfer learning; and an early stopping mechanism and a learning rate decay strategy are set to prevent overfitting.

[0069] During the model validation phase, key performance indicators are monitored on the validation set, including: mean intersection-over-union ratio (mIoU), overall classification accuracy (OA), precision for each category, recall, and F1 score; model hyperparameters are adjusted and network structure is optimized based on the validation results.

[0070] During the model testing phase, the final model performance was evaluated on an independent test set. The model outputs a confusion matrix, a visual segmentation map, and a score. The visual segmentation map marks the outline of the identified regions with segmentation lines and labels the land type and score of each region. The confusion matrix provides a detailed statistical analysis of the correspondence between the model's classification prediction for each pixel in the image and the true label. The score is based on a percentage system, with higher scores indicating higher reliability in land type identification. The performance differences between the single model and the ensemble model were compared to verify the effectiveness of the ensemble strategy.

[0071] To construct a feedback optimization mechanism for the model, the following steps are also included: optimizing the labeled dataset based on the test results, i.e., analyzing misclassified samples and their causes (such as the influence of lighting, similarity of land features, etc.); and feeding back the results to the data collection and labeling stages to guide subsequent sample supplementation and optimization. This step aims to build a structurally sound and stable land cover identification system, ensuring high accuracy and strong generalization ability of the model in land cover identification through multi-model integration and rigorous evaluation and feedback mechanisms.

[0072] After training each model in the AI ​​land cover identification model group, it is necessary to statistically analyze the test results of each AI land cover identification model on the validation set. The statistical data includes the identification accuracy. By analyzing the identification accuracy of each AI land cover identification model, the performance of each AI land cover identification model is judged. Based on the performance of the AI ​​land cover identification model, the weights are set in step S300 for weighted calculation. For AI land cover identification models with superior performance, the weight values ​​are set relatively high, and for AI land cover identification models with poor performance, the weight values ​​are set relatively low.

[0073] For example, the weight w_i of the i-th AI land category identification model can be expressed as:

[0074]

[0075] Where w_i represents the weight of the i-th AI land type identification model, A i Let A be the validation set accuracy of the i-th AI land type identification model. j Let be the validation set accuracy of the j-th AI land use identification model, where j = 1, 2, 3...N, and N is the total number of AI land use identification models.

[0076] Suppose there are three models with accuracies of 90%, 85%, and 80% on the validation set, then:

[0077] w_1 = 0.9 / (0.9 + 0.85 + 0.8) = 0.35

[0078] w_2 = 0.85 / (0.9 + 0.85 + 0.8) = 0.33

[0079] w_3=0.8 / (0.9+0.85+0.8)=0.31.

[0080] The image to be processed obtained in step S100 is input into a pre-trained AI land use identification model group. Each AI land use identification model in the AI ​​land use identification model group is used to identify the land use of the image to be processed, thereby obtaining multiple identification results. Among them, due to the differences in model structure and training method of each AI land use identification model, the identification results will also have certain differences.

[0081] AI land cover recognition models classify each pixel in an image. Therefore, the contour points in the recognition results of AI land cover recognition models are relatively fine, with many contour points, and the resulting contour lines may be relatively thick and not clear enough. In order to make the contours in the recognition results clearer, the contours of the recognition regions in the recognition results are simplified. Specifically, the Douglas-Peucker Algorithm is used. The principle of this algorithm is: recursively find the point farthest from the starting line segment. If the distance is greater than ε, the point is retained and the segment is further simplified; otherwise, it is discarded. The simplified contour point set significantly reduces the number of points but retains the main shape features, thereby making the contours more regular and clear through boundary simplification.

[0082] In a further embodiment, the method also includes: establishing a land category verification rule base. Specifically, based on the land and resources survey specifications and land category identification standards, a land category verification rule base is constructed that includes various land feature characteristics, spectral characteristics, texture characteristics, and land category transformation constraints, serving as the basis for multi-model judgment.

[0083] S300: Weighted calculation is performed on multiple identification results to obtain the land category and score for each identified area.

[0084] First, it is necessary to calculate the area of ​​each recognition region in each recognition result in step S200. In this embodiment, the method in step S500 is used for calculation, as follows:

[0085] Step S1: Obtain the shooting parameter information of the image to be processed;

[0086] Specifically, the shooting parameters of images captured by handheld devices or aerial images are read out. The shooting parameter information includes focal length, pixels, camera position, and shooting posture information.

[0087] Step S2: Calculate the transformation parameters from pixel coordinates to the world coordinate system based on the shooting parameter information of the image to be processed;

[0088] Specifically, the image capture parameters (including camera parameters, pose parameters, image information, etc.) and the contour coordinates of the recognition area in the recognition results of each image are calculated to obtain the transformation parameters from the pixel coordinates of the two-dimensional image to the world coordinate system.

[0089] Step S3: Obtain the world coordinates of the image based on the transformation parameters from the pixel coordinates of the two-dimensional image to the world coordinate system.

[0090] Specifically, the transformation parameters from the pixel coordinates of the 2D image to the world coordinate system are calculated along with the contour pixel coordinates in the image recognition results to obtain the world coordinates corresponding to the contour range (i.e., the recognition area).

[0091] Step S4: Calculate the specific area of ​​each landform in the image based on the world coordinates of the image to be processed, that is, calculate the specific area of ​​each recognition region.

[0092] Specifically, the world coordinate latitude and longitude values ​​of the image to be calculated are converted into the corresponding projected coordinate reference system, and the area of ​​each recognition region is calculated according to the point coordinate values, so as to calculate the specific area of ​​each recognition region very accurately, that is, the area of ​​each type of land in the recognition result.

[0093] For example, during the camera calibration phase, the camera's intrinsic parameter matrix and radial distortion parameters can be estimated by photographing a standard checkerboard pattern containing known three-dimensional geometry. Next, using the correspondence between the world coordinate system (containing at least three points) and the image pixel coordinates, the camera's extrinsic parameter matrix is ​​calculated, thus establishing a transformation relationship from the image pixel coordinate system to the camera coordinate system. Once the connection between the image pixel coordinates and the camera coordinates is established, the process of light being projected from the three-dimensional world coordinate system onto the two-dimensional imaging plane can be simulated by applying a pinhole camera model. The pinhole model assumes that light passes through a small hole in the center of the camera in a straight line, projecting onto the imaging plane to form an inverted and reduced image. Using this model, the boundary of the segmented target area in the image pixel coordinate system can be converted into a three-dimensional point cloud in the camera coordinate system. The next step is to map the three-dimensional point cloud in the camera coordinate system to a unified world coordinate system through rigid body transformation. In this way, the actual three-dimensional spatial extent of the target area can be calculated in the world coordinate system; for two-dimensional land cover surveys, only the projected area with uniform ground elevation needs to be considered.

[0094] To calculate the actual area, the boundary points of the target area in the world coordinate system are processed, and their projected area on the ground is calculated using simple geometric algorithms (such as the convex hull algorithm or the line / curve integral method). This completes the conversion and calculation from image segmentation results to the actual physical area of ​​the target region, thus providing accurate quantitative data support for farmland surveys and evidence collection.

[0095] For each identified region, not only are the identification results from all models in the AI ​​land use identification model group collected and weighted for voting, but the precisely calculated land use area is also used as an important constraint. The voting formula is:

[0096] S(c) = ∑(w_i × v_i(c)) × A_f(c)

[0097] Where S(c) is the score of land category c, w_i is the weight of the i-th model set according to model performance in step S200, v_i(c) is the voting result of the i-th model for land category c (0 or 1, where 0 represents that the land category identified by the i-th model is not land category c, and 1 represents that the land category identified by the i-th model is land category c), and A_f(c) is the area adjustment factor for land category c, which is calculated based on the reasonableness of the area of ​​the land category. When the identified area falls within the reasonable area range of the land category, A_f(c) takes the larger value.

[0098] The formula for calculating the area adjustment factor is:

[0099]

[0100] Where A(c) is the identified area of ​​land class c in the current identified region, A min (c) represents the minimum reasonable area for land category c, A max (c) represents the maximum reasonable area for land type c, A nom (c) is the standard reasonable area of ​​land category c (i.e., the ideal value, which can be the median or the historical average).

[0101] S400. Using the composition method, the boundary area with discrepancies is decomposed into at least two sub-regions. Based on the historical land use information of each sub-region and the land use information of adjacent regions, the discrepancies are redefined to identify the area with discrepancies.

[0102] Specifically, for boundary areas where different AI land use identification models diverge, a compositional approach is used for fine-grained analysis. This decomposes the boundary area into smaller judgment units, comprehensively considering the land use attributes, terrain features, historical change information, and area rationality of adjacent areas. When boundary adjustments cause abnormal changes in the area of ​​a certain land use type, the approach tends to maintain the integrity and area rationality of the land use type, avoiding fragmented areas that do not conform to reality. For example, if a very small area suddenly appears, and this very small area clearly does not conform to the normal area of ​​that land use type, then this very small identified area is merged into a more likely adjacent area to maintain the integrity and area rationality of the land use type. Another example: if a divergent area is surrounded by woodland with significant topographic relief and historical data indicating it was once woodland, it tends to be classified as woodland. Simultaneously, if a divergent area is adjacent to farmland, has flat terrain, and its spectral characteristics are similar to farmland, it is classified as farmland.

[0103] S500: Based on the shooting parameter information of the image to be processed, transform its pixel coordinates to the world coordinate system and calculate the area of ​​each recognition region in the world coordinate system. Specifically, this includes the following steps:

[0104] Step S1: Obtain the shooting parameter information of the image to be processed; specifically, read out the shooting parameters of the image captured by the handheld device or the aerial image. The shooting parameter information includes focal length, pixels, camera position and shooting attitude information, that is, the camera's intrinsic parameters: focal length and pixels; and the camera's extrinsic parameters: camera position and shooting attitude information (shooting attitude information includes pitch angle, roll angle and yaw angle).

[0105] Step S2: Calculate the transformation parameters from pixel coordinates to the world coordinate system based on the shooting parameter information of the image to be processed; specifically, calculate the shooting parameter information of the image to be processed (including camera parameters, pose parameters, image information, etc.) and the contour coordinates of the recognition area in the recognition result of each image to be processed to obtain the transformation parameters from pixel coordinates of the two-dimensional image to the world coordinate system.

[0106] Step S3: Transform the pixel coordinates of the image to be processed to the corresponding world coordinates according to the transformation parameters in step S2; specifically, calculate the transformation parameters from the pixel coordinates of the two-dimensional image to the world coordinate system and the contour pixel coordinates in the recognition result of the image to be processed to obtain the world coordinates corresponding to the contour range (i.e., the recognition area).

[0107] Step S4: Calculate the specific area of ​​each recognition region in the image based on the world coordinates of the image to be processed. Specifically, the latitude and longitude values ​​of the world coordinates of the image to be processed are converted into the corresponding projected coordinate reference system, and the area of ​​each recognition region is calculated according to the point coordinate values, so as to calculate the specific area of ​​each recognition region very accurately.

[0108] In a further embodiment, the method further includes: verifying the recognition result based on the area features of the recognition region, specifically including:

[0109] Establish a statistical feature database of area for each land type, which includes the typical area range, shape complexity index, and area ratio of different land types with surrounding land types.

[0110] Features are extracted from the area of ​​each identified region, including: area size, perimeter-to-area ratio (the perimeter-to-area ratio is used to represent the shape complexity index), and area ratio with surrounding land types.

[0111] The features are matched with an area statistical feature database to obtain an area rationality score, which serves as the basis for land category verification.

[0112] S600. Determine whether each identified area is less than the preset threshold corresponding to the land type. If the determination result is yes, then merge the identified area into a suitable adjacent area and update the identification result. That is, delete the identified areas and land type information in the identification result that are less than the preset threshold corresponding to the land type, and merge the area into a suitable adjacent area.

[0113] Specifically, in step S500, the specific area of ​​the identified region corresponding to the land type can be obtained. Based on the minimum area requirements for different land types in the land and resources survey specifications, the identification results are filtered. When the area of ​​the identified region for a certain land type is less than the specified threshold, the identified region is merged into the most suitable adjacent land type according to the surrounding land types, avoiding the appearance of small-area patches that do not meet the specifications, thereby improving the standardization of the verification results.

[0114] S700. Compare the identification results with historical land use information to obtain information on land use changes, and calculate the confidence score of land use change based on the preset land use conversion probability matrix and area change trend.

[0115] Specifically, the current identification result is compared with the historical land use information obtained in step S100 to detect land use changes, and the rationality of the change is judged based on the preset land use conversion probability matrix and area change trend. Suspicious change areas are marked and provided with confidence scores for manual review.

[0116] It should be noted that the confidence score is used to measure the credibility of land use changes in the identification results, and is calculated by comprehensively considering factors such as land use conversion probability, area change trend, and spatial consistency. The specific rules are as follows: First, based on the preset land use conversion probability matrix, the rationality of the conversion between the current land use and historical land use is judged, and corresponding matching weights are assigned; second, the area change ratio of the changed area is analyzed. If the change is within a reasonable fluctuation range (e.g., ±10%), the confidence score is increased; otherwise, it is appropriately decreased; finally, the spatial distribution characteristics of surrounding land use and the area rationality score are combined to evaluate the consistency of the change in geographic semantics. After weighted fusion of various indicators, a confidence score between 0 and 1 is output, where a score close to 1 indicates a highly credible change, and areas below a set threshold (e.g., 0.6) are marked as suspicious change areas, and manual review is recommended.

[0117] S800. Based on the historical land use information of the verification area, and combined with the area, land use type and score of each identified area, a verification report is generated. The verification report includes the location of the changed area, the type of change, and the land use type, confidence score and area of ​​each identified area.

[0118] Specifically, based on the historical land use information of the verification area, combined with the area, land use type and score of each identified area, a detailed report is automatically generated, including the location of the changed area, the type of change, the confidence score, the land use type and area of ​​each identified area, and the basis for verification. Targeted on-site verification suggestions are also proposed for areas with low confidence, in order to improve the overall verification efficiency and accuracy.

[0119] This embodiment fully demonstrates the technical implementation path of the verification rules and voting mechanism module. Through multi-model collaborative judgment, composition method optimization and time series data analysis, efficient and automated verification of land change surveys is achieved, while retaining necessary manual intervention mechanisms to ensure the reliability and practicality of the verification results.

[0120] The above are merely preferred embodiments of this application; however, the scope of protection of this application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in this application, based on the technical solution and its improved concept, should be covered within the scope of protection of this application.

Claims

1. A method for verifying land use changes based on computer vision models, characterized in that, include: Obtain historical land classification information, images to be processed, and corresponding shooting parameter information for the verification area; The image to be processed is used to identify land use categories using a pre-trained AI land use identification model set to obtain multiple identification results, wherein the identification results include the outline of at least one identification region, as well as the land use category and score corresponding to the identification region; The land category and score for each identified area are obtained by weighting multiple identification results; The composition method is used to decompose the divergent boundary area into at least two sub-regions. Based on the historical land use information of each sub-region and the land use information of adjacent regions, the divergent identification area is redefined. Based on the shooting parameter information of the image to be processed, its pixel coordinates are transformed to the world coordinate system, and the area of ​​each recognition region is calculated in the world coordinate system; Determine whether each identified area is smaller than the preset threshold corresponding to the land type. If the determination result is yes, merge the identified area into a suitable adjacent area and update the identification result. The identification results are compared with historical land use information to obtain information on land use changes, and the confidence score of land use change is calculated based on the preset land use conversion probability matrix and area change trend. Based on the historical land use information of the verification area, a verification report is generated by combining the area, land use type and score of each identified area. The verification report includes the location of the changed area, the type of change, the land use type, confidence score and area of ​​each identified area. It also includes: verifying the recognition results based on the area features of the recognition region, specifically including: Establish a statistical feature database of area for each land type, which includes the typical area range, shape complexity index, and area ratio of different land types with surrounding land types. Features are extracted from the area of ​​each identified region, including: area size, perimeter-to-area ratio, and area ratio with surrounding land types; The features are matched with an area statistical feature library to obtain an area rationality score, which serves as the basis for land category verification. Specifically, during land category identification, the area rationality score is calculated by analyzing the area ratio between the target land category and surrounding land categories, and matching its features with a pre-constructed typical land category combination feature library. First, the number of pixels for each category in the current image is counted and converted into a normalized area proportion to form an area feature vector. Then, the cosine similarity of this vector with reference vectors in the feature library is calculated to measure the degree of matching of their distribution patterns. Several reference items with the highest matching degree are selected and weighted to obtain the overall matching score. Finally, the matching score is linearly mapped to the [0,1] interval to generate an area rationality score, which serves as an auxiliary indicator for land category determination. In the process of weighting multiple recognition results, an area adjustment factor is introduced, and the specific calculation formula is as follows: S(c) = ∑(w_i × v_i(c)) × A_f(c) Where S(c) is the score of land category c, w_i is the weight of the i-th AI land category identification model, v_i(c) is the voting result of the i-th AI land category identification model for land category c, and A_f(c) is the area adjustment factor of land category c.

2. The land change survey and verification method based on computer vision model according to claim 1, characterized in that, Also includes: Establish a land use verification rule base, which includes the characteristics of various land features, spectral properties, texture features, and land use conversion constraints; The land category verification rule base is used as the basis for judgment in the AI ​​land category identification model group.

3. The land change survey and verification method based on computer vision model according to claim 2, characterized in that, The AI ​​land use identification model group includes multiple AI land use identification models with different model structures or different training methods. The training methods for the AI ​​land use identification models include: Choose the basic framework of the model; Acquire land images containing different land cover types; The land images are labeled with land cover types to obtain a labeled dataset, and the labeled dataset is reviewed and corrected. The labeled dataset is divided into training, validation and test sets according to a preset ratio. The selected model is trained using a preset training method to obtain a trained AI land use identification model.

4. The land change survey and verification method based on computer vision model according to claim 3, characterized in that, It also includes image enhancement of the acquired land images to achieve image augmentation, the image enhancement including: geometric transformation, illumination adjustment, adding noise and adding color jitter.

5. The land change survey and verification method based on computer vision model according to claim 3 or 4, characterized in that, The training method for the AI ​​land category identification model also includes: The trained AI land use identification model was validated using a validation set; The accuracy of various AI land type recognition models in identifying different land types was statistically analyzed. The weight of each AI land use identification model for different land use types is determined based on the accuracy of each model in identifying different land use types.

6. The land change survey and verification method based on computer vision model according to claim 5, characterized in that, The training method for the AI ​​land category identification model also includes: Optimize the labeled dataset based on the test results; The trained AI land cover identification model was optimized using the optimized labeled dataset.

7. The land change survey and verification method based on computer vision model according to any one of claims 1-4, 6, characterized in that, Also includes: The contours of the identified regions in the recognition results are simplified.

8. The land change survey and verification method based on computer vision model according to claim 7, characterized in that, Based on the image capture parameters, the pixel coordinates are transformed to the world coordinate system, and the area of ​​each recognition region is calculated in the world coordinate system, including: Acquire the shooting parameter information of the image to be processed, including focal length, pixels, camera position and shooting posture information; Calculate the transformation parameters from pixel coordinates to the world coordinate system based on the shooting parameter information; The pixel coordinates of the image to be processed are transformed into world coordinates according to the transformation parameters. The area of ​​each recognition region in the image to be processed is calculated based on world coordinates.

Images