Land change investigation and verification method based on computer vision model

Abstract

The invention discloses a land change investigation and verification method based on a computer vision model, relates to the technical field of land change verification, and solves the problems that in the prior art, manual verification is high in cost and low in efficiency, and artificial intelligence verification land class recognition precision is low and land class segmentation and area calculation are inaccurate. The method comprises the following steps: utilizing a plurality of AI land type identification models to carry out land type identification and weighted voting so as to improve the accuracy of an identification result, utilizing a world coordinate system to more accurately calculate the area of land types, redelimiting a boundary region with divergence, and combining with a land type minimum threshold to filter out an error identification region with a very small area so as to improve the accuracy of the identification result. According to the historical land type information of the check area, the area, the land type and the score of each identification area are combined to generate the check report, the reliability and the accuracy of land type judgment are greatly improved, and meanwhile, the high automation and the high intelligence of land change investigation are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of land change investigation and verification, and particularly to a land change investigation and verification method based on a computer vision model. BACKGROUND

[0002] With the increasing demand for land and resource management, traditional land change investigation methods such as manual verification, ground investigation and traditional image analysis have exposed problems of low efficiency, high cost and large error. Existing image processing technologies mainly rely on manual intervention and cannot achieve full automation, and have problems of inaccurate land class identification and large spatial data transformation error.

[0003] At present, although some image recognition technologies based on artificial intelligence have been gradually applied to geographic data processing, a large number of experienced interpreters are needed to visually interpret and interpret the classification of geographic features, and there are problems of low accuracy of land class identification and inaccurate land class segmentation and area calculation. SUMMARY

[0004] The present application aims to overcome the problems of high cost and low efficiency of manual verification, and low accuracy of land class identification, inaccurate land class segmentation and area calculation in artificial intelligence verification in the prior art, and provides a land change investigation and verification method based on a computer vision model.

[0005] In a first aspect, a land change investigation and verification method based on a computer vision model is provided, comprising:

[0006] obtaining historical land class information of a verification area, a to-be-processed image and corresponding shooting parameter information;

[0007] using a pre-trained AI land class identification model group to identify the land class of the to-be-processed image to obtain a plurality of identification results, wherein the identification results include the contour of at least one identification region, and the land class and score corresponding to the identification region;

[0008] performing weighted calculation on the plurality of identification results to obtain the land class and score of each identification region;

[0009] using a composition method to divide the boundary region with disagreement into at least two sub-regions, and re-delineating the identification region with disagreement according to the historical land class information of each sub-region and the land class information of adjacent regions;

[0010] According to the shooting parameter information of the to-be-processed image, the pixel coordinates are converted to the world coordinate system, and the area of each identification region is calculated in the world coordinate system;

[0011] determining whether each identified region is smaller than a preset threshold corresponding to the land class, and if the result is yes, merging the identified region into a suitable adjacent region and updating the identification result;

[0012] comparing the identification result with historical land class information to obtain land class change information, and calculating a confidence score of the land class change according to a preset land class conversion probability matrix and an area change trend;

[0013] generating a verification report according to the historical land class information of the verification region, the area, the land class and the score of each identified region, wherein the verification report includes the location of the changed region, the type of the changed region, the land class, the confidence score and the area of each identified region.

[0014] In some possible implementations, the method further includes:

[0015] establishing a land class verification rule library, wherein the land class verification rule library includes features of various land objects, spectral characteristics, texture features and land class conversion constraints;

[0016] using the land class verification rule library as a basis for determining the AI land class identification model group.

[0017] In some possible implementations, the AI land class identification model group includes a plurality of AI land class identification models with different model structures or different training methods, wherein the training method of the AI land class identification model includes:

[0018] selecting a basic framework of the model;

[0019] obtaining land images containing different land object types;

[0020] annotating the land images to obtain an annotated data set, and reviewing and correcting the annotated data set;

[0021] dividing the annotated data set into a training set, a validation set and a test set according to a preset ratio, and training the selected model through a preset training method to obtain a trained AI land class identification model.

[0022] In some possible implementations, the method further includes image enhancement to achieve image expansion, wherein the image enhancement includes geometric transformation, light adjustment, noise addition and color jittering.

[0023] In some possible implementations, the training method of the AI land class identification model further includes:

[0024] verifying the trained AI land class identification model using the validation set;

[0025] statistically determining the identification accuracy of each AI land class identification model for different land classes.

[0026] According to the recognition accuracy of each AI land class recognition model on different land classes, weights of scores of the AI land class recognition model on different land classes are determined.

[0027] In some possible implementation manners, the training method of the AI land class recognition model further includes:

[0028] optimizing the labeled data set according to the test result;

[0029] optimizing training of the trained AI land class recognition model by using the optimized labeled data set.

[0030] In some possible implementation manners, the method further includes: performing boundary simplification on the contour of the identified region in the recognition result.

[0031] In some possible implementation manners, the area of each identified region is calculated in the world coordinate system by converting the pixel coordinates of the to-be-processed image to the world coordinate system according to the shooting parameter information of the to-be-processed image, and the conversion includes:

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

[0033] calculating transformation parameters from the pixel coordinates to the world coordinate system according to the shooting parameter information;

[0034] transforming the pixel coordinates of the to-be-processed image into world coordinates according to the transformation parameters;

[0035] calculating the area of each identified region in the to-be-processed image according to the world coordinates.

[0036] In some possible implementation manners, the method further includes: verifying the recognition result based on the area feature of the identified region, and specifically includes:

[0037] establishing an area statistical feature library of each land class, the area statistical feature library including a typical area range, a shape complexity index, and an area proportion relationship with a surrounding land class of different land classes;

[0038] extracting features of the area of each identified region, the features including an area size, a perimeter-to-area ratio, and an area proportion relationship with a surrounding land class;

[0039] matching the features with the area statistical feature library to obtain an area rationality score, and the area rationality score is used as a basis for land class verification.

[0040] Specifically, in the land class identification process, the area rationality score is obtained by analyzing the area proportion relationship between the target land class and the surrounding land class, and matching the features with the pre-constructed typical land class combination feature library. The specific method is as follows: first, the pixel number of each classification result in the current image is counted, and the normalized area proportion is converted to form an area feature vector; then, the vector is calculated with the reference vector in the feature library to measure the matching degree of the distribution pattern; the matching degree of the highest several reference items is selected, and the overall matching score is calculated by weighting; finally, the matching score is linearly mapped to the interval [0, 1] to generate the area rationality score as an auxiliary index for land class determination.

[0041] For example, the area proportions of buildings, roads, cultivated land, trees, construction sites, water areas and water engineering buildings in a certain image segmentation result are 35%, 20%, 10%, 15%, 10% and 10% respectively. After constructing the corresponding area feature vector, the reference vector of "typical urban residential area" and "typical farmland area" in the feature library is compared, and it is found that the cosine similarity with "urban residential area" is 0.98, and the similarity with "farmland area" is only 0.65. Taking the first optimal matching item, combined with the matching degree range [0.5, 1.0] set in the feature library, the area rationality score is finally calculated as 0.96, indicating that the land class combination structure of this area is highly consistent with the urban residential area, and has high spatial rationality.

[0042] In some possible implementations, an area adjustment factor is also introduced in the process of weighted calculation of multiple identification results, and the specific calculation formula is:

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

[0044] Where S(c) is the score of land class c, w_i is the weight of the i th AI land class identification model, v_i(c) is the voting result of the i th AI land class identification model for land class c, and A_f(c) is the area adjustment factor of land class c. A_f(c) is calculated according to the area rationality of the land class c. When the area of the identified land class c falls within the reasonable area range of the land class 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 class determination are greatly improved.

[0045] The present application has the following beneficial effects:

[0046] 1. The application realizes high-precision identification of land classes in handheld device images and aerial images by combining multi-model identification results with accurate land class model data, overcoming the problems of low labor cost and low efficiency in traditional manual identification methods, and significantly improving the reliability and accuracy of land class identification through a weighted voting mechanism.

[0047] 2. The 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, significantly improving the accuracy of area calculation, especially in complex terrain areas such as mountains and hills.

[0048] 3. The application constructs a complete technical chain from identification to calculation to verification, realizing high automation and intelligence of land change survey through area feature verification, boundary optimization, and area threshold filtering. In addition, the application is suitable for various handheld device images and aerial image data, can adaptively process according to different terrain characteristics and land class distribution characteristics, and has wide application prospects. BRIEF DESCRIPTION OF DRAWINGS

[0049] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, and are incorporated herein for purposes of explanation and are not intended to limit the application.

[0050] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on these drawings.

[0051] Figure 1 is a flowchart of the land change survey verification method based on computer vision model of the embodiments of the present application. DETAILED DESCRIPTION

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

[0053] Embodiment 1

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

[0055] S100, acquire historical land class information (including land class information obtained from a third party and in historical recognition results), an image to be processed, and corresponding shooting parameter information of a verification area.

[0056] If it is necessary to verify territorial change data of an area, an image of the area is shot by using a handheld shooting device or a drone aerial shooting device, as an image to be processed, and shooting parameter information at the time of shooting is collected, wherein the shooting parameter information includes focal length, pixels, camera position, and shooting posture information, and the purpose of acquiring the shooting parameter information is to prepare for subsequent coordinate system conversion. In addition, in order to realize recognition of territorial change, historical land class information of the area also needs to be acquired, so as to determine the change by comparing the current land class information with the historical land class information, wherein the historical land class information can be a historical model recognition result or land class data of the verification area provided by a third party with credibility.

[0057] S200, land class recognition is performed on the image to be processed by using a pre-trained AI land class recognition model group to obtain a plurality of recognition results, wherein the recognition results include the outline of at least one recognition area, and the land class and score corresponding to the recognition area.

[0058] Specifically, the pre-trained AI land class recognition model group includes AI land class recognition models with different model structures or different training methods, wherein the model structures can be improved U-Net (attention mechanism is introduced on the basis of U-Net), PSPNet, DeepLabV3+, and the like, and the training methods include supervised learning, transfer learning, reinforcement learning, and the like. The specific training process is as follows:

[0059] First, the basic framework of the model needs to be selected, for example, PSPNet is selected as the model framework structure.

[0060] In order to prepare the data set for multi-model training, a large number of land images need to be acquired, for example, evidence images are acquired from a national land cloud platform, and it is necessary to ensure that the samples cover different land class categories, different seasonal periods, and different geographical regions. The selected images need to have clear ground feature characteristics and high resolution to ensure the subsequent labeling quality.

[0061] Then, the acquired images are divided into water area, water conservancy facilities land, road, farmland (including paddy field, dry land, and watered land), forest land, garden land, construction site, building and its auxiliary land, and the like according to the ground object types. The divided images are labeled by using a labelme tool to obtain a labeled data set, and it needs to be noted that a unified labeling specification should be followed in the labeling process.

[0062] To ensure the quality of the annotation, an annotation quality inspection mechanism is established, random sampling review or comprehensive review is adopted, annotation errors are found and recorded and returned for modification, and it is ensured that the samples cover different land class categories, different seasonal periods and different geographical regions. The selected images should have clear ground feature characteristics and high resolution to ensure the annotation quality in the annotation data set.

[0063] The annotation data set constructed in the above steps is divided into a training set, a validation set and a test set according to a proportion, for example: the proportion of the training set: the validation set: the test set is 70%: 15%: 15%, and it is ensured that each type of ground object is evenly distributed in each set.

[0064] In an optional embodiment, in order to expand the data set, the following methods are used to perform image enhancement on the obtained land images to realize image expansion to enhance the adaptability of the model to different shooting conditions:

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

[0066] Light adjustment (brightness, contrast change, etc.);

[0067] Add noise, color jitter, etc.

[0068] In the model training stage, the cross-entropy loss function (Cross-Entropy Loss) is used as the main optimization target; the Dice Loss or IoU Loss is introduced to improve the segmentation effect of small class ground objects; the pre-trained weight (for example: ImageNet) is used for transfer learning; the early stopping mechanism (Early Stopping) and the learning rate decay strategy (LearningRate Scheduler) are set to prevent overfitting.

[0069] In the model verification stage, the key performance indicators are monitored on the validation set, including: average intersection over union (mIoU), overall classification accuracy (OA), precision of each class (Precision), recall (Recall) and F1 score; according to the verification result, the model hyperparameters are adjusted and the network structure is optimized.

[0070] In the model testing stage, the performance of the final model is evaluated on an independent test set; the model outputs a confusion matrix, a visual segmentation map and a score, wherein the visual segmentation map is annotated with the outline of the identified area and the ground class and score of the identified area through the segmentation line, wherein the confusion matrix details the correspondence between the classification prediction result of each pixel point in the image and the true label, the score is in percentage, and the higher the score, the higher the ground class recognition credibility; and the performance difference between the single model and the integrated model is compared to verify the effectiveness of the integration strategy.

[0071] To build the feedback optimization mechanism of the model, it also includes optimizing the labeled data set according to the results of the test, that is, analyzing the misclassified samples and their causes (such as light influence, ground object similarity, etc.); feedback to the data collection and labeling link to guide the subsequent sample supplement and optimization. This step aims to build a reasonable structure and stable performance land class recognition system, through multi-model integration and strict evaluation and feedback mechanism, to ensure the high precision and strong generalization ability of the model in land class recognition.

[0072] After the training of each model in the AI land class recognition model group is completed, the test results of each AI land class recognition model in the validation set need to be counted and analyzed. The data for statistics includes recognition accuracy. The performance of each AI land class recognition model is judged by analyzing the recognition accuracy of each AI land class recognition model, and the weight in the weighted calculation in step S300 is set according to the performance of the AI land class recognition model. The weight value set for the AI land class recognition model with superior performance is relatively high, and the weight value set for the AI land class recognition model with poor performance is relatively low.

[0073] For example, the weight w_i of the i-th AI land class recognition model can be represented as:

[0074]

[0075] where w_i is the weight of the i-th AI land class recognition model, A i is the validation set accuracy of the i-th AI land class recognition model, A j is the validation set accuracy of the j-th AI land class recognition model, where j = 1, 2, 3…N, and N is the total number of AI land class recognition models.

[0076] Suppose the accuracy of the three models on the validation set is 90%, 85% and 80% respectively, 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 the pre-trained AI land class recognition model group, and each AI land class recognition model in the AI land class recognition model group is used to perform land class recognition on the image to be processed, so that multiple recognition results can be obtained. Since the model structure and training method of each AI land class recognition model are different, the recognition results will also have certain differences.

[0081] The AI land class recognition model classifies each pixel point in the image, so the outline points of the recognition area in the recognition result of the AI land class recognition model are relatively fine, there are many outline points, the formed outline line can be relatively thick, and the outline line can not be clear enough. In order to make the outline in the recognition result more clear, the boundary of the outline of the recognition area in the recognition result is simplified, specifically: a Douglas-Peucker algorithm is used, the principle of which is: recursively finding the point farthest from the starting segment, if the distance is greater than epsilon, the point is retained and the segmentation continues to simplify, otherwise it is discarded; the simplified outline point set significantly reduces the number of points but retains the main shape features, so that the outline can be made more regular and clear through boundary simplification.

[0082] In a further embodiment, further comprising: establishing a land class verification rule library, specifically, according to the land resource survey specification and the land class discrimination standard, a land class verification rule library containing various types of ground object features, spectral characteristics, texture features and land class transformation constraint conditions is constructed as the basis for multi-model judgment.

[0083] S300, weighted calculation of the multiple recognition results to obtain the land class and score of each recognition area.

[0084] First, the area of each recognition area in each recognition result in step S200 needs to be calculated, which is calculated in this embodiment by the method in step S500, specifically as follows:

[0085] Step S1, obtaining the shooting parameter information of the to-be-processed image;

[0086] Specifically, the shooting parameters of the handheld device photographed image or the aerial photographed image are read out, and the shooting parameter information includes focal length, pixel, camera position and shooting posture information.

[0087] Step S2, calculating the transformation parameters from the pixel coordinates to the world coordinate system according to the shooting parameter information of the to-be-processed image;

[0088] Specifically, the shooting parameter information of the image (including camera parameters, posture parameters, image information, etc.) and the outline coordinates of the recognition area in the recognition result 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, obtaining the world coordinates corresponding to the image according to 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 two-dimensional image to the world coordinate system and the outline pixel coordinates in the recognition result of the image are calculated to obtain the world coordinates corresponding to the outline range (i.e. the recognition area).

[0091] Step S4, calculating the specific area of each land class in the image according to the world coordinates of the image to be processed, i.e. calculating the specific area of each recognition region;

[0092] Specifically, the world coordinate latitude and longitude value of the image to be calculated is converted into the current corresponding projection coordinate reference system, and the area of each recognition region is calculated according to the point coordinate value, so that the specific area of each recognition region, i.e. the area of each land class in the recognition result, can be calculated very accurately.

[0093] For example, in the camera calibration stage, by shooting a standard checkerboard pattern containing a known three-dimensional geometric shape, the intrinsic matrix and radial distortion parameters of the camera can be estimated. Then, by using the correspondence between the world coordinate system containing at least three points and the image pixel coordinates, the extrinsic matrix of the camera is solved, thereby establishing the conversion relationship from the image pixel coordinate system to the camera coordinate system. Once the relationship between the image pixel coordinates and the camera coordinates is established, the process of light rays from the three-dimensional world coordinate system to the two-dimensional imaging plane can be simulated by applying the pinhole camera model. The pinhole model assumes that the light rays pass through the small hole in the center of the camera along a straight line, and form an inverted and reduced image on the imaging plane. Through this model, the boundaries of the segmented target region in the image pixel coordinate system can be converted into a three-dimensional point cloud in the camera coordinate system. Next, the three-dimensional point cloud in the camera coordinate system is mapped to the unified world coordinate system through rigid body transformation. In this way, the actual three-dimensional space range of the target region can be calculated in the world coordinate system, and for two-dimensional land class investigation, only the projection area with consistent ground height needs to be concerned.

[0094] In order to calculate the actual area, the boundary points of the target region in the world coordinate system are processed, and the projection area on the ground is calculated by using simple geometric algorithms (such as convex hull algorithm or straight line / curve integral method). In this way, the conversion and calculation from the image segmentation result to the actual physical area of the target region are completed, thereby providing accurate quantitative data support for the evidence of cultivated land investigation.

[0095] For each recognition region, not only the recognition results of all models in the AI land class recognition model group are collected for weighted voting, but also the accurate land class area calculated above is used as an important constraint condition. The voting formula is:

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

[0097] Wherein, S(c) is the score of the land class c, w i is the weight of the i th model set in step S200 according to the model performance, v i (c) is the voting result of the i th model to the land class c (0 or 1, 0 represents that the land class in the i th model recognition result is not the land class c, and 1 represents that the land class in the i th model recognition result is the land class c), A f (c) is the area adjustment factor of the land class c, which is calculated according to the area rationality of the land class. When the recognition area falls within the reasonable area range of the land class, A f (c) takes a larger value.

[0098] The calculation formula of the area adjustment factor is:

[0099]

[0100] Wherein, A (c) is the recognition area of the land class c in the current recognition area, A min (c) is the minimum reasonable area of the land class c, A max (c) is the maximum reasonable area of the land class c, A nom (c) is the standard reasonable area of the land class c (i.e. ideal value, which can take intermediate value or historical average value).

[0101] S400, the boundary area with disagreement is divided into at least two sub-areas by using the composition method, and the recognition area with disagreement is re-determined according to the historical land class information of each sub-area and the land class information of the adjacent area.

[0102] Specifically, for the boundary area with disagreement between different AI land class recognition models, the composition method is used for fine-grained analysis, the boundary area is divided into smaller judgment units, and the land class properties, topographic features, historical change information and area rationality of the adjacent area are comprehensively considered. When the boundary adjustment leads to abnormal change of the area of a land class, the integrity and area rationality of the land class will be maintained, and the fragmentation area which does not conform to the actual situation will be avoided, for example: if a very small area suddenly appears, which is obviously not consistent with the regular area of the land class, the very small recognition area will be merged into the adjacent area which is more likely to maintain the integrity and area rationality of the land class; for another example: if a disagreement area is surrounded by forest land, and the topography is large and the historical data shows that it was forest land, it will be inclined to be classified as forest land; at the same time, if a disagreement area is adjacent to cultivated land, the topography is flat and the spectral characteristics are close to farmland, it will be classified as cultivated land.

[0103] S500, according to the shooting parameter information of the image to be processed, the pixel coordinates are converted to the world coordinate system, and the areas of the recognition areas are calculated in the world coordinate system, which specifically includes the following steps:

[0104] Step S1, obtaining the shooting parameter information of the image to be processed; specifically, the shooting parameters of the handheld device image or the aerial image are read out, and the shooting parameter information includes the focal length, the pixel, the camera position and the shooting posture information, i.e. the internal parameters of the camera: the focal length and the pixel; and the external parameters of the camera: the position of the camera and the shooting posture information (the shooting posture information includes the pitch angle, the roll angle and the yaw angle).

[0105] Step S2, calculating the transformation parameters from the pixel coordinates to the world coordinate system according to the shooting parameter information of the image to be processed; specifically, the shooting parameter information (including the camera parameters, the posture parameters, the image information, etc.) of the image to be processed and the contour coordinates of the identified area in the identification result of each image to be processed are calculated to obtain the transformation parameters from the pixel coordinates of the two-dimensional image to the world coordinate system.

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

[0107] Step S4, calculating the specific area of each identified area in the image according to the world coordinates of the image to be processed; specifically, the world coordinate latitude and longitude values of the image to be processed are converted into the current corresponding projection coordinate reference system, and the area of each identified area is calculated according to the point coordinate values, so that the specific area of each identified area can be calculated very accurately.

[0108] In further embodiments, it further includes verifying the identification result based on the area features of the identified area, specifically including:

[0109] Establishing an area statistical feature library of each land class, the area statistical feature library including the typical area range of different land classes, the shape complexity index and the area proportion relationship with the surrounding land classes;

[0110] Extracting the features of the area of each identified area, the features including the area size, the perimeter-area ratio (the perimeter-area ratio is used to represent the shape complexity index) and the area proportion relationship with the surrounding land classes;

[0111] Matching the features with the area statistical feature library to obtain the area rationality score, which is used as the basis for land class verification.

[0112] S600, judging whether each identified region is smaller than a preset threshold corresponding to the land class, if the result of the judgment is yes, merging the identified region into a suitable adjacent region, and updating the identification result, that is, deleting the identified region and land class information smaller than the preset threshold corresponding to the land class in the identification result, and merging the region into a suitable adjacent region.

[0113] Specifically, in step S500, the specific area of the identified region corresponding to the land class can be obtained, and the identification result is filtered based on the minimum plot area requirement for different land classes in the land resources survey specification. When the land class area of an identified region is smaller than a specified threshold, the identified region is merged into the most suitable adjacent land class according to the surrounding land class situation, avoiding the appearance of small area plots that do not meet the specification, thereby improving the standardization of the verification result.

[0114] S700, comparing the identification result with historical land class information to obtain land class change information, and calculating the confidence score of the land class change according to a preset land class conversion probability matrix and area change trend.

[0115] Specifically, the current identification result is compared with the historical land class information obtained in step S100 to detect the land class change, and the rationality of the change is judged according to the preset land class conversion probability matrix and area change trend. For suspicious change areas, a confidence score is provided for manual review.

[0116] It should be noted that the confidence score is used to measure the credibility of the land class change in the identification result, and is calculated by considering factors such as land class conversion probability, area change trend and spatial consistency. The specific rules are as follows: first, according to the preset land class conversion probability matrix, the rationality of the conversion between the current land class and the historical land class is judged, and the corresponding matching weight is given; second, the area change ratio of the changed region is analyzed, if the change is within a reasonable fluctuation range (such as ±10%), the confidence is increased, otherwise it is appropriately reduced; finally, the spatial distribution characteristics of the surrounding land class and the area rationality score are combined to evaluate the coordination of the change in geographical semantics. After weighting and fusion of various indicators, a confidence score between 0 and 1 is output, where a value close to 1 indicates a high degree of confidence in the change, and a region below a certain threshold (such as 0.6) is marked as a suspicious change area, which is recommended for manual review.

[0117] S800, generating a verification report according to the historical land class information of the verification region, combining the area, land class and score of each identified region, the verification report including the location of the changed region, the type of the change, and the land class, the confidence score and the area of each identified region.

[0118] Specifically, according to the historical land type information of the verification area, combined with the area, land type and score of each identified area, the detailed report containing the change area position, change type, confidence score, land type and area of each identified area and verification basis is automatically generated, and specific on-site verification suggestions are proposed for low-confidence areas, so as to improve the overall verification efficiency and accuracy.

[0119] In the embodiment, the technical implementation path of the verification rule and voting mechanism module is fully embodied, the efficient and automatic verification of the land change survey is realized through multi-model collaborative judgment, composition method optimization and time sequence data analysis, and at the same time, the necessary manual intervention mechanism is reserved to ensure the reliability and practicability of the verification result.

[0120] The above is only a preferred specific embodiment of the present application; however, the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application, according to the technical scheme of the present application and the improvement concept thereof, can be replaced or changed equivalently, which should be covered in the protection scope of the present application.

Claims

1. A land change investigation and verification method based on computer vision model, characterized in that: include: Obtain historical land classification information, images to be processed, and corresponding shooting parameter information of the verification area; Using a pre-trained AI land classification recognition model group to perform land classification 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, and the land classification and score corresponding to the recognition area; Perform weighted calculation on multiple identification results to obtain the land type and score of each identification area; The composition method is used to decompose the boundary area with disagreements into at least two sub-areas, and the identification area with disagreements is redefined based on the historical land classification information of each sub-area and the land classification information of the adjacent areas; According to the shooting parameter information of the image to be processed, its pixel coordinates are converted into the world coordinate system and the area of ​​each recognition area 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 judgment result is yes, merge the identified area into a suitable adjacent area and update the recognition result; Compare the identification results with historical land classification information to obtain land classification changes, and calculate the confidence score of land classification changes based on the preset land classification conversion probability matrix and area change trend; Based on the historical land classification information of the verification area, combined with the area, land classification 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 classification, confidence score and area of ​​each identified area.

2. The land change investigation and verification method based on computer vision model according to claim 1 is characterized in that: Also includes: Establishing a land classification verification rule base, wherein the land classification verification rule base includes various land feature characteristics, spectral characteristics, texture characteristics and land classification conversion constraint conditions; The land classification verification rule base is used as the judgment basis of the AI ​​land classification recognition model group.

3. The land change investigation and verification method based on computer vision model according to claim 2 is characterized in that: The AI ​​land classification recognition model group includes multiple AI land classification recognition models with different model structures or different training methods, wherein the training method of the AI ​​land classification recognition model includes: Select the basic framework of the model; Acquire land images containing different types of land features; Annotating the land image with land object types to obtain an annotated dataset, and reviewing and revising the annotated dataset; The labeled data set is divided into training set, validation set and test set according to the preset ratio, and the selected model is trained using the preset training method to obtain a trained AI land classification model.

4. The land change investigation and verification method based on computer vision model according to claim 3 is characterized in that: The method also includes performing image enhancement on the acquired land image to achieve image expansion, wherein the image enhancement includes: geometric transformation, illumination adjustment, noise addition, and color dithering.

5. The land change investigation and verification method based on computer vision model according to claim 3 or 4 is characterized in that: The training method of the AI ​​land classification model further includes: Use the validation set to verify the trained AI land classification model; Count the recognition accuracy of each AI land classification model for different land classifications; The weight of each AI land classification recognition model's score for different land classes is determined based on the recognition accuracy of each AI land classification recognition model for different land classes.

6. The land change investigation and verification method based on computer vision model according to claim 5 is characterized in that: The training method of the AI ​​land classification model further includes: Optimize the labeled dataset based on the test results; The trained AI land classification model is optimized and trained using the optimized labeled dataset.

7. The land change investigation and verification method based on computer vision model according to any one of claims 1 to 4 and 6, characterized in that: Also includes: Simplify the boundaries of the identified areas in the recognition results.

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

9. The land change investigation and verification method based on computer vision model according to claim 1 is characterized in that: Also includes: Verify the recognition results based on the area characteristics of the recognition area, including: Establishing an area statistical feature database for each land type, wherein the area statistical feature database includes the typical area range, shape complexity index and area ratio relationship of different land types with surrounding land types; Extract features of the area of ​​each identified region, including area size, perimeter-to-area ratio, and area ratio relationship with surrounding land types; The features are matched with an area statistical feature library to obtain an area rationality score, which serves as a basis for land classification verification.

10. The land change investigation and verification method based on computer vision model according to claim 1 or 8, characterized in that: The area adjustment factor is also introduced in the process of weighted calculation of multiple recognition results. The specific calculation formula is: S(c) = ∑(w_i × v_i(c)) × A_f(c) Among them, S(c) is the score of land class c, w_i is the weight of the i-th AI land class recognition model, vi_i(c) is the voting result of the i-th AI land class recognition model for land class c, and A_f(c) is the area adjustment factor of land class c.

Images