Remote sensing change detection method and device, computer device, readable storage medium and program product

Abstract

The application relates to a remote sensing change detection method and device, computer equipment, a readable storage medium and a program product. The method comprises the following steps: processing double-phase remote sensing images of a target business scene based on a preset change detection model to obtain a plurality of change polygons; determining the business attribute features corresponding to each change polygon; processing the business attribute features and local image pairs corresponding to each change polygon based on a preset probability scoring model corresponding to the target business scene to determine the acceptance probability scores of each change polygon in the target business scene; and determining the vector geometric data and business attribute features corresponding to the change polygon with an acceptance probability score greater than or equal to a first acceptance threshold value as the change detection result in the target business scene. The method can improve the change detection accuracy of remote sensing images.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology for remote sensing imagery, and in particular to a method, apparatus, computer equipment, readable storage medium, and program product for detecting remote sensing changes. Background Technology

[0002] In actual operational deployments, remote sensing change detection systems need to be applied to different operational procedures to detect changes in remote sensing images. However, different operational procedures have different acceptance requirements for change detection standards in remote sensing images, that is, different operational procedures have different institutional definitions of "reportable change events".

[0003] When related technologies are used to detect changes in remote sensing images corresponding to different business processes, customized change detection models are often required for different business processes, which makes it difficult to deploy and maintain multiple business processes. If the same change detection model is used, the accuracy of change detection for business processes will be low. Summary of the Invention

[0004] Therefore, it is necessary to address the aforementioned technical issues and provide a remote sensing change detection method, apparatus, computer equipment, readable storage medium, and program product that can improve the accuracy of change detection in remote sensing images for different business applications.

[0005] Firstly, this application provides a remote sensing change detection method, including:

[0006] Based on a preset change detection model, dual-temporal remote sensing images of the target business scenario are processed to obtain multiple change polygons; the business attribute features corresponding to each change polygon are determined.

[0007] Based on the preset probability scoring model corresponding to the target business scenario, the business attribute features and local image pairs corresponding to each of the changing polygons are processed to determine the acceptance probability score of each of the changing polygons in the target business scenario.

[0008] The vector geometric data and business attribute features corresponding to the changed polygons whose acceptance probability scores are greater than or equal to the first acceptance threshold are determined as the change detection results in the target business scenario.

[0009] In one embodiment, the business attribute features include deterministic constraint features and probabilistic evaluation features; the preset probability scoring model is obtained by updating the initial probability scoring model based on the initial training sample data under the target business scenario.

[0010] In one embodiment, the step of processing the business attribute features and local image pairs corresponding to each of the changing polygons based on a preset probability scoring model corresponding to the target business scenario to determine the acceptance probability score of each of the changing polygons in the target business scenario includes:

[0011] From each of the described variable polygons, select target variable polygons whose deterministic constraint features satisfy the initial filtering conditions;

[0012] Based on a preset classifier, feature extraction is performed on the probabilistic evaluation features and local image pairs corresponding to each of the target variable polygons to obtain the acceptance probability score of each variable polygon in the target business scenario.

[0013] In one embodiment, the method further includes:

[0014] The polygon whose acceptance probability score is less than the first acceptance threshold and greater than the second acceptance threshold is identified as an uncertain polygon.

[0015] The business attribute features and business identifiers of the target business scenario corresponding to each of the uncertain polygons are sent to a preset review queue; the review label and the corresponding review confidence level of each of the uncertain polygons are determined through the preset review queue;

[0016] If the confidence level of the audit is greater than the preset confidence threshold, an updated sample set is generated, and each updated sample data in the updated sample set is stored in the model update queue corresponding to the business identifier. The updated sample data includes business attribute features, audit tags, audit time, and business identifier.

[0017] In one embodiment, both the first acceptance threshold and the second acceptance threshold are determined based on the performance trade-off curve corresponding to the initial training set, or based on a preset false negative rate, or determined by determining the adjustment value corresponding to the offset when the ratio of the number of uncertain polygons is greater than a ratio threshold or the number of uncertain polygons in the preset review queue is greater than a number threshold within a preset period.

[0018] In one embodiment, the method further includes:

[0019] If the number of updated sample data in the model update queue corresponding to the business identifier is greater than or equal to the update threshold, then the comprehensive loss value is calculated based on the acceptance probability score, review label and review confidence level corresponding to each updated sample data.

[0020] Based on the comprehensive loss value, the model parameters of the preset probability scoring model corresponding to the business identifier are updated to obtain the updated probability scoring model corresponding to the business identifier. The updated probability scoring model is used to execute the steps based on the preset probability scoring model corresponding to the target business scenario.

[0021] Secondly, this application also provides a remote sensing change detection device, comprising:

[0022] The processing module is used to process the dual-temporal remote sensing images of the target business scenario based on a preset change detection model to obtain multiple change polygons; and to determine the business attribute features corresponding to each change polygon.

[0023] The first determining module is used to process the business attribute features and local image pairs corresponding to each of the changing polygons based on the preset probability scoring model corresponding to the target business scenario, and determine the acceptance probability score of each of the changing polygons in the target business scenario.

[0024] The second determining module is used to determine the vector geometric data and business attribute features corresponding to the changed polygon whose acceptance probability score is greater than or equal to the first acceptance threshold, which are the change detection results under the target business scenario.

[0025] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0026] Based on a preset change detection model, dual-temporal remote sensing images of the target business scenario are processed to obtain multiple change polygons; the business attribute features corresponding to each change polygon are determined.

[0027] Based on the preset probability scoring model corresponding to the target business scenario, the business attribute features and local image pairs corresponding to each of the changing polygons are processed to determine the acceptance probability score of each of the changing polygons in the target business scenario.

[0028] The vector geometric data and business attribute features corresponding to the changed polygons whose acceptance probability scores are greater than or equal to the first acceptance threshold are determined as the change detection results in the target business scenario.

[0029] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0030] Based on a preset change detection model, dual-temporal remote sensing images of the target business scenario are processed to obtain multiple change polygons; the business attribute features corresponding to each change polygon are determined.

[0031] Based on the preset probability scoring model corresponding to the target business scenario, the business attribute features and local image pairs corresponding to each of the changing polygons are processed to determine the acceptance probability score of each of the changing polygons in the target business scenario.

[0032] The vector geometric data and business attribute features corresponding to the changed polygons whose acceptance probability scores are greater than or equal to the first acceptance threshold are determined as the change detection results in the target business scenario.

[0033] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0034] Based on a preset change detection model, dual-temporal remote sensing images of the target business scenario are processed to obtain multiple change polygons; the business attribute features corresponding to each change polygon are determined.

[0035] Based on the preset probability scoring model corresponding to the target business scenario, the business attribute features and local image pairs corresponding to each of the changing polygons are processed to determine the acceptance probability score of each of the changing polygons in the target business scenario.

[0036] The vector geometric data and business attribute features corresponding to the changed polygons whose acceptance probability scores are greater than or equal to the first acceptance threshold are determined as the change detection results in the target business scenario.

[0037] The aforementioned remote sensing change detection method, apparatus, computer equipment, readable storage medium, and program product process dual-temporal remote sensing images of a target business scenario using a preset change detection model to obtain multiple change polygons. They determine the business attribute features corresponding to each change polygon; based on a preset probability scoring model corresponding to the target business scenario, they process the business attribute features and local image pairs corresponding to each change polygon to determine the acceptance probability score of each change polygon in the target business scenario. This achieves a separation architecture between detection and business alignment. The change detection model possesses spatial recognition capabilities, and the probability scoring model is adapted to the business standards of the corresponding business scenario, decoupling change detection and business scoring judgment. The vector geometric data and business attribute features corresponding to change polygons with acceptance probability scores greater than or equal to a first acceptance threshold are determined as the change detection results in the target business scenario. The model output is then transformed into business decisions through probability scoring. The change polygons are stored in the form of vector geometric data and business attribute features, facilitating subsequent spatial analysis and business workflow, thereby improving the accuracy of change detection. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating a remote sensing change detection method in one embodiment;

[0040] Figure 2 This is a flowchart illustrating step 102 in one embodiment;

[0041] Figure 3 This is a schematic diagram of a remote sensing change detection method in one embodiment;

[0042] Figure 4 This is a schematic diagram of a decoupled detection-alignment architecture in one embodiment;

[0043] Figure 5 This is a structural block diagram of a remote sensing change detection device in one embodiment;

[0044] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] In one exemplary embodiment, such as Figure 1 As shown, a remote sensing change detection method is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0047] Step 101: Process the dual-temporal remote sensing images of the target business scene based on the preset change detection model to obtain multiple change polygons; determine the business attribute features corresponding to each change polygon.

[0048] The preset change detection model is a detection model that performs pixel-level change detection on dual-temporal remote sensing images and determines the connected regions of the changed areas. The target business scenario refers to the current change detection business procedure. Dual-temporal remote sensing images are remote sensing images of the same area acquired at two different times. The change polygon refers to the changed region formed by vectorizing connected changed pixel regions. Business attribute features can be structured attribute feature vectors, which can include deterministic constraint features and probabilistic evaluation features. Deterministic constraint features are attributes that can be defined and judged by explicit rules (hard rules). Probabilistic evaluation features are subjective standard features that cannot be explicitly defined by hard rules and need to be learned by the model from training data.

[0049] Specifically, the terminal performs pixel-level change detection on dual-temporal remote sensing images of the target business scenario based on a preset change detection model, obtaining a change probability map. The change probability map refers to the probability that the position of each pixel in the dual-temporal remote sensing image has changed. The change probability map is then subjected to binarization and thresholding processing, marking pixels with values ​​greater than a pixel threshold as changed pixels. Furthermore, vectorization processing is performed on the modified changed pixel regions, extracting the outer contours of connected regions, with each connected region representing a changed polygon. The specific value of this pixel threshold is related to the specific business scenario and is not limited here. Optionally, this preset change detection model focuses on recall, aiming to discover as many changed polygons as possible, which can be achieved by adjusting the pixel threshold.

[0050] The terminal can extract the business attribute features of the changing polygon corresponding to the target business scenario. It should be understood that the type of business attribute features extracted under different target business scenarios is determined based on the requirements of the corresponding business.

[0051] Step 102: Based on the preset probability scoring model corresponding to the target business scenario, process the business attribute features and local image pairs corresponding to each changing polygon to determine the acceptance probability score of each changing polygon in the target business scenario.

[0052] Among them, the preset probability scoring model refers to a pre-trained lightweight classification model; the acceptance probability score represents the confidence level that the changed polygon is accepted as a changed patch in the target business scenario.

[0053] Specifically, the terminal can determine the preset probability scoring model corresponding to the service identifier of the target service scenario based on the correspondence between service identifiers and probability scoring models. Then, based on this preset probability scoring model, it processes the service attribute features and local image pairs corresponding to the changing polygons to determine the acceptance probability score of each changing polygon in the target service scenario. The specific processing procedure is described in subsequent embodiments and will not be repeated here.

[0054] In one example, the preset probability scoring model is obtained by updating the initial probability scoring model based on the initial training sample data under the target business scenario. Alternatively, the preset probability scoring model can also be trained based on the updated sample set output during operation. The process of determining the updated sample set and training the preset probability scoring model will be described in detail in subsequent embodiments and will not be repeated here.

[0055] Step 103: Determine the vector geometric data and business attribute features corresponding to the changed polygons whose acceptance probability score is greater than or equal to the first acceptance threshold, which are the change detection results under the target business scenario.

[0056] Here, the first acceptance threshold refers to the lower limit threshold for determining a change as valid. The vector geometric data corresponding to the changed polygon refers to the spatial location data of the changed polygon. The change detection result can be a changed patch.

[0057] Specifically, when the terminal determines that the acceptance probability score is greater than or equal to the first acceptance threshold, it uses the vector geometric data and business attribute features of the changed polygon as the change detection result for the current business scenario. That is, the changed polygon has undergone a change that meets the acceptance criteria for this business scenario. This change detection result is then stored in the production database. This change detection result is the business change patch.

[0058] Furthermore, this change detection data can be used in subsequent business processes. For example, illegal construction patches in urban construction monitoring will be incorporated into the enforcement process, and change patches in farmland protection will be included in the annual farmland change statistics. It should be understood that the examples here are for illustrative purposes only and do not constitute specific limitations.

[0059] The aforementioned remote sensing change detection method processes dual-temporal remote sensing images of a target business scenario using a pre-set change detection model to obtain multiple change polygons. It then determines the business attribute features corresponding to each change polygon. Based on a pre-set probability scoring model corresponding to the target business scenario, it processes the business attribute features and local image pairs corresponding to each change polygon to determine the acceptance probability score of each change polygon in the target business scenario. This achieves a separation architecture between detection and business alignment. The change detection model possesses spatial recognition capabilities, and the probability scoring model is adapted to the business standards of the corresponding business scenario, decoupling change detection and business scoring judgment. The method identifies the vector geometric data and business attribute features corresponding to change polygons with acceptance probability scores greater than or equal to a first acceptance threshold, which are then used as the change detection results for the target business scenario. The model output is transformed into business decisions through probability scoring. The change polygons are stored in the form of vector geometric data and business attribute features, facilitating subsequent spatial analysis and business workflow, thereby improving the accuracy of change detection.

[0060] In one exemplary embodiment, such as Figure 2As shown, the specific implementation process of step 102, "based on the preset probability scoring model corresponding to the target business scenario, processing the business attribute features and local image pairs corresponding to each changed polygon to determine the acceptance probability score of each changed polygon in the target business scenario," may include:

[0061] Step 201: From each variable polygon, select the target variable polygon whose deterministic constraint features satisfy the initial filtering conditions.

[0062] Step 202: Based on the preset classifier, extract features from the probabilistic evaluation features and local image pairs corresponding to each target change polygon to obtain the acceptance probability score of each change polygon in the target business scenario.

[0063] Here, the initial filtering condition refers to the hard rule judgment condition of the deterministic filtering layer; the target variation polygon refers to the variation polygon after being filtered by the initial filtering condition, that is, the candidate variation that meets all formalizable criteria; the preset classifier is used to perform soft criterion scoring on the target variation polygon, and the preset classifier can be a lightweight classifier. The local image pair refers to the local image slice acquired from the dual-temporal image based on the bounding rectangle of the variation polygon.

[0064] Specifically, the terminal can select target variable polygons from the variable polygons whose deterministic constraint features meet the initial filtering conditions. In addition, the acceptance probability score of variable polygons whose deterministic constraint features do not meet the preset filtering conditions is set to 0, and these variable polygons are stored in the high-confidence rejection queue corresponding to each service identifier, and are not processed.

[0065] Based on a pre-defined classifier, feature extraction is performed on the probabilistic evaluation features and local image pairs corresponding to each target change polygon to obtain the acceptance probability score of each change polygon in the target business scenario. It should be understood that the pre-defined classifier can learn business judgment criteria that cannot be hard-coded from the training data to classify changes in the probabilistic evaluation features and output the corresponding acceptance probability score. Optionally, the acceptance probability score can range from (0,1).

[0066] In addition, the preset probability scoring model can be an alignment module, which includes a deterministic filtering layer and a lightweight classifier. The deterministic filtering layer filters each variant polygon based on deterministic constraint features. The acceptance probability score of variant polygons whose deterministic constraint features do not meet the formalizable criteria is set as a target value. This target value indicates that the variant polygon does not meet the change condition. For example, the target value can be 0, and the classifier is skipped, that is, no further processing is performed on the variant polygon. Variant polygons whose deterministic constraint features meet the formalizable criteria are input into the lightweight classifier to obtain the acceptance probability score.

[0067] Optionally, the standard parameters of the deterministic filter layer and the lightweight classifier are stored independently. The standard parameters of the deterministic filter layer can be stored in the corresponding configuration file, and the standard parameters of the lightweight classifier can be stored in the corresponding model file. Adjustments are made to the standard parameters in the formalizable criteria corresponding to the deterministic filter layer and the learning soft criteria in the classifier. For example, when the business rules in the target business scenario change, the standard parameters of the formalizable criteria are adjusted based on the changed business rules. The model weights in the classifier are updated based on the updated sample data. Optionally, the updated sample data can be paired training samples, including business attribute features corresponding to polygons, approval labels, and business identifiers.

[0068] In one example, the business attribute feature can be Attr, where Attr = [Area, Shape_regularity, Conf_score, Change_type, Spatial_context], and Area is the changed area (square meters); Shape_regularity is the shape regularity (a geometric indicator of the compactness of a polygon); Conf_score is the mean of the change probability map within the polygon; Change_type is the change type category code, which is the dimension with the highest weight in alignment judgment, and the differences in the institutional definition of reportable change types among different business procedures are mainly reflected here; Spatial_context is the spatial relationship feature with surrounding features, where change type and changed area are the most critical factors affecting business acceptance decisions. It should be understood that Attr can be extended according to business needs, such as adding historical record features, neighborhood context features, and spatial relationship features with specific thematic layers; the alignment module architecture supports feature dimension expansion without changing the overall process. Change area, change type category code, and detection confidence are necessary features for alignment judgment, while Shape_regularity and Spatial_context are important extended dimensions.

[0069] Deterministic constraint features include at least one or more of the following: changed area, changed type, and changed region location. For example, if the changed area is greater than or equal to an area threshold, the changed type belongs to a preset whitelist, and the changed region location falls within the jurisdiction, the initial filtering conditions are determined to be met. If any one of these conditions is not met, the initial filtering conditions are determined not to be met, and the acceptance probability score of the corresponding changed polygon is set to 0, and no further processing is performed on that changed polygon. It should be understood that when business rules change, the area threshold, preset whitelist, and jurisdiction can be adjusted accordingly. Specific adjustments related to changes in business rules are not specifically limited here.

[0070] Probabilistic evaluation features include at least one or more of the following: shape regularity, detection confidence, spatial context features, and image appearance features. Shape regularity refers to the geometric shape of the changing polygon, used to determine whether it conforms to typical shape characteristics of real changes in the business scenario; shape regularity can be determined based on compactness. Detection confidence is the degree of certainty that the changing polygon region detected by the change detection model has changed; this detection confidence can be the average probability of change for each pixel in the changing polygon. Spatial context features are the spatial relationship features between the changing polygon and its surrounding features, used to determine whether it conforms to common patterns in the business process. Through spatial analysis, the distance, adjacency relationship, and land use percentage within the buffer zone are calculated between the changing polygon and surrounding features. The land use percentage within the buffer zone refers to the area proportion of adjacent land use types of the changing polygon, used to determine the geographical environment in the target business scenario. Image appearance features refer to the visual feature vectors obtained by encoding local image slices, reflecting the true physical state of the land surface.

[0071] It should be understood that the explanations here are for illustrative purposes only. The specific types of business attributes and the specific values ​​corresponding to each feature, as well as the specific thresholds involved in the hard rule judgment conditions, are all related to the business requirements of the target business scenario. They are only used as examples here and do not constitute specific limitations.

[0072] Optionally, in the preset review queue, a judgment decision can be made for each uncertain polygon. The judgment decision is used to determine whether the uncertain polygon is accepted or rejected, and the reason for the decision can also be returned.

[0073] In this embodiment, by first deterministically filtering to quickly eliminate obviously unqualified variable polygons, and then performing fine scoring on the remaining variable polygons, the efficiency and accuracy of variable polygon detection are improved. Furthermore, by comprehensively judging polygons based on local imagery and attribute features, considering both geometric statistics and regularity features, as well as visual semantic features that cannot be structured and encoded, the complementary nature of these two aspects forms a complete information basis for aligned judgment, enhancing the model's ability to discriminate complex scenes.

[0074] In one exemplary embodiment, the remote sensing change detection method further includes:

[0075] The polygons with acceptance probability scores less than the first acceptance threshold and greater than the second acceptance threshold are identified as uncertain polygons. The business attribute features and business identifiers of the target business scenario corresponding to each uncertain polygon are sent to a preset review queue. The review label and corresponding review confidence level of each uncertain polygon are determined through the preset review queue. If the review confidence level is greater than the preset confidence level threshold, an update sample set is generated, and each update sample data in the update sample set is stored in the model update queue corresponding to the business identifier.

[0076] The second acceptance threshold is the upper limit for the acceptance probability score of a polygon to meet the change requirements. The business identifier indicates the business procedure (target business scenario) to which the changed polygon belongs. The review label is the decision result after re-reviewing the changed polygon based on the review decision; for example, it can be an acceptance label and a rejection label. Review confidence refers to the credibility weight of the review decision. For example, review confidence is data used to verify the consistency of review results through multiple reviews of the same uncertain polygon. The updated sample set is a collection of uncertain polygons that meet the review confidence standard and their review labels, used for incremental training of the preset change detection model and / or the preset probability scoring model. The model update queue is a sample queue stored in isolation according to business identifiers, accumulating updated samples under the same business procedure, triggering a model update after reaching a threshold. Updated sample data includes business attribute features, review labels, review time, and business identifiers. The review time refers to the timestamp of the review decision, used for sample timeliness management.

[0077] Specifically, the terminal can identify polygons with acceptance probability scores less than a first acceptance threshold but greater than a second acceptance threshold as uncertain polygons; send the business attribute features and business identifiers of the target business scenario corresponding to each uncertain polygon to a preset review queue; determine the review tags for each uncertain polygon through the preset review queue; and determine the review confidence levels corresponding to multiple review tags for each uncertain polygon. If the review confidence level is greater than a preset confidence threshold, the business attribute features, review tags, review time, and business identifier are identified as update sample data, and each update sample data in the update sample set is stored in the model update queue corresponding to the business identifier. In one example, such as... Figure 3As shown, decision-making can be performed manually on uncertain polygons in a preset review queue to obtain a review label for each uncertain polygon. The review confidence score for each label is calculated, and this confidence score is used to calculate the loss value in the subsequent training process of the probability scoring model. When multiple review labels are consistent and the review confidence score is greater than a preset confidence threshold, the business attribute label, review label, review time, and business identifier corresponding to the uncertain polygon are determined as update sample data, which carries the review confidence score. Based on the business identifier, the corresponding update sample data is stored in the corresponding model update queue. When the number of update sample data in the model update queue is greater than or equal to a preset threshold, the probability scoring model is updated using the update sample dataset to obtain the updated probability scoring model. In subsequent business processing, the changed polygons are processed based on this updated probability scoring model. The model update queue for each model is a private queue for that business program. That is, it includes multiple business programs (A, ..., N), with a preset threshold of N_update. It should be understood that the preset threshold for each business program can be the same or different, and can be determined based on the actual application scenario. In addition, the updated sample data corresponding to each business identifier is stored in the basic model update queue corresponding to the change detection model. When a preset period is reached, the model parameters of the preset change detection model are updated based on the updated sample data in the basic model update queue, and the model parameters of the corresponding probability scoring model are updated based on the corresponding model update queue. For example, the preset period can be every quarter.

[0078] Additionally, if the audit confidence level is less than or equal to a preset confidence threshold, the data is transferred to a manual review pool for further review. A review label is then generated. If the review label's representation is acceptable, the audit confidence level is determined based on the review label. Furthermore, the business attribute features, review label, audit time, and business identifier corresponding to the review label's polygon are identified as updated sample data. The audit confidence level corresponding to this polygon can be determined based on the audit label and the review label.

[0079] The terminal can identify polygons with acceptance probability scores greater than or equal to a first acceptance threshold as high-confidence acceptance data; and polygons with acceptance probability scores less than or equal to a second acceptance threshold as high-confidence rejection data. The terminal can perform random checks on each set of high-confidence acceptance and rejection data at a preset period, obtain the check results, and use the check results, business attribute characteristics, check time, and business identifier as update sample data to verify the accuracy of the preset probability scoring model's automatic judgment and continuously supplement training data.

[0080] In this embodiment, variable polygons are distinguished using a first acceptance threshold and a second acceptance threshold, improving the efficiency of the review process. Uncertain polygons are reviewed again, and this data is used as feedback data. Furthermore, highly reliable review labels are selected as training samples based on the review confidence level, improving the accuracy of the model's judgment. Update samples are stored and isolated according to business identifiers to ensure that training data between different business procedures does not overlap, maintaining model independence.

[0081] In an exemplary embodiment, both the first acceptance threshold and the second acceptance threshold are determined based on the performance trade-off curve corresponding to the initial training set, or based on a preset false negative rate, or determined by an adjustment value corresponding to the offset when the ratio of the number of uncertain polygons within a preset period is greater than a ratio threshold or the number of uncertain polygons in a preset review queue is greater than a quantity threshold.

[0082] The initial training set is the dataset used to train the initial probability scoring model, including labeled business attribute features and corresponding ground truth labels; the performance trade-off curve can be a precision-recall curve (PRC), with recall on the horizontal axis and precision on the vertical axis, used to evaluate model performance under different thresholds; the preset false negative rate is the upper limit of the false negative rate; the preset period refers to the time window for detection, used to periodically observe the changing trend of the proportion of uncertain polygons; the number ratio is the proportion of uncertain polygons to the total number of changing polygons within a certain period.

[0083] Specifically, the terminal can process the polygons in the initial training set based on the probability scoring model to obtain the acceptance probability score of each sample polygon; for each initial threshold, calculate the precision and recall corresponding to each acceptance probability score at that initial threshold, and draw a precision-recall curve with recall on the horizontal axis and precision on the vertical axis; find the corresponding point on the curve where the precision reaches the business-acceptable level, and set its score value as the first acceptance threshold; find the corresponding point on the curve where the recall reaches the business-acceptable level, and set its score value as the second acceptance threshold.

[0084] Determine the upper limit of the false negative rate stipulated in the business procedure, and determine the minimum recall rate based on the upper limit of the false negative rate. Optionally, the minimum recall rate = 1 - false negative rate; determine the acceptance probability score corresponding to the minimum recall rate on the precision-recall curve, and determine the acceptance probability score as the second acceptance threshold; and determine the first acceptance threshold based on the second acceptance threshold and the auditable quantity.

[0085] Within a preset period, the proportion of uncertain polygons and the number of pending cases in the review queue are obtained. If the proportion of uncertain cases remains higher than the preset upper limit for several consecutive periods, or if the backlog in the review queue exceeds the capacity limit, the first acceptance threshold and / or the second acceptance threshold are adjusted. If the proportion of uncertain cases remains lower than the preset lower limit for several consecutive periods, the second acceptance threshold is lowered or the first acceptance threshold is raised. The specific adjustment value is related to the actual operating conditions, and the terminal can determine the corresponding adjustment value based on the rate at which uncertain cases increase or decrease.

[0086] In addition, in business procedures that require high recall rates in regulatory tasks, the terminal can adjust the first acceptance threshold based on the recall rate.

[0087] In one example, the accuracy of the probability scoring model (alignment module) corresponding to each business procedure is evaluated periodically. When the accuracy is lower than the target value, the first acceptance threshold is automatically lowered to increase the coverage of manual review and to trigger incremental fine-tuning first.

[0088] In this embodiment, by initializing and adapting the first and second acceptance thresholds, the acceptance thresholds are made to meet the needs of the business scenario.

[0089] In one exemplary embodiment, the remote sensing change detection method further includes:

[0090] If the number of updated sample data in the model update queue corresponding to the business identifier is greater than or equal to the update threshold, then the comprehensive loss value is calculated based on the acceptance probability score, review label, and review confidence level corresponding to each updated sample data. Based on the comprehensive loss value, the model parameters of the preset probability scoring model corresponding to the business identifier are updated to obtain the updated probability scoring model corresponding to the business identifier.

[0091] The updated probability scoring model is used to execute steps based on the preset probability scoring model corresponding to the target business scenario; model parameters refer to the learnable parameters in the probability scoring model, such as model weights.

[0092] Specifically, if the number of updated sample data in the model update queue corresponding to the business identifier is greater than or equal to the update threshold, then a comprehensive loss value is calculated based on the acceptance probability score, review label, and review confidence level corresponding to each updated sample data, using a loss function. For example, the loss value corresponding to each updated sample data can be calculated based on the cross-entropy function, and the weight of each loss value can be calculated based on the review confidence level. The loss values ​​are then weighted and summed to obtain the comprehensive loss value. Based on the comprehensive loss value, the model parameters of the preset probability scoring model corresponding to the business identifier are updated to obtain the updated probability scoring model corresponding to the business identifier. Optionally, the model weights of the probability scoring model can be adjusted using gradient descent. It should be understood that the parameters of the change detection model are not adjusted here.

[0093] Optionally, when the regulatory rules for a certain business procedure change, only the alignment module of that procedure (probability scoring model) is incrementally fine-tuned, while the alignment modules of other business procedures and the parameters of the shared basic detection model remain unchanged.

[0094] In addition, the terminal can update the preset change detection model at a specified period based on the updated sample set corresponding to each business identifier to obtain the updated change detection model. It should be understood that after the update, since the numerical distribution of business attribute features may change, each alignment module needs to perform an adaptive fine-tuning using recent backflow samples to realign the feature space.

[0095] In one example, the initial remote sensing change detection model may include a base change detection model and a base probability scoring model. The process of training the initial remote sensing change detection model may include the following steps.

[0096] Obtain an initial training set, which includes historical dual-temporal imagery and historical operational patch data;

[0097] Based on the basic change detection model, historical dual-temporal images are processed to obtain multiple sets of historical polygons. Each historical polygon is then spatially overlaid with its corresponding historical business patch data. Specifically, the spatial overlay analysis is calculated as follows: spatial intersection is performed on the historical polygons and historical business patches from the same period, and the ratio of the intersection area to the area of ​​the historical polygon is calculated. For example, this ratio can be the Intersection over Union (IoU) or the coverage ratio. This ratio is determined as the spatial overlap. Based on this spatial overlap and a matching threshold, the label of each historical polygon is determined: if the spatial overlap exceeds the matching threshold, the label is determined as accept; otherwise, it is marked as reject.

[0098] The business attribute features of each historical polygon are extracted from the corresponding historical business patch data. These features can be attribute feature vectors (Attr) and paired with corresponding labels to form paired samples {Attr, label, program_id}. These samples are then stored in the pre-defined training queue for each business program, categorized by program_id. It should be understood that each business program can construct hundreds to thousands of valid paired samples for the initial training of a lightweight classifier, resulting in a pre-trained classifier. The historical business patches have been verified by domain experts, ensuring high label reliability. The large volume of historical data allows for the construction of a substantial initial training set at once. The data originates from real-world business scenarios, naturally aligning with the operational environment, and is superior to simulated data and general labeled data.

[0099] In another instance, for situations where historical data archives are incomplete or where historical business patches do not match the detection polygons in time, historical polygons can be manually labeled. In this process, business personnel determine the corresponding accept / reject judgments for several typical AI detection polygons to construct a training set. The amount of labeling required at this stage is far less than pixel-level sample labeling; typically, tens to hundreds of pairs are sufficient to support the initial training of the alignment module.

[0100] In one instance, such as Figure 4 As shown, Figure 4This is a schematic diagram of a decoupled detection and alignment architecture. The specific architecture may include a change detection model and an alignment module group. The change detection module is used to process dual-temporal remote sensing images to obtain multiple change polygons and the business attribute features corresponding to each change polygon. The alignment module group includes multiple probability scoring models corresponding to different business identifiers. For example, different business identifiers could be the identifier for urban construction detection A, the identifier for farmland protection B, the identifier for ecological conservation C, etc. This is only an example and does not constitute a specific limitation. Each probability scoring model is used to process the business attribute features corresponding to the corresponding polygon and the corresponding local image pair to obtain the corresponding acceptance probability score P_accept. The model then divides each acceptance probability score using a first acceptance threshold Th_high and a second acceptance threshold Th_low. Cases with P_accept ≥ Th_high are identified as high-confidence cases and stored in the production database for subsequent business processing. Cases with P_accept ≤ Th_low are identified as low-confidence cases and discarded. Cases with Th_low ≤ P_accept ≤ Th_high are identified as uncertain cases and stored in a preset review queue. Multiple review labels are obtained for each uncertain case by reviewing the polygons of the same uncertain case. When multiple review labels are consistent, the corresponding paired training sample data is stored in the corresponding model update queue. When the model update condition is met, the model parameters of the corresponding probability scoring model and / or change detection model are updated. The model update condition is that when the sample data in the model update queue of each probability scoring model reaches a preset threshold, the model parameters of that probability scoring model are updated. At this time, other probability update queues and change detection models are not updated. When the preset period is reached, the model parameters of the change detection model and each probability scoring model are updated. It should be understood that the updated model is used in the subsequent change detection business process.

[0101] In this embodiment, the detection and alignment functions are decoupled. The detection function aims for maximum recall across all objective changes and is handled by a single, shared base detection model. The alignment function aims for maximum matching to the acceptance standards of specific business procedures and is handled by a dedicated, lightweight alignment module for each procedure. The two functions are connected by structured information at the polygon level. The alignment module operates on a panoramic pixel-level processing flow independent of the base detection model, making judgments based solely on attribute feature vectors and locally cropped image slices. This allows the two components to be designed, trained, and updated independently. The advantages of using polygon-level information as the data interface are: the alignment module only processes the attribute features and local slices of a single polygon, resulting in significantly less computation than panoramic pixel-level processing, supporting lightweight incremental learning; the interface layer completely decouples the detection model and alignment module, eliminating the need to rewrite the alignment module when upgrading the base model architecture; and attribute feature vectors can be expanded to new dimensions as needed while the overall process remains unchanged.

[0102] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0103] Based on the same inventive concept, this application also provides a remote sensing change detection device for implementing the remote sensing change detection method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the remote sensing change detection device provided below can be found in the limitations of the remote sensing change detection method described above, and will not be repeated here.

[0104] In one exemplary embodiment, such as Figure 5 As shown, a remote sensing change detection device 50 is provided, including: a processing module 51, a first determining module 52, and a second determining module 53, wherein:

[0105] Processing module 51 is used to process dual-temporal remote sensing images of the target business scene based on a preset change detection model to obtain multiple change polygons; and to determine the business attribute features corresponding to each change polygon.

[0106] The first determining module 52 is used to process the business attribute features and local image pairs corresponding to each of the changing polygons based on the preset probability scoring model corresponding to the target business scenario, and determine the acceptance probability score of each of the changing polygons in the target business scenario.

[0107] The second determining module 53 is used to determine the vector geometric data and business attribute features corresponding to the changed polygon whose acceptance probability score is greater than or equal to the first acceptance threshold, which are the change detection results under the target business scenario.

[0108] In one embodiment, the business attribute features include deterministic constraint features and probabilistic evaluation features; the preset probability scoring model is obtained by updating the initial probability scoring model based on the initial training sample data under the target business scenario.

[0109] In one embodiment, the first determining module 52 is used to filter target variable polygons whose deterministic constraint features satisfy the initial filtering conditions from each of the variable polygons.

[0110] Based on a preset classifier, feature extraction is performed on the probabilistic evaluation features and local image pairs corresponding to each of the target variable polygons to obtain the acceptance probability score of each variable polygon in the target business scenario.

[0111] In one embodiment, the second determining module 53 is further configured to determine that the changing polygon whose acceptance probability score is less than a first acceptance threshold and greater than a second acceptance threshold is an uncertain polygon;

[0112] The business attribute features and business identifiers of the target business scenario corresponding to each of the uncertain polygons are sent to a preset review queue; the review label and the corresponding review confidence level of each of the uncertain polygons are determined through the preset review queue;

[0113] If the confidence level of the audit is greater than the preset confidence threshold, an updated sample set is generated, and each updated sample data in the updated sample set is stored in the model update queue corresponding to the business identifier. The updated sample data includes business attribute features, audit tags, audit time, and business identifier.

[0114] In one embodiment, both the first acceptance threshold and the second acceptance threshold are determined based on the performance trade-off curve corresponding to the initial training set, or based on a preset false negative rate, or determined by determining the adjustment value corresponding to the offset when the ratio of the number of uncertain polygons is greater than a ratio threshold or the number of uncertain polygons in the preset review queue is greater than a number threshold within a preset period.

[0115] In one embodiment, the device further includes:

[0116] The calculation module is used to calculate the comprehensive loss value based on the acceptance probability score, review label and review confidence level corresponding to each update sample data if the number of update sample data in the model update queue corresponding to the business identifier is greater than or equal to the update threshold.

[0117] The update module is used to update the model parameters of the preset probability scoring model corresponding to the business identifier based on the comprehensive loss value, so as to obtain the updated probability scoring model corresponding to the business identifier. The updated probability scoring model is used to execute the step of the preset probability scoring model corresponding to the target business scenario.

[0118] Each module in the aforementioned remote sensing change detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0119] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a remote sensing change detection method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0120] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0121] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0122] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0123] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0124] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0125] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0126] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0127] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A remote sensing change detection method, characterized in that, The method includes: Based on a preset change detection model, dual-temporal remote sensing images of the target business scenario are processed to obtain multiple change polygons; the business attribute features corresponding to each change polygon are determined. Based on the preset probability scoring model corresponding to the target business scenario, the business attribute features and local image pairs corresponding to each of the changing polygons are processed to determine the acceptance probability score of each of the changing polygons in the target business scenario. The vector geometric data and business attribute features corresponding to the changed polygons whose acceptance probability scores are greater than or equal to the first acceptance threshold are determined as the change detection results in the target business scenario.

2. The method according to claim 1, characterized in that, The business attribute features include deterministic constraint features and probabilistic evaluation features; the preset probability scoring model is obtained by updating the initial probability scoring model based on the initial training sample data under the target business scenario.

3. The method according to claim 2, characterized in that, The process of processing the business attribute features and local image pairs corresponding to each of the changing polygons based on the preset probability scoring model corresponding to the target business scenario to determine the acceptance probability score of each of the changing polygons in the target business scenario includes: From each of the described variable polygons, select target variable polygons whose deterministic constraint features satisfy the initial filtering conditions; Based on a preset classifier, feature extraction is performed on the probabilistic evaluation features and local image pairs corresponding to each of the target variable polygons to obtain the acceptance probability score of each variable polygon in the target business scenario.

4. The method according to claim 1, characterized in that, The method further includes: The polygon whose acceptance probability score is less than the first acceptance threshold and greater than the second acceptance threshold is identified as an uncertain polygon. The business attribute features and business identifiers of the target business scenario corresponding to each of the uncertain polygons are sent to a preset review queue; the review label and the corresponding review confidence level of each of the uncertain polygons are determined through the preset review queue; If the confidence level of the audit is greater than the preset confidence threshold, an updated sample set is generated, and each updated sample data in the updated sample set is stored in the model update queue corresponding to the business identifier. The updated sample data includes business attribute features, audit tags, audit time, and business identifier.

5. The method according to claim 4, characterized in that, Both the first acceptance threshold and the second acceptance threshold are determined based on the performance trade-off curve corresponding to the initial training set, or based on a preset false negative rate, or determined by an adjustment value corresponding to the offset when the ratio of the number of uncertain polygons is greater than a ratio threshold or the number of uncertain polygons in the preset review queue is greater than a number threshold.

6. The method according to claim 4, characterized in that, The method further includes: If the number of updated sample data in the model update queue corresponding to the business identifier is greater than or equal to the update threshold, then the comprehensive loss value is calculated based on the acceptance probability score, review label and review confidence level corresponding to each updated sample data. Based on the comprehensive loss value, the model parameters of the preset probability scoring model corresponding to the business identifier are updated to obtain the updated probability scoring model corresponding to the business identifier. The updated probability scoring model is used to execute the steps based on the preset probability scoring model corresponding to the target business scenario.

7. A remote sensing change detection device, characterized in that, The device includes: The processing module is used to process the dual-temporal remote sensing images of the target business scenario based on a preset change detection model to obtain multiple change polygons; and to determine the business attribute features corresponding to each change polygon. The first determining module is used to process the business attribute features and local image pairs corresponding to each of the changing polygons based on the preset probability scoring model corresponding to the target business scenario, and determine the acceptance probability score of each of the changing polygons in the target business scenario. The second determining module is used to determine the vector geometric data and business attribute features corresponding to the changed polygon whose acceptance probability score is greater than or equal to the first acceptance threshold, which are the change detection results under the target business scenario.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images