Anomaly detection method and device for planar geographic entities

Abstract

The present invention provides an anomaly detection method and device for planar geographic entities. The method comprises: constructing and training a fused semantic segmentation model including at least two fused semantic segmentation models set in parallel and using an additive fusion method to fuse the prediction results of each fused semantic segmentation model as a classification prediction result; constructing and training a fused change detection model including at least two change detection models set in parallel and using an additive fusion method to fuse the prediction results of each change detection model as a change prediction result; acquiring image data of a to-be-detected area and planar geographic entity data to be detected, obtaining a to-be-detected data set and a land class label set; inputting the to-be-detected data set into the fused semantic segmentation model and the fused change detection model, obtaining a classification prediction result and a change prediction result; using the classification prediction result to assign category attributes to the change prediction result, obtaining a fused prediction result; and comparing the fused prediction result with the elements with the same name in the land class label set.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image data processing, and in particular to an anomaly detection method and device for planar geographic entities. Background Art

[0002] With the rapid development of remote sensing technology and geographic information systems, data processing and analysis based on satellite images have been widely used in natural resource monitoring, environmental change assessment, urban planning and other fields. As an important part of remote sensing image application, the automated detection and quality inspection technology of planar geographic entities is directly related to the accurate extraction of ground object information and the reliability of spatial data. However, the current methods of detecting planar geographic entities mainly rely on manual operation or automated detection of a single model, which has the following technical challenges:

[0003] 1. Insufficient detection accuracy: A single semantic segmentation model or change detection model has limited performance in complex terrain, especially in large-scale, different phases, and multi-type landforms, which is prone to detection errors or missed detections. Different types of land objects have different performances in images, and a single model is difficult to adapt to the diverse land features, resulting in low detection accuracy.

[0004] 2. Complex integration and processing of multi-source data: In practical applications, the detection of planar geographic entities involves remote sensing image data from multiple time periods, different sensors, and different resolutions. Since these image data differ in lighting conditions, resolution, and acquisition time, a single detection model is difficult to effectively handle these inter-domain differences, thus affecting the accuracy of anomaly detection.

[0005] 3. Insufficient coordination between category prediction and change detection: In traditional facet geographic entity detection, semantic segmentation models and change detection models often work independently. Semantic segmentation is used to classify each pixel in an image, while change detection is used to identify changes between images of different periods. Due to the lack of fusion and unified framework between the two, the detection results cannot fully reflect the actual changes, reducing the accuracy and reliability of anomaly detection.

[0006] 4. Low automation of quality inspection process: Current quality inspection of surface geographic entities mostly relies on manual comparison and analysis, which is labor-intensive, inefficient, and subjectivistic, making it difficult to adapt to the rapid processing requirements of large-scale data. This inefficient working mode seriously restricts the quality inspection efficiency and accuracy of geographic entity data results. Summary of the invention

[0007] In order to solve the deficiencies existing in the above-mentioned prior art, the present invention aims to provide a method and device for detecting anomalies of planar geographic entities based on the fusion of semantic segmentation and change detection models. The accuracy of planar geographic entity detection is improved by the collaborative work of multiple semantic segmentation models and change detection models, and the model fusion technology is used to solve the domain difference problem of multi-source data, ultimately realizing efficient and automated quality inspection of complex planar geographic entities.

[0008] In order to achieve the above-mentioned invention object, the technical solution provided by the present invention includes:

[0009] The anomaly detection method for planar geographic entities comprises the following steps:

[0010] Constructing and training at least two semantic segmentation models arranged in parallel and fusing the prediction results of the fused semantic segmentation models using an additive fusion method as a fused semantic segmentation model of classification prediction results;

[0011] Constructing and training at least two change detection models arranged in parallel and fusing the prediction results of the change detection models using an additive fusion method as a fused change detection model of change prediction results;

[0012] Obtain image data of the area to be detected and the planar geographic entity to be detected, and obtain the data set to be detected and the land class label set after preprocessing;

[0013] Inputting the to-be-detected data set into the fused semantic segmentation model and the fused change detection model respectively to obtain a classification prediction result and a change prediction result;

[0014] Assigning category attributes to the change prediction result using the classification prediction result to obtain a fusion prediction result;

[0015] The fusion prediction result is compared with the elements with the same name in the land class label set to complete anomaly detection.

[0016] Preferably, the training method of the fused semantic segmentation model and the fused change detection model includes:

[0017] Acquiring a sample data set with prior information to complete preliminary training of the fused semantic segmentation model and the fused change detection model;

[0018] The to-be-detected data set and the sample data set are mixed as a new training sample set to complete the fine-tuning training of the fused semantic segmentation model and the fused change detection model.

[0019] Preferably, the preprocessing method of the image data of the area to be detected and the planar geographic entity to be detected includes:

[0020] The planar geographic entity to be detected is rasterized and, together with the image data of the area to be detected, is divided into a number of image blocks and label blocks that are consistent with the element resolution and size in the sample data set to form the data set to be detected and the land class label set.

[0021] Preferably, the method for determining the elements with the same name includes:

[0022] Each label block in the fusion prediction result is traversed and the overlap degree is calculated respectively, and the label blocks with the overlap degree higher than the preset overlap threshold are regarded as the elements with the same name.

[0023] Preferably, before comparing the fused prediction results with the elements with the same name in the land class label set, the method further includes removing the fused prediction results whose areas are smaller than a preset area threshold.

[0024] Preferably, performing image conversion on the image data of the area to be detected and / or the sample data set comprises the steps of:

[0025] Select several images containing the planar geographic entity to be detected from the image data of the area to be detected and / or the sample data set as target images, convert the target images to the LAB color space and calculate the average value and variance, convert the image data of the area to be detected and / or the sample data set to the LAB color space, and perform image conversion using the following formula:

[0026] ;

[0027] in, and are the mean and variance of the target image, and are respectively the image data of the area to be detected and\or the mean value and variance of the sample data set, and are respectively the image data of the area to be detected and\or the original value and the converted value of the image of the sample data set.

[0028] Preferably, the additive fusion method includes: superimposing the predicted probability maps output by all semantic segmentation models, and if the superimposed pixel value is greater than a preset fusion threshold, setting the value of the pixel to 1, otherwise it is set to 0; wherein 1 is the predicted category and 0 is the background.

[0029] The present invention also provides a device, which is used to implement the above method.

[0030] Beneficial Effects

[0031] 1. Improve detection accuracy: By fusing the results of multiple semantic segmentation models and change detection models, the present invention can fully utilize the advantages of different models in different scenarios, overcome the limitations of a single model, and effectively enhance the accuracy of prediction results. Especially in complex terrain, different seasons or lighting conditions, the present invention significantly improves the recognition accuracy of abnormal planar entities and reduces missed detections and false detections.

[0032] 2. Adapt to multi-source heterogeneous data: The present invention can effectively cope with the challenges brought by multi-source heterogeneous data by pre-processing remote sensing image data from different times, sensors and spatial resolutions, and combining the adaptive training and fine-tuning technology of the model. It solves the problems of differences in spectrum, resolution and phase of different image data, ensures the consistency and reliability of detection results, and is suitable for large-scale, cross-time period surface geographic entity detection.

[0033] 3. Refined processing of anomaly detection results: The present invention assigns category attributes to change detection results through the classification prediction results of the semantic segmentation model, and compares the elements with the same name based on standards such as overlap rate and area threshold, thereby realizing refined processing of anomaly detection and effectively combining category information with change information, making anomaly detection results more accurate and reducing false alarm problems caused by mismatch of category information.

[0034] 4. Improve the efficiency of automated quality inspection: Through automated training, reasoning and fusion of models, the present invention can automatically inspect large-scale data without human intervention, greatly improving the efficiency of quality inspection. Compared with traditional manual quality inspection methods, the present invention not only significantly reduces manual workload, but also reduces subjective judgment errors, providing an efficient and objective solution for the quality inspection of geographic entity data results. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic flow chart of an anomaly detection method for planar geographic entities provided by a preferred embodiment of the present invention;

[0036] Figure 2 A schematic diagram of the model structure and data flow direction provided for a preferred embodiment of the present invention;

[0037] Figure 3 A schematic diagram for comparing the semantic segmentation results of each model is provided for another preferred embodiment of the present invention;

[0038] Figure 4 A schematic diagram for comparing the change prediction results of each model is provided for another preferred embodiment of the present invention; DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described below in conjunction with the accompanying drawings. In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside", "outside" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention.

[0040] like Figure 1 , Figure 2 As shown, this embodiment provides an anomaly detection method for planar geographic entities, comprising the steps of:

[0041] S1. Construct and train at least two semantic segmentation models set in parallel and use the additive fusion method to fuse the prediction results of each fused semantic segmentation model as a fused semantic segmentation model of classification prediction results.

[0042] The fused semantic segmentation model is commonly used in typical object segmentation scenarios in this field. Mainstream semantic segmentation models include but are not limited to HRNet, DeepLab V3+, Swin Transformer, OCRNet, etc. The present invention considers fusing the results of multiple semantic segmentation models, making full use of the advantages of different models in different scenarios, overcoming the limitations of a single model, and effectively enhancing the accuracy of the prediction results. In some preferred embodiments, the following models are considered for constructing a fused semantic segmentation model:

[0043] 1. Swin Transformer model.

[0044] The Swin Transformer model is based on the Transformer architecture and can learn global and local features through its sliding window mechanism. This mechanism enables it to process large-scale high-resolution image data, and is particularly suitable for modeling long-distance dependencies in complex scenes, such as global feature extraction of objects such as buildings, roads, and water bodies. Swin Transformer effectively reduces computational complexity through layered design and downsampling technology, and is particularly suitable for the detection of large-scale planar geographic entities. It can greatly improve the reasoning efficiency of the model while ensuring accuracy. Through multi-level Patch Merging, Swin Transformer can extract features of different resolutions layer by layer, improving the recognition ability of complex objects.

[0045] 2. OCRNet model.

[0046] The core idea of ​​OCRNet is to enhance the contextual information of objects. The model transforms the pixel classification problem into the object region classification problem. By calculating the similarity between the object region representation and the pixel feature representation, it can more accurately segment objects with complex shapes and blurred boundaries, such as forests and cultivated land. OCRNet has good processing of boundary information and can enhance the boundary segmentation accuracy of objects. It is suitable for scenes that require refined segmentation, such as building outlines or road boundaries.

[0047] 3. HRNet model.

[0048] The HRNet (High-Resolution Network) model is able to maintain high-resolution feature representation throughout the network's reasoning process. This feature enables the model to accurately capture the details of objects in space, and has great advantages for accurately locating object boundaries and scenes with rich details (such as the segmentation of water bodies and buildings). HRNet connects high-resolution and low-resolution subnetworks in parallel and repeatedly fuses multi-scale features, which allows it to have strong semantic expression capabilities without losing spatial information. This allows HRNet to ensure both semantic strength and accuracy of spatial resolution when processing planar geographic entities.

[0049] This embodiment selects the above three models to construct a fusion semantic segmentation model based on the following technical considerations:

[0050] 1. Combination of global and local information: Swin Transformer is good at extracting global information, while HRNet maintains high-resolution local details, and OCRNet can enhance regional context information. The fusion of the three can integrate global and local features, greatly improving the model's adaptability to complex scenes.

[0051] 2. Complementarity of different types of objects: Different models perform well when dealing with different types of objects. For example, Swin Transformer has strong recognition capabilities for large-scale, long-distance dependent objects (such as roads and water bodies), HRNet has advantages in fine segmentation (such as building boundaries), and OCRNet performs well in fuzzy boundaries and complex shapes (such as forests and cultivated land). Through model fusion, the overall segmentation accuracy of various types of objects can be improved.

[0052] 3. Enhance the robustness of the model: By fusing the output results of different structural models, the present invention can improve the robustness of the model and reduce the false detection and missed detection problems caused by data deviation or scene complexity of a single model.

[0053] In the present invention, the prediction results (probability maps) output by the fusion model are adopted in a back-end fusion manner. The additive fusion method includes: superimposing the prediction probability maps output by all semantic segmentation models, and if the pixel value after superposition is greater than the preset fusion threshold, the value of the pixel is set to 1, otherwise it is set to 0; wherein 1 is the prediction category and 0 is the background. The additive fusion method integrates the prediction results of multiple models by accumulating the prediction probability maps of multiple semantic segmentation models or change detection models. Compared with the output of a single model, this method can effectively integrate the advantages of different models in different scenarios, make up for the shortcomings of a single model in certain scenarios, and thus improve the overall prediction accuracy. When processing complex planar geographic entities, different models may have different prediction advantages in specific land object categories. Through the additive fusion method, the prediction results of multiple models can be integrated, and their advantages can be taken, thereby improving the detection accuracy of multiple land object categories. Compared with other complex model fusion methods (such as decision-making fusion or voting method), the additive fusion method is simple to operate, has low computational cost, and is easy to implement. It only needs to add and process the prediction probability maps output by different models, and does not require additional complex operations or a large amount of computing resources. Therefore, in practical applications, multi-model fusion can be achieved quickly and efficiently, which is suitable for the rapid processing of large-scale data.

[0054] S2. Construct and train at least two change detection models set in parallel and use an additive fusion method to fuse the prediction results of each change detection model as a fused change detection model of change prediction results.

[0055] Mainstream change detection models in this field include, but are not limited to, Faster R-CNN, STANet, SNUNet, CDNET, P2V, and FCCDN. Similar to the reason for building a fusion semantic segmentation model, the present invention integrates the characteristics of multiple change detection models to improve the accuracy and robustness of change detection, adapt to complex and diverse surface geographic entity change detection needs, and achieve an efficient and fast detection process. In some preferred embodiments, consider selecting the following models to build a fusion change detection model:

[0056] 1. STANet model.

[0057] STANet (Spatial-Temporal Attention Network) introduces a spatiotemporal attention mechanism that can more accurately model the spatial and temporal features of different time series images. By calculating the correlation between any two pixels in a time series image, STANet can more effectively capture the changed area. This is very useful for handling change detection tasks in different time periods and complex scenes, such as the addition of buildings or changes in cultivated land. Through its spatiotemporal attention mechanism, STANet can accurately identify small changes in images. This feature gives it an advantage in scenes with small or subtle changes, and can effectively improve the accuracy of change detection and reduce false detections and missed detections.

[0058] 2. BIT model.

[0059] The BIT (Big Transfer) model is based on the ResNet-v2 structure. It has undergone a lot of pre-training and has enhanced its performance in change detection through transfer learning. It uses optimization techniques such as Group Normalization and WeightStandardization to improve the model's feature extraction capabilities in change detection tasks. By introducing large-scale feature transfer technology, the BIT model can better extract the global change features of images and is suitable for processing large-scale and complex structural changes in land objects, such as urban expansion and road changes. Through feature transfer technology, BIT can effectively cope with data domain differences across time periods and regions, and improve the generalization ability of the model.

[0060] This embodiment selects the above two models to construct a fusion change detection model based on the following technical considerations:

[0061] 1. Complementarity of local and global features: STANet is good at capturing subtle local changes, especially in areas of inconspicuous changes, such as changes in cultivated land and changes in water boundaries, and has high recognition accuracy. The BIT model is good at identifying large-scale changes in objects from a global perspective, especially in the detection of changes in large-scale objects such as urban buildings and roads. The combination of the two can achieve comprehensive coverage of different types of change scenarios, thereby improving the overall robustness and accuracy of change detection.

[0062] 2. Adaptability to different detection scenarios: In the change detection of planar geographic entities, the temporal span of image data and the complexity of the scene vary greatly. The combination of STANet's ability to capture details of short-term changes and BIT's global perception of long-term changes can meet the needs of change detection in various complex scenarios and ensure adaptability in different detection scenarios.

[0063] 3. Strong cross-domain migration capability: STANet can adapt to images at different times and under different lighting conditions through the spatiotemporal attention mechanism, especially for change detection tasks with large inter-domain differences. BIT has strong cross-domain feature migration capabilities through large-scale pre-training, and can maintain high detection accuracy under conditions of different data sources and image resolutions. The combination of the two effectively improves the generalization ability of the fusion change detection model.

[0064] 4. Enhance the stability of change detection: By integrating the STANet and BIT models, we can fully utilize the advantages of the two models in different scenarios, reduce the errors that may occur in a single model, and improve the stability and accuracy of overall change detection. The additive fusion rule can smooth the prediction results between different models and further improve the accuracy of detection.

[0065] It should be understood that, under normal circumstances, the data to be detected and the model training data have different geographical environments, image sources, time phases, spectral characteristics, etc. There will be a large domain gap between the two types of data, which will seriously affect the reasoning accuracy of the convolutional neural network model, thereby reducing the accuracy of ground surface entity detection. Therefore, in addition to the conventional neural network training method, the present invention also provides a better model training method, which specifically includes:

[0066] A sample data set with prior information is obtained to complete the preliminary training of the fused semantic segmentation model and the fused change detection model. The present invention performs preliminary training by using a remote sensing image sample library containing prior information. The sample data set covers a rich range of ground object categories, different image sources, seasons, regions and other information. Therefore, during the preliminary training, the model can fully learn the diverse characteristics of planar geographic entities and lay a solid foundation. The use of such prior information enables the model to have a strong generalization ability in the early stage and can cope with different scenarios and data sources. The data in the prior information sample library comes from a wide range of sources, covering a variety of ground object types (such as buildings, roads, cultivated land, water bodies, etc.), and the data comes from different remote sensing sensors and different time periods. Therefore, this preliminary training based on multi-source data makes the model more robust when processing heterogeneous data and can reduce errors caused by differences in data sources.

[0067] The data set to be detected and the sample data set are mixed as a new training sample set to complete the fine-tuning training of the fused semantic segmentation model and the fused change detection model. After the initial training of the model, the data to be detected is mixed with the sample data set to form a new training sample set. By mixing the data to be detected with the high-precision sample data set, the model can better adapt to the data characteristics in the actual detection task during the fine-tuning process, and further improve the detection accuracy of the model and the adaptability to new data. Through mixed training, the model of the present invention can be adaptively adjusted in the features of land objects in different land types, different regions and different time periods, thereby improving the generalization ability of the model in complex scenarios. Compared with using a pre-trained model alone, this mixed training method enables the model to better cope with the uncertainty and data diversity in detection tasks.

[0068] S3. Obtain image data of the area to be detected and the planar geographic entity to be detected, and obtain the data set to be detected and the land class label set after preprocessing.

[0069] A planar geographic entity refers to a division of different areas of the earth's surface into several independent plots or regions according to land use or land cover types in remote sensing images or geographic information systems (GIS). These plots or regions have relatively consistent attributes or characteristics, such as buildings, cultivated land, water bodies, roads, etc. Planar geographic entities are usually used in the fields of natural resource monitoring, land use planning, environmental protection, public services, etc. to classify and manage land objects. The core of the present invention is that by comparing new image data with existing planar geographic entities, changes or anomalies in the category of objects within the geographic entity can be discovered. Planar geographic entities serve as a reference benchmark in this process, which is used to measure whether the image data conforms to the expected land distribution. For example, a geographic entity is classified as cultivated land, but in the new image data, the area may have been transformed into a building, and this change will be identified as an anomaly.

[0070] It should be understood that since the resolution and size of the image data of the area to be detected and the area geographic entity to be detected are different, and the area geographic entity to be detected is vector format data, preprocessing is required to ensure the quality and consistency of the input data. Preprocessing methods commonly used in this field include denoising, standardization and clipping. In some preferred embodiments, a specific preprocessing method is provided, including: rasterizing the area geographic entity to be detected, and cutting it together with the image data of the area to be detected into a number of image blocks and label blocks consistent with the element resolution and size in the sample data set to form the data set to be detected and the land class label set. Subsequently, the segmented image blocks of the data to be detected are mixed with high-precision training data of equal proportion to reduce the proportion of erroneous labels in the data to be detected. The segmented label blocks can be more convenient for the subsequent comparison steps.

[0071] Furthermore, the brightness and color of images obtained by different phases, weather, and sensors vary greatly, which seriously affects the reasoning ability of the model during quality inspection. The images used for training, verification, and reasoning should generally be normalized. However, in order to further reduce the impact of the color differences of different images on model reasoning, in some preferred embodiments, it is also necessary to perform image conversion on the image data of the area to be detected and / or the sample data set to improve the consistency of color and brightness between images, reduce the difference between cross-domain images, and thus improve the reasoning accuracy of the model, which specifically includes the following steps:

[0072] Select several images containing the planar geographic entity to be detected from the image data of the area to be detected and / or the sample data set as target images, convert the target images to the LAB color space and calculate the average value and variance, convert the image data of the area to be detected and / or the sample data set to the LAB color space, and perform image conversion using the following formula:

[0073] ;

[0074] in, and are the mean and variance of the target image, and are respectively the image data of the area to be detected and\or the mean value and variance of the sample data set, and are respectively the image data of the area to be detected and\or the original value and the converted value of the image of the sample data set.

[0075] S4. Input the data set to be detected into the fused semantic segmentation model and the fused change detection model respectively to obtain a classification prediction result and a change prediction result.

[0076] like Figure 3 As shown, in some preferred embodiments, 5 target images are selected from the data set to be detected, and the building roofs are extracted using a separate Swin Transformer model, an OCRNet model, an HRNet model and the fused semantic segmentation model proposed in the present invention (marked as Union in the figure). It can be seen that the building roofs obtained by the fused semantic segmentation model proposed in the present invention are basically consistent with the true value, and are more accurate than the results predicted by the separate HRNet, Swin and OCRNet models.

[0077] like Figure 4As shown, in some preferred embodiments, 5 groups of chronological images are selected from the data set to be detected, and change detection is performed using the BIT model, the P2V model and the fusion change detection model proposed in the present invention (marked as Union in the figure). It can be seen that the building change prediction map obtained by the fusion change detection model proposed in the present invention is basically consistent with the actual change value, and has higher accuracy than the results predicted by the separate BIT and P2V models.

[0078] S5. Use the classification prediction result to assign category attributes to the change prediction result to obtain a fusion prediction result.

[0079] In this step, the classification prediction results characterize the classification prediction results of each pixel in the image data, and these results contain the classification information of the ground objects (such as buildings, roads, cultivated land, etc.). The change prediction results mark which areas in the image data have changed, but these changed areas themselves do not have specific category information. Assigning these classification information to the change prediction results means that the changed areas are specifically labeled with categories. For example, if an area is marked as a changed area in change detection, and the classification prediction results show that the area has changed from cultivated land to a building, then through step S5, the category attribute of "building" can be assigned to the changed area. The fused prediction results are the basis for anomaly detection and comparison in subsequent steps. By assigning category attributes, the system can compare the fused prediction results with the existing land class label set to determine whether there are any abnormalities (for example, if an area was originally cultivated land, but is now classified as a building, it may indicate that an abnormal change has occurred in the area).

[0080] S6. Compare the fusion prediction result with the elements with the same name in the land class label set to complete anomaly detection.

[0081] It should be understood that the fusion prediction results may include some tiny surface entities. Such surface entities that are obviously inconsistent with the actual situation are interference items in monitoring. Therefore, in some preferred embodiments, it is considered to filter by setting a surface entity area threshold, specifically including: removing the fusion prediction results whose area is smaller than the preset area threshold.

[0082] The same-name elements refer to elements that represent the same spatial area or category in the fusion prediction results and the land class label set (i.e., the existing classification standards or benchmark data). For example, if a surface entity is marked as "cultivated land" in the existing land class label set, and the area is still "cultivated land" in the fusion prediction results, then this part of the surface entity is called a "same-name element". Another purpose of the comparison is to verify whether the fusion prediction results of the model are consistent with the existing land class label data, and to ensure that the model's classification prediction of the changed area is accurate. By comparing the same-name elements, it is possible to verify whether the prediction results match the actual situation, which helps to further optimize the accuracy of the model. The core of the comparison is to check whether the surface geographic entity meets the standards of the same-name element by matching the spatial position and category attributes. Generally, whether two elements are same-name elements is determined by calculating the spatial overlap and category consistency. If two surface entities are highly overlapped in space and have the same category attributes, they are same-name elements. In some preferred embodiments, a specific method for confirming elements with the same name is provided, including: traversing each label block in the fusion prediction result and calculating the overlap, and taking the label blocks with the overlap higher than a preset overlap threshold as elements with the same name.

[0083] In some other preferred embodiments, a device is also provided, which is used to implement the above-mentioned anomaly detection method for planar geographic entities.

[0084] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for detecting anomalies of planar geographic entities, characterized in that: Includes steps: Constructing and training at least two semantic segmentation models arranged in parallel and fusing the prediction results of the fused semantic segmentation models using an additive fusion method as a fused semantic segmentation model of classification prediction results; Constructing and training at least two change detection models arranged in parallel and fusing the prediction results of the change detection models using an additive fusion method as a fused change detection model of change prediction results; Obtain image data of the area to be detected and the planar geographic entity to be detected, and obtain the data set to be detected and the land class label set after preprocessing; Inputting the to-be-detected data set into the fused semantic segmentation model and the fused change detection model respectively to obtain a classification prediction result and a change prediction result; Assigning category attributes to the change prediction result using the classification prediction result to obtain a fusion prediction result; The fusion prediction result is compared with the elements with the same name in the land class label set to complete anomaly detection.

2. The method for detecting anomalies of planar geographic entities according to claim 1, characterized in that: The training method of the fused semantic segmentation model and the fused change detection model includes: Acquiring a sample data set with prior information to complete preliminary training of the fused semantic segmentation model and the fused change detection model; The to-be-detected data set and the sample data set are mixed as a new training sample set to complete the fine-tuning training of the fused semantic segmentation model and the fused change detection model.

3. The method for detecting anomalies of planar geographic entities according to claim 2, characterized in that: The preprocessing method of the image data of the area to be detected and the planar geographic entity to be detected includes: The planar geographic entity to be detected is rasterized and, together with the image data of the area to be detected, is divided into a number of image blocks and label blocks that are consistent with the element resolution and size in the sample data set to form the data set to be detected and the land class label set.

4. The method for detecting anomalies of planar geographic entities according to claim 3, characterized in that: The method for determining the elements with the same name includes: Each label block in the fusion prediction result is traversed and the overlap degree is calculated respectively, and the label blocks with the overlap degree higher than the preset overlap threshold are regarded as the elements with the same name.

5. The method for detecting anomalies of planar geographic entities according to claim 1, characterized in that: Before comparing the fused prediction results with the elements with the same name in the land class label set, the method further includes removing the fused prediction results whose areas are smaller than a preset area threshold.

6. The method for detecting anomalies of planar geographic entities according to claim 2, characterized in that: Performing image conversion on the image data of the area to be detected and / or the sample data set comprises the steps of: Select several images containing the planar geographic entity to be detected from the image data of the area to be detected and / or the sample data set as target images, convert the target images to the LAB color space and calculate the average value and variance, convert the image data of the area to be detected and / or the sample data set to the LAB color space, and perform image conversion using the following formula: ； in, and are the mean and variance of the target image, and are respectively the image data of the area to be detected and\or the mean value and variance of the sample data set, and are respectively the image data of the area to be detected and\or the original value and the converted value of the image of the sample data set.

7. The method for detecting anomalies of planar geographic entities according to claim 1, characterized in that: The additive fusion method includes: superimposing the predicted probability maps output by all semantic segmentation models, and if the superimposed pixel value is greater than a preset fusion threshold, setting the value of the pixel to 1, otherwise it is set to 0; wherein 1 is the predicted category and 0 is the background.

8. An anomaly detection device for planar geographic entities, characterized in that: The device is used to implement the method according to any one of claims 1 to 7.

Images