Vector and image data consistency detection method based on prior knowledge guidance

Abstract

The present application relates to surveying and mapping geographic information data processing and artificial intelligence technical field, disclose a kind of vector and image data consistency detection method based on prior knowledge guidance, comprising: obtaining the vector data and corresponding image data to be detected, generate prior knowledge layer;Prior knowledge layer and the original wave band of image data are fused to generate fused image;Provide multiple consistency detection models;Fused image is respectively input into multiple consistency detection models, so that each consistency detection model is based on the prior knowledge in fused image, identifies and outputs the abnormal area that it is respectively judged as content inconsistency;The abnormal area outputted by multiple consistency detection models is fused and processed, to generate final consistency detection result.The present application realizes full-process automation, improves detection efficiency, shortens data quality inspection period, saves a lot of manpower cost, and improves the overall quality level of geospatial data product.

Description

Technical Field

[0001] The present invention relates to the field of geographic information processing and artificial intelligence technology, and in particular to a method for detecting consistency between vector and image data based on prior knowledge guidance. Background Art

[0002] Geospatial data, particularly vector geographic entity data (or survey and monitoring patches) and high-resolution remote sensing imagery (such as digital orthophotos (DOMs)) as the digital foundation, are key to supporting numerous applications related to national economy and people's livelihoods, including natural resource planning and management, land and space monitoring, urban construction, and emergency response. The accuracy, currency, and completeness of data—in other words, data quality—directly determine its application value.

[0003] In the practice of data production and updating, inconsistencies between vector data and contemporaneous imagery are inevitable due to the complex and variable nature of surface features, differences in experience between field and field personnel, operational oversights, and limitations in data processing workflows. These inconsistencies primarily manifest as: incorrect feature category attributes (e.g., a house already built in the image may be labeled "arable land" in the vector data), missed feature elements (e.g., a new road appears in the image but is missing in the vector data), and redundant feature elements (e.g., a building labeled in the vector data has been demolished in the image).

[0004] To ensure data quality, traditional data quality inspection processes rely heavily on manual visual inspection and sampling. On a computer screen, quality inspectors overlay vector layers on top of image layers, visually comparing and identifying inconsistencies. The drawbacks of this approach are becoming increasingly apparent:

[0005] 1. Inefficiency and high cost: Faced with massive amounts of data covering an entire city or even a province, manual inspection is time-consuming and labor-intensive, requiring a large investment of manpower costs and a long project cycle.

[0006] 2. High subjectivity and inconsistent standards: Inspection results are affected by the personal experience, professional level and even mental state of quality inspectors. Different people may make different judgments on some ambiguous situations, making it difficult to unify quality inspection standards and resulting in poor consistency in results.

[0007] 3. Low coverage and significant hidden dangers: Due to cost and time constraints, quality inspections typically rely on sampling, making it impossible to conduct 100% data coverage checks. This means that a large number of potential errors in areas not sampled will be missed, creating serious data quality risks for subsequent applications.

[0008] In recent years, with the development of artificial intelligence (AI), some technologies have attempted to leverage computer vision models (such as object detection and semantic segmentation) to automatically extract features from remote sensing imagery. However, these methods typically rely on "starting from scratch" recognition, aiming to "generate" new vector data rather than "verify" the consistency between existing vector data and imagery. They often lack effective mechanisms to integrate and leverage the rich "prior knowledge" contained in existing vector data to directly and efficiently locate conflicts and contradictions between the two types of data.

[0009] Therefore, this field urgently needs a new data consistency detection method that is efficient, objective, and comprehensive and can overcome the limitations of traditional manual quality inspection and existing automation technology. Summary of the Invention

[0010] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a technical solution for automatically and intelligently detecting the consistency between vector geographic data describing the categories and boundaries of land objects and remote sensing image data reflecting the actual conditions of the surface through a deep neural network.

[0011] In order to achieve the above-mentioned object of the invention, the technical solution provided by the present invention includes:

[0012] A method for detecting consistency between vector and image data based on prior knowledge guidance includes the following steps:

[0013] Obtain the vector data to be detected and the corresponding image data, rasterize the vector data according to its category attributes, and generate a priori knowledge layer;

[0014] The prior knowledge layer is used as at least one new band to fuse with the original band of the image data to generate a fused image containing the prior knowledge of the category attributes of the vector data;

[0015] Providing multiple consistency detection models trained using a sample library, wherein the sample library contains consistent samples and inconsistent samples constructed by modifying vector data categories;

[0016] Inputting the fused image into the plurality of consistency detection models respectively, so that each consistency detection model identifies and outputs an abnormal region determined by each consistency detection model as having inconsistent content based on prior knowledge in the fused image;

[0017] The abnormal regions respectively output by the plurality of consistency detection models are fused to generate a final consistency detection result.

[0018] Preferably, the method for generating the fused image includes: adding the prior knowledge layer to the original band of the image data, and then performing registration according to the geographical location.

[0019] Preferably, the plurality of consistency detection models include at least two types selected from the following group: a semantic segmentation model, a target detection model, and an industrial anomaly detection model.

[0020] Preferably, the training of the plurality of consistency detection models further includes: using the output results of each training or each detection to produce fine-tuning samples, and fine-tuning the plurality of consistency detection models.

[0021] Preferably, the category attributes include construction land related feature categories and non-construction land related feature categories.

[0022] Preferably, the fusion processing method includes: determining a final consistency detection result based on the position overlap and / or area size between abnormal regions output by a plurality of consistency detection models.

[0023] Preferably, the ratio of consistent samples to inconsistent samples contained in the sample library is 1:1.

[0024] Beneficial effects

[0025] 1. This invention utilizes an end-to-end deep learning model to automate the entire process, from data fusion to outlier area output. Compared to traditional manual visual inspection, this significantly improves detection efficiency, shortens the data quality inspection cycle, and saves significant labor costs. Furthermore, it enables comprehensive automated inspection of data within the inspection area, rather than traditional spot checks, improving the overall quality of geospatial data products. This invention can be used not only for consistency monitoring of vector and image data but also for change detection.

[0026] 2. The present invention uses a detection model trained by data-driven testing to conduct testing, which completely eliminates the problems of inconsistent judgment standards and fluctuating results caused by subjective factors (such as experience and fatigue) in manual inspection, ensuring the objectivity, consistency and reproducibility of quality inspection results.

[0027] 3. This invention uniquely incorporates the categorical attributes of vector data into the image data as prior knowledge bands, guiding the model to directly learn specific patterns of inconsistency. This enables the model to not only detect simple geometric deviations but also accurately identify deeper semantic conflicts such as mislabeled feature categories. Furthermore, by expanding the training sample library with inconsistent samples, it addresses the common problem of scarcity and difficulty in collecting abnormal sample data in the real world. This provides a solid data foundation for the effective training of detection models, demonstrating its strong practicality and potential for widespread adoption. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1Schematic diagram of the flow of a method for detecting consistency between vector and image data based on prior knowledge provided in a preferred embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of the mapping between category attributes and commonly used feature category names provided in another preferred embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] In one embodiment of the present invention, the complete process of the method for detecting consistency between vector and image data based on prior knowledge is as follows: Figure 1 As shown in Figure 1, it mainly includes several key stages, including data fusion, model training, anomaly detection and result fusion.

[0032] S1. Obtain the vector data to be detected and the corresponding image data, rasterize the vector data according to its category attributes, and generate a priori knowledge layer.

[0033] The purpose of this step is to integrate discrete, object-based vector data with pixel-based image data to generate a fused data that can be directly processed by a deep learning model and contains prior knowledge. The vector data is a geographic feature with clear geographic coordinates and category attributes. For example, in a land survey application, a polygonal vector patch may be assigned the category attribute of "construction land" or "arable land". The image data is a contemporaneous remote sensing image that can truly reflect the current condition of the surface. Specifically, in an application scenario, the vector data can be a survey and monitoring patch or geographic entity data in the shapefile format obtained from a natural resources survey and monitoring database. The corresponding image data can be a digital orthophoto (DOM) of the area at the same time phase obtained from the surveying and mapping department.

[0034] Rasterization is the process of converting vector structured data into a raster (i.e., pixel grid) structure. In the present invention, this process is not a simple graphic conversion, but rather converts different category attributes into different, fixed pixel values ​​based on preset mapping rules. For example, in a preferred embodiment, the category attributes may include at least the category of land features related to construction land and the category of land features related to non-construction land. Figure 2As shown, the category attributes can be preset by establishing a mapping table between the category attributes and the names of commonly used land features. Furthermore, by uniformly assigning the pixel values ​​of the corresponding area after rasterization, the generated single-channel grayscale image, that is, the prior knowledge layer, carries the core semantic information of the original vector data in the form of pixels. It should be understood that the raster assignment can be any value between 1-254, but there must be a certain difference between the label values ​​of each category to make it easier to distinguish after normalization. For example, the raster value of construction land can be 192, and the raster value of non-construction land can be 64.

[0035] S2. Taking the prior knowledge layer as at least one new band, the layer is fused with the original band of the image data to generate a fused image containing the prior knowledge of the category attributes of the vector data.

[0036] If the original remote sensing image is a color image containing three bands: red, green, and blue, the fused image will become a four-band image, where the first three bands are the original visual information, and the newly added fourth band is the above-mentioned prior knowledge layer that carries the category attributes.

[0037] To ensure fusion accuracy, in one specific embodiment, the fusion step further includes: adding the prior knowledge layer as at least one new band to the original band of the image data, and then performing registration based on geographic location. Registration is a standard operation in surveying and mapping geographic information data processing. It aims to ensure that every pixel in different layers (bands) is precisely aligned in space through techniques such as coordinate transformation and resampling. That is, pixels at the same location in different bands correspond to the exact same real-world geographic location.

[0038] S3. Provide multiple consistency detection models trained using a sample library, wherein the sample library contains consistent samples and inconsistent samples constructed by modifying vector data categories.

[0039] It should be understood that the present invention adopts a multi-model integration strategy, aiming to utilize the complementary advantages of different types of models to improve the robustness and accuracy of the detection results. In a preferred embodiment, the plurality of consistency detection models include at least two types selected from the following groups: semantic segmentation model, target detection model, and industrial anomaly detection model. For example, a semantic segmentation model such as HRNet or DeepLab V3+ can be used in combination to accurately outline the pixel-level contours of the abnormal area; at the same time, a target detection model such as the YOLO series or Faster R-CNN can be used in combination to quickly locate the position of the abnormal target (output bounding box); an industrial anomaly detection model such as STFPM or GANomaly can also be used in combination to further locate the abnormal target and output its position information to provide a basis for subsequent abnormal result judgment. The collaborative work of different models can capture abnormal features from different dimensions, effectively reducing the false positive and false negative rates of a single model.

[0040] The key to training the aforementioned detection model lies in the sample library it utilizes. A key feature of this library is that it not only contains consistent samples (i.e., samples whose vector patch categories match the imaged features), but more importantly, it also includes inconsistent samples constructed by modifying the vector data categories. This addresses the scarcity and difficulty of collecting outlier samples (i.e., inconsistent samples) in the real world. Specifically, these samples can be constructed by randomly and proportionally modifying the category attributes of otherwise correct vector patches. For example, a portion of patches originally labeled "construction land" can be intentionally modified to "non-construction land" when creating training samples, thereby artificially creating samples whose vector data is inconsistent with the image data. For example, a patch with a house in an image can be modified to "cultivated land" when its vector attribute should be "construction land," thereby creating an inconsistent sample. Inconsistent samples can also be generated by collecting survey and monitoring data or identifying quality issues such as missed or redundant data collection during quality inspection of geographic entity results, using the aforementioned steps. In a preferred embodiment, to ensure balanced model training, the ratio of consistent to inconsistent samples in the sample library can be set to approximately 1:1. By training on this carefully constructed sample library containing a large number of positive and negative samples, the model can efficiently learn the deep discriminative rules between "consistency" and "inconsistency."

[0041] In order to further improve the performance of the model on specific detection tasks, an optional enhancement step is to fine-tune the model, specifically including: using the output results of each training or each detection to create fine-tuning samples, and fine-tuning the multiple consistency detection models. This can be understood as an iterative optimization or active learning mechanism. For example, after each training or each detection, the areas that the model determines to be abnormal with high confidence can be added to the sample library as high-quality negative samples after manual confirmation; or, for areas that are difficult for the model to judge and have low confidence, they can be submitted to manual labeling, and the labeled results can be used as new training data. In this way, the output results of the model itself are used to continuously expand and optimize the training set, which enables the model to self-evolve in continuous application and dynamically adapt to new data characteristics and more complex abnormal patterns.

[0042] S4. Input the fused image into the plurality of consistency detection models respectively, so that each consistency detection model identifies and outputs abnormal regions determined to be inconsistent in content based on prior knowledge in the fused image.

[0043] Consider an input fused image where the fourth band (the prior knowledge band) has a pixel value of 64 (indicating "non-construction land") in a certain area, but the first three bands (the image bands) clearly indicate that the area is a newly built factory. The model has learned during training that this pattern is inconsistent. Therefore, the semantic segmentation model might output a precise pixel mask of the factory area as an anomaly, while the object detection model might output a bounding box enclosing the factory as an anomaly.

[0044] S5. The abnormal areas outputted by the multiple consistency detection models are fused to generate a final consistency detection result. This step takes into account that different models may identify abnormalities from different angles, and their output results may be different, thereby obtaining a unified and reliable final conclusion. In a specific embodiment, the fusion processing method includes: determining the final consistency detection result based on the position overlap and / or area size between the abnormal areas outputted by the multiple consistency detection models. For example, a rule can be set: only when the overlapping area of ​​the abnormal areas outputted by at least two models at the same position exceeds a certain threshold (such as 80%), the area is confirmed as the final abnormality; or, all small abnormal areas with an area less than a certain threshold (such as 1000 square meters) can be filtered out to focus on major and meaningful changes. This rule-based fusion decision can effectively integrate multi-source information and improve the confidence of the detection results.

[0045] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting consistency between vector and image data based on prior knowledge, characterized in that: Including steps: Obtain the vector data to be detected and the corresponding image data, rasterize the vector data according to its category attributes, and generate a priori knowledge layer; The prior knowledge layer is used as at least one new band to fuse with the original band of the image data to generate a fused image containing the prior knowledge of the category attributes of the vector data; Providing multiple consistency detection models trained using a sample library, wherein the sample library contains consistent samples and inconsistent samples constructed by modifying vector data categories; Inputting the fused image into the plurality of consistency detection models respectively, so that each consistency detection model identifies and outputs an abnormal region determined by each consistency detection model as having inconsistent content based on prior knowledge in the fused image; The abnormal regions respectively output by the plurality of consistency detection models are fused to generate a final consistency detection result.

2. The method for detecting consistency between vector and image data based on prior knowledge as claimed in claim 1, wherein: The step of fusing the prior knowledge layer with the original band of the image data specifically includes: adding the prior knowledge layer as at least one new band to the original band of the image data, and then performing registration according to the geographic location.

3. The method for detecting consistency between vector and image data based on prior knowledge as claimed in claim 1, wherein: The plurality of consistency detection models include at least two types selected from the following group: a semantic segmentation model, a target detection model, and an industrial anomaly detection model.

4. The method for detecting consistency between vector and image data based on prior knowledge as claimed in claim 1, wherein: The training of the plurality of consistency detection models further includes: using the output results of the training or detection to create fine-tuning samples, and fine-tuning the plurality of consistency detection models.

5. The method for detecting consistency between vector and image data based on prior knowledge as claimed in claim 1, wherein: The category attributes include construction land related feature categories and non-construction land related feature categories.

6. The method for detecting consistency between vector and image data based on prior knowledge as claimed in claim 1, wherein: The fusion processing method includes: determining a final consistency detection result based on the position overlap and / or area size between abnormal regions output by multiple consistency detection models.

7. The method for detecting consistency between vector and image data based on prior knowledge as claimed in claim 1, wherein: The ratio of consistent samples to inconsistent samples included in the sample library is 1:1.

Images