Urban green land invasion and degradation identification system and method based on computer visual identification

Abstract

The invention provides an urban green land invasion and occupation degradation identification system and method based on computer vision identification, and the method comprises the steps: collecting multi-temporal images of an urban green land, sorting the images according to timestamps, and generating a difference atlas to reflect the dynamic change of the green land; carrying out vectorization processing on the dynamic information, establishing a digital model as a reference standard of a greenbelt boundary, and carrying out superposition comparison with the latest remote sensing image to obtain a corrected greenbelt boundary; using the target detection network to identify a non-vegetation coverage area in the multi-temporal image, generating a preliminary green land invasion location map, then performing shape regularization and size filtering, and optimizing the green land invasion location map; and based on the optimization graph, time and space distribution characteristics of each factor are analyzed, prediction is carried out in combination with the change trend of the non-vegetation coverage area, and a trend prediction result is generated. And finally, carrying out quantitative evaluation on a prediction result, and generating a comprehensive degradation severity report. According to the invention, the identification efficiency and identification precision of urban green land invasion and degradation are improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer vision recognition technology, and in particular to a system and method for identifying encroachment and degradation of urban green space based on computer vision recognition. Background Art

[0002] Against the backdrop of rapid urbanization, urban green spaces, as an important component of improving environmental quality and enhancing residents' quality of life, are facing serious problems of encroachment and degradation.

[0003] In order to effectively monitor these changes, traditional monitoring methods mainly include field surveys, remote sensing image analysis, and analysis of geographic information systems combined with historical data.

[0004] However, field surveys rely on manual visit records, which can provide intuitive information, but are inefficient and difficult to cover large areas; remote sensing image analysis uses data collected by satellites or drones to monitor green space changes. Although it can expand the monitoring range, the manual annotation and simple classification algorithms in traditional methods have a low degree of automation and are difficult to accurately handle changes in complex environments; and historical data analysis based on GIS can provide some reference, but it has limitations in capturing real-time changes. In general, the traditional identification of urban green space encroachment and degradation has the problem of low efficiency. Summary of the invention

[0005] The embodiments of the present application provide a system and method for identifying encroachment and degradation of urban green space based on computer vision recognition, so as to solve the problem of low efficiency in identifying encroachment and degradation of urban green space in the prior art.

[0006] In a first aspect, the present invention provides an urban green space encroachment and degradation identification system based on computer vision recognition, comprising:

[0007] The acquisition module (i.e., the UAV imaging detection platform) is used to periodically acquire multi-phase images of urban green spaces, sort the multi-phase images according to timestamps, generate difference maps between the images of each phase, and obtain dynamic information reflecting the changes of urban green spaces over time based on the difference maps between the images of each phase;

[0008] A processing module (i.e., a high-performance server) is used to vectorize the dynamic information reflecting the changes of green space over time based on the historical geographic information system records and the land boundary information stipulated by the law, so as to establish a digital model as a reference for the green space boundary, and to superimpose and compare the digital model with the latest remote sensing image obtained to obtain a corrected green space boundary. The latest remote sensing image refers to an image obtained by modern remote sensing technology at the most recent time point, which is used to provide the latest surface condition information;

[0009] an identification module, for identifying non-vegetation covered areas in the multi-temporal image based on the corrected green space boundary by using a deep learning target detection network to generate a preliminary green space encroachment location map, and performing shape regularization and size filtering processing on the preliminary green space encroachment location map by using morphological operations and connected domain analysis technology to generate an optimized green space encroachment location map, wherein the non-vegetation covered areas include artificial structures and hardened ground in the green space image;

[0010] A generation module is used to obtain the time and space distribution characteristics of each factor based on the optimized green space encroachment location map, using a spatiotemporal data analysis framework and combining environmental parameters, predict the time and space distribution characteristics of each factor in combination with the change trend of the non-vegetation covered area, generate trend prediction results, quantitatively evaluate the trend prediction results, and generate a degradation severity report. The environmental parameters include climate data and human activity intensity, and each factor refers to various variables or conditions that may affect green space encroachment and degradation.

[0011] In a second aspect, an embodiment of the present application provides a method for identifying urban green space encroachment and degradation based on computer vision recognition, comprising:

[0012] Periodically collect multi-phase images of urban green space, sort the multi-phase images according to timestamps, generate difference maps between the images of each phase, and obtain dynamic information reflecting the changes of urban green space over time based on the difference maps between the images of each phase;

[0013] Based on the historical geographic information system records and the land boundary information stipulated by the law, the dynamic information reflecting the changes of the green space over time is vectorized to establish a digital model as a reference for the green space boundary, and the digital model is superimposed and compared with the latest remote sensing image obtained to obtain a corrected green space boundary. The latest remote sensing image refers to an image obtained by modern remote sensing technology at the most recent time point to provide the latest surface condition information;

[0014] Based on the corrected green space boundary, a deep learning target detection network is used to identify non-vegetation covered areas in the multi-temporal image to generate a preliminary green space encroachment location map, and morphological operations and connected domain analysis techniques are used to perform shape regularization and size filtering on the preliminary green space encroachment location map to generate an optimized green space encroachment location map, wherein the non-vegetation covered areas include artificial structures and hardened ground in the green space image;

[0015] Based on the optimized green space encroachment location map, the spatiotemporal data analysis framework is used in combination with environmental parameters to obtain the temporal and spatial distribution characteristics of each factor, and the temporal and spatial distribution characteristics of each factor are combined with the changing trend of the non-vegetation covered area to perform forecast processing to generate trend forecast results, and the trend forecast results are quantitatively evaluated to generate a degradation severity report. The environmental parameters include climate data and the intensity of human activities, and each factor refers to various variables or conditions that may affect green space encroachment and degradation.

[0016] Optionally, based on the corrected green space boundary, a deep learning target detection network is used to identify non-vegetation covered areas in the multi-temporal image to generate a preliminary green space encroachment location map, and morphological operations and connected domain analysis techniques are used to perform shape regularization and size filtering on the preliminary green space encroachment location map to generate an optimized green space encroachment location map, including:

[0017] Based on the corrected green space boundary, a pre-trained deep learning target detection network is used in combination with a transfer learning algorithm to optimize model parameters, and each green space image in the multi-temporal image dataset is automatically identified for non-vegetation covered areas to obtain a preliminary green space encroachment location map;

[0018] According to the preliminary green space encroachment location map, the detected non-vegetation covered area is regularized by using morphological operation technology combined with a watershed algorithm to obtain a shape-regularized area, and the shape-regularized area is filtered by using a connected domain analysis technology combined with a distance transformation algorithm to generate an accurate connected domain that meets the actual green space encroachment standard;

[0019] The information of the accurate connected domain that meets the actual green space encroachment standard is comprehensively processed, and the spatial relationship between each connected domain in the accurate connected domain is evaluated using a spatial statistical analysis method to output a final green space encroachment location map.

[0020] Optionally, based on the corrected green space boundary, a pre-trained deep learning target detection network is used in combination with a transfer learning algorithm to optimize model parameters, and each green space image in the multi-temporal image dataset is automatically identified for non-vegetation covered areas to obtain a preliminary green space encroachment location map, including:

[0021] Using a deep learning target detection network suitable for urban green space scenes, preprocessing each green space image in the multi-temporal image dataset to generate prepared image data, wherein the preprocessing includes resizing, color space conversion, and normalization processing;

[0022] Obtain a small sample data set collected locally, use a transfer learning algorithm to fine-tune the pre-trained model, optimize the last layer or layers of the pre-trained model, adapt the pre-trained model to the data distribution of the new task, optimize the parameters of the pre-trained model, and obtain a target detection model optimized by transfer learning;

[0023] Applying the object detection model optimized by transfer learning, performing pixel-level classification and bounding box prediction on the prepared green space image, using non-maximum suppression technology to remove redundant detection boxes, retaining the most likely candidate areas, and generating an image marked with non-vegetation covered areas;

[0024] Based on the image marked with non-vegetation covered areas, a preliminary green space encroachment location map is formed, and the image marked with non-vegetation covered areas shows the location and scope of the suspected encroachment areas.

[0025] Optionally, applying the object detection model optimized by transfer learning to perform pixel-level classification and bounding box prediction on the prepared green space image, using non-maximum suppression technology to remove redundant detection boxes, retaining the most likely candidate areas, and generating an image marked with non-vegetation covered areas, includes:

[0026] Using the object detection model optimized by transfer learning, the prepared green space image is classified at the pixel level to obtain a probability map of each pixel belonging to the background or non-vegetation covered area;

[0027] Based on the probability map of each pixel belonging to the background or non-vegetation covered area, the detected non-vegetation covered area is obtained by using the regression information output by the target detection model, and a bounding box prediction is performed on the detected non-vegetation covered area to generate a series of candidate area bounding boxes including position information and confidence scores, wherein the position information includes center coordinates, width and height;

[0028] Based on the series of candidate region bounding boxes containing position information and confidence scores, applying non-maximum suppression technology, sorting according to the confidence scores, sequentially comparing the intersection-and-union ratios between adjacent bounding boxes of the candidate region bounding boxes, and retaining the one with the highest confidence for the bounding boxes whose intersection-and-union ratios are higher than a preset threshold, and removing other redundant detection boxes to obtain the retained bounding boxes;

[0029] The retained bounding box is drawn onto the periodic multi-temporal image dataset to form an image with non-vegetation covered areas marked.

[0030] Optionally, based on the optimized green space encroachment location map, the spatiotemporal data analysis framework is used in combination with environmental parameters to obtain the temporal and spatial distribution characteristics of each factor, the temporal and spatial distribution characteristics of each factor are combined with the change trend of the non-vegetation covered area to perform forecasting processing, and the trend forecasting results are quantitatively evaluated to generate a degradation severity report, including:

[0031] Using the optimized green space encroachment location map data, the information of the non-vegetation covered area is integrated and processed to obtain a data set suitable for a spatiotemporal data analysis framework;

[0032] Collecting and preprocessing environmental parameters according to the data set suitable for the spatiotemporal data analysis framework to generate multidimensional spatiotemporal data related to changes in non-vegetated areas;

[0033] Based on the multidimensional spatiotemporal data related to the changes in non-vegetation covered areas, the temporal and spatial distribution characteristics of each factor in the environmental parameters are analyzed and modeled to obtain a relationship model between different environmental parameters and the changes in non-vegetation covered areas;

[0034] Using the relationship model between the different environmental parameters and the change of the non-vegetation covered area, the change trend of the non-vegetation covered area is predicted and processed to generate trend prediction results for a period of time in the future;

[0035] Based on the trend prediction results for the future period, a quantitative assessment is conducted on the expansion speed of the non-vegetation covered area, the scope of influence and the degree of ecological impact of the non-vegetation covered area to generate a degradation severity report.

[0036] Optionally, the using of the relationship model between the different environmental parameters and the non-vegetation covered area change to predict the change trend of the non-vegetation covered area and generate trend prediction results for a period of time in the future includes:

[0037] Using the relationship model between the different environmental parameters and the changes in the non-vegetation covered area, selecting a suitable machine learning or statistical model, training the change data of the non-vegetation covered area, and constructing a prediction model that captures the temporal and spatial change rules;

[0038] According to the requirements of the prediction model that captures the temporal and spatial variation patterns, the pre-processed multi-dimensional spatiotemporal data are integrated, and the non-vegetation covered area variation data including the historical period and the latest observations are used as input to generate an input data set suitable for prediction;

[0039] Based on the input data set suitable for prediction, different future scenario assumptions are set to simulate the changing trends of different non-vegetation covered areas that may occur in the future, and generate multiple prediction scenarios covering the best case, the most likely case and the worst case. The different future scenario assumptions include different paths of climate change and changes in the intensity of human activities;

[0040] By using the multiple prediction scenarios covering the best case, the most likely case and the worst case, the short-term and long-term change trends of the non-vegetation covered area are predicted and processed to generate a series of trend prediction results at future time points.

[0041] Optionally, the dynamic information reflecting the change of green space over time is vectorized based on the land boundary information recorded in the historical geographic information system and stipulated by law, so as to establish a digital model as a reference for the green space boundary, and the digital model is superimposed and compared with the latest remote sensing image to obtain the corrected green space boundary, including:

[0042] Using historical geographic information system records and legally required land boundary information, vectorize the dynamic information reflecting the changes of green space over time to obtain vector data that accurately represents the changes of green space and the legal boundaries of green space changes;

[0043] Based on the vector data accurately representing the green space transition and the legal boundaries of the green space transition, a digital model is established and used as a reference for the green space boundary, converted into a digital format, and a high-precision digital model of the green space transition and boundary position is generated;

[0044] Acquire the latest remote sensing images with high resolution, high precision and high timeliness, make the coordinate system of the digital model consistent with that of the latest remote sensing images through image registration technology, and perform superposition and comparison processing to obtain the detection result of the green space boundary change;

[0045] Based on the detection result of the green space boundary change, combined with field investigation or other auxiliary data sources, the dynamic information reflecting the change of the green space over time is adjusted and corrected to generate a corrected green space boundary.

[0046] In a third aspect, an embodiment of the present application provides a computing device comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a method for identifying encroachment and degradation of urban green space based on computer vision recognition as described in any one of the first aspects.

[0047] In a fourth aspect, an embodiment of the present application provides a computer storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement a method for identifying encroachment and degradation of urban green space based on computer vision recognition as described in any one of the first aspects.

[0048] In the embodiment of the present application, multi-phase images of urban green space are periodically collected, the multi-phase images are sorted according to timestamps, and difference maps between the images of each phase are generated. Based on the difference maps between the images of each phase, dynamic information reflecting the changes of urban green space over time is obtained; according to the land boundary information recorded in the historical geographic information system and stipulated by law, the dynamic information reflecting the changes of green space over time is vectorized to establish a digital model as a reference for the boundary of the green space, and the digital model is superimposed and compared with the latest remote sensing image obtained to obtain the corrected green space boundary, wherein the latest remote sensing image refers to an image obtained by modern remote sensing technology at the most recent time point, which is used to provide the latest surface condition information;

[0049] Based on the corrected green space boundary, a deep learning target detection network is used to identify non-vegetation covered areas in the multi-temporal image to generate a preliminary green space encroachment location map, and morphological operations and connected domain analysis techniques are used to perform shape regularization and size filtering on the preliminary green space encroachment location map to generate an optimized green space encroachment location map, wherein the non-vegetation covered areas include artificial structures and hardened ground in the green space image;

[0050] Based on the optimized green space encroachment location map, the spatiotemporal data analysis framework is used in combination with environmental parameters to obtain the temporal and spatial distribution characteristics of each factor, and the temporal and spatial distribution characteristics of each factor are combined with the changing trend of the non-vegetation covered area to perform forecast processing to generate trend forecast results, and the trend forecast results are quantitatively evaluated to generate a degradation severity report. The environmental parameters include climate data and the intensity of human activities, and each factor refers to various variables or conditions that may affect green space encroachment and degradation.

[0051] The technical solution of this application has the following beneficial effects:

[0052] This application can accurately capture the dynamic changes of green space over time by periodically collecting multi-temporal images of urban green space and generating difference maps by sorting by timestamp. This method not only improves the temporal resolution of the data, but also provides rich historical information for subsequent analysis. The dynamic information of green space is vectorized using historical geographic information system records and legally required land boundary information, and a high-precision digital model is established as a reference benchmark. By superimposing and comparing this model with the latest remote sensing images, the accuracy and timeliness of the green space boundaries are ensured, human errors are reduced, and monitoring accuracy is improved.

[0053] Based on the corrected green space boundaries, the deep learning target detection network is used to automatically identify non-vegetation covered areas, such as artificial structures and hardened ground, in multi-temporal images to generate preliminary green space encroachment location maps. These location maps are further optimized by combining morphological operations and connected domain analysis techniques to ensure the accuracy of shape regularization and size filtering processing, thereby improving the reliability and practicality of the recognition results. Using the spatiotemporal data analysis framework, combined with environmental parameters such as climate data and human activity intensity, the temporal and spatial distribution characteristics of various factors are analyzed to predict the changing trend of non-vegetation covered areas. This comprehensive analysis method not only considers the influence of natural factors, but also covers the role of human activities, so as to more comprehensively evaluate the severity of green space degradation. The prediction results are quantitatively evaluated to generate detailed degradation severity reports. These reports are presented in the form of intuitive electronic maps and charts, which are easy for decision makers to understand and apply. At the same time, combined with the rule base and case base of the expert system, scientific and reasonable protection suggestions are provided, providing strong support for urban planning and environmental protection.

[0054] Furthermore, by combining the transfer learning algorithm to optimize the model parameters, the pre-trained deep learning target detection network is used to automatically identify the non-vegetation covered area for each green space image in the multi-temporal image dataset. This method not only improves the adaptability and generalization ability of the model in specific urban green space scenarios, but also significantly improves the accuracy of identifying non-vegetation covered areas. The preliminary green space encroachment location map is processed by shape regularization using morphological operation technology combined with the watershed algorithm to ensure that the detected non-vegetation covered area has a more regular shape and clear boundaries. In addition, the connected domain analysis technology and distance transformation algorithm are used to filter the size of the shape regularized area to generate an accurate connected domain that meets the actual green space encroachment standard. This processing method effectively removes small-area false detection areas and retains large-area areas that truly have encroachment characteristics, further improving the reliability of the results. The accurate connected domain information that meets the actual green space encroachment standard is comprehensively processed, and the spatial relationship between each connected domain is evaluated using spatial statistical analysis methods, and finally a comprehensive and detailed green space encroachment location map is output. This not only provides intuitive visualization results, but also provides a scientific basis for subsequent spatial relationship research and decision support, which helps to formulate more reasonable protection measures. The entire process, from image preprocessing, non-vegetation area identification to shape regularization and size filtering processing, and finally to spatial relationship evaluation, is highly automated. It reduces the need for manual intervention, improves work efficiency, and makes large-scale green space monitoring possible.

[0055] Furthermore, a deep learning target detection network suitable for urban green space scenes is used to preprocess each green space image in the multi-phase image dataset, including resizing, color space conversion and normalization. These preprocessing steps ensure the consistency and high quality of the input image, laying a good foundation for subsequent recognition processing. At the same time, a small sample data set collected locally is obtained, and the pre-trained model is fine-tuned using a transfer learning algorithm so that the model can better adapt to the data distribution of the new task, thereby optimizing the model parameters and improving the adaptability of the model in specific scenarios. At the same time, this embodiment applies the target detection model optimized by transfer learning to perform pixel-level classification and bounding box prediction on the prepared green space image, uses non-maximum suppression technology to remove redundant detection frames, and retains the most likely candidate areas. This process not only improves the accuracy of identifying non-vegetation covered areas, but also ensures the precise positioning of the detection frame, reduces the possibility of false detection and missed detection, generates images marked with non-vegetation covered areas, and shows the location and range of suspected encroachment areas.

[0056] Finally, based on the above images marked with non-vegetation covered areas, a preliminary green space encroachment location map is formed. This map not only intuitively shows the situation of green space encroachment, but also provides a reliable basis for subsequent shape regularization and size filtering processing. Analysis of the above technical solutions shows that the entire recognition method can complete the analysis of large-scale green space images in a relatively short time, greatly improving the efficiency of green space monitoring. At the same time, by introducing local small sample data sets for model fine-tuning, and using regularization terms to prevent overfitting, it is ensured that the model has good robustness and stability when facing complex and changeable urban green space scenes. In particular, when processing different batches of data, different weights are given to different batches, which improves the overall performance of the model and ensures the stability and consistency of the recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A flowchart of a method for identifying urban green space encroachment and degradation based on computer vision recognition provided in an embodiment of the present application;

[0058] Figure 2 A schematic diagram of the structure of an urban green space encroachment and degradation identification system based on computer vision recognition provided in an embodiment of the present application;

[0059] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0061] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0062] Based on this, this application provides a method for identifying urban green space encroachment and degradation based on computer vision recognition, such as Figure 1 ,include:

[0063] Step 101: Periodically collect multi-phase images of urban green space, sort the multi-phase images according to timestamps, generate difference maps between the images of each phase, and obtain dynamic information reflecting the change of urban green space over time based on the difference maps between the images of each phase;

[0064] In this step, periodic acquisition refers to repeated observations of a specific area at fixed time intervals, such as monthly, quarterly or annually, to obtain data on changes in the area over time. Multi-temporal images refer to images of the same area obtained at different time points. After these images are sorted by timestamp, a difference map can be generated, which reflects the changes in surface cover by comparing the differences between images of each phase. This dynamic information can reveal the development trend of urban green space over time.

[0065] In the embodiment of the present application, it is assumed that a city's greening bureau wants to monitor changes in green spaces in downtown parks. They can use satellite remote sensing technology, such as the Landsat series of satellites or high-resolution commercial satellites, to collect images of the area every three months. Then, using specialized software tools such as ENVI, ERDAS, etc., these images are preprocessed and analyzed to calculate the change in each pixel value over time, thereby generating a difference map. This will help identify changes in green space area, changes in vegetation health, and so on.

[0066] The embodiment of the present application provides a quantitative way to observe and record the process of green space changes, laying a foundation for subsequent more in-depth analysis. It can help decision makers to understand the status of green space resources in a timely manner and take necessary protective measures.

[0067] Step 102: Based on the historical geographic information system records and the land boundary information stipulated by the law, the dynamic information reflecting the change of the green space over time is vectorized to establish a digital model as a reference for the green space boundary, and the digital model is superimposed and compared with the latest remote sensing image to obtain a corrected green space boundary;

[0068] In this step, the latest remote sensing image refers to the image obtained by modern remote sensing technology at the most recent time point, which is used to provide the latest surface condition information; vectorization processing refers to the process of converting raster map data into vector format to facilitate spatial analysis. The reference benchmark of the green space boundary is a digital model created based on historical Geographic Information System (GIS) records and legal land boundary information. The latest remote sensing image refers to the most recently obtained high-quality ground image, which is used to update the existing green space boundary information.

[0069] In the embodiment of the present application, continuing with the above example, the Greening Bureau can build a vector database containing all known green space boundaries based on the GIS records of the past few years and the official land boundary information provided by the urban planning department. Then, this database is superimposed with the latest high-resolution remote sensing image to check whether there is a deviation between the two. If there is, the boundaries in the vector model are adjusted to match the actual situation.

[0070] The embodiments of the present application can obtain a more accurate definition of green space boundaries, ensuring that the basic data for subsequent analysis is up-to-date and reliable. This is crucial for detecting green space encroachment, as any inaccurate boundaries may lead to misjudgment.

[0071] Step 103: Based on the corrected green space boundary, a deep learning target detection network is used to identify non-vegetation covered areas in the multi-temporal image to generate a preliminary green space encroachment location map, and morphological operations and connected domain analysis techniques are used to perform shape regularization and size filtering on the preliminary green space encroachment location map to generate an optimized green space encroachment location map;

[0072] In this step, non-vegetation covered areas include artificial structures and hardened ground in green space images; the deep learning object detection network is an artificial intelligence algorithm that can automatically find specific objects or patterns from images. Non-vegetation covered areas refer to parts of green space images that do not belong to natural plants, such as buildings, roads or other artificial structures. Morphological operations and connected domain analysis techniques are used to improve the quality of preliminary detection results, eliminate noise points and maintain the main features.

[0073] In an embodiment of the present application, the Greening Bureau can use a trained convolutional neural network (CNN network) to identify non-vegetation covered areas in multi-temporal images. First, a pre-trained model suitable for the urban environment is selected and fine-tuned with a local small sample data set. After that, the optimized model is applied to classify the prepared green space image and mark the locations of all non-vegetation covered areas. Finally, morphological operations are used to remove isolated small patches, and the final green space encroachment location map is determined through connected domain analysis.

[0074] The embodiment of the present application improves the accuracy of identifying non-vegetation covered areas and reduces the false alarm rate, making the green space encroachment location map more reliable. At the same time, the automated process greatly improves work efficiency and reduces labor costs.

[0075] Step 104: Based on the optimized green space encroachment location map, using the spatiotemporal data analysis framework and combining environmental parameters, the temporal and spatial distribution characteristics of each factor are obtained, and the temporal and spatial distribution characteristics of each factor are combined with the change trend of the non-vegetation covered area to predict and process, generate trend prediction results, and quantitatively evaluate the trend prediction results to generate a degradation severity report.

[0076] In this step, environmental parameters include climate data and the intensity of human activities. Factors refer to various variables or conditions that may affect the encroachment and degradation of green spaces. The spatiotemporal data analysis framework combines information from both time and space dimensions to analyze and predict the development trends of geographical phenomena. Environmental parameters include natural factors and socioeconomic factors that affect the health of green spaces. By understanding these parameters and their interactions, we can better predict possible problems in the future and develop corresponding prevention strategies.

[0077] In the embodiment of the present application, the Greening Bureau can integrate data from multiple sources, such as climate data such as temperature and precipitation provided by meteorological stations, as well as data reflecting the intensity of human activities such as traffic flow and the number of building permits issued. Then, the relationship between these factors and green space encroachment is explored using the spatiotemporal data analysis framework. Based on the results obtained, the Greening Bureau can predict future trends in green space encroachment, quantitatively assess the potential impact, and generate a degradation severity report. For example, if it is found that the intensity of human activities in a certain area is increasing year by year, while the area of ​​green space is decreasing during the same period, then it can be warned that the area may face more serious green space degradation problems.

[0078] The embodiments of the present application can obtain scientific predictions about the future development of green areas and provide strong support for policy making. In addition, quantitative assessment can help decision makers intuitively understand the severity of the problem, thereby giving priority to areas that need intervention the most.

[0079] Through the above four steps, the program builds a complete urban green space monitoring and early warning system. It can track green space changes in real time, accurately identify encroachment behaviors, predict future development trends, and provide effective management strategies. This method greatly improves the efficiency and scientific nature of urban green space protection work, and helps maintain the ecological balance of the city and the quality of life of residents.

[0080] Optionally, in step 103, based on the corrected green space boundary, a deep learning target detection network is used to identify non-vegetation covered areas in the multi-temporal image to generate a preliminary green space encroachment location map, and morphological operations and connected domain analysis techniques are used to perform shape regularization and size filtering on the preliminary green space encroachment location map to generate an optimized green space encroachment location map, including: based on the corrected green space boundary, using a pre-trained deep learning target detection network, combined with a transfer learning algorithm to optimize model parameters, automatically identifying non-vegetation covered areas for each green space image in the multi-temporal image dataset The green space encroachment location map is obtained by performing special processing on the detected non-vegetation covered areas according to the preliminary green space encroachment location map; according to the preliminary green space encroachment location map, the morphological operation technology is adopted, and the watershed algorithm is combined to perform shape regularization processing on the detected non-vegetation covered areas to obtain the shape regularized areas; the connected domain analysis technology is applied, and the distance transformation algorithm is combined to perform size filtering processing on the shape regularized areas to generate accurate connected domains that meet the actual green space encroachment standards; the information of the accurate connected domains that meet the actual green space encroachment standards is comprehensively processed, and the spatial statistical analysis method is used to evaluate the spatial relationship between each connected domain in the accurate connected domain, and the final green space encroachment location map is output.

[0081] In this step, the transfer learning algorithm is a machine learning method that accelerates model training and improves performance by applying knowledge gained in one field or task to another related field or task. In this scenario, transfer learning is used to optimize the model parameters of the deep learning object detection network so that it can better adapt to the multi-temporal image dataset of green space in a specific city. Morphological operation technology refers to a series of image processing operations, such as dilation and erosion, which are used to improve the quality of binary images; the watershed algorithm is a segmentation method that can effectively separate adjacent but different areas and help shape regularization. The distance transformation algorithm can calculate the distance from each pixel to the nearest background pixel. Combined with the connected domain analysis technology, it can help screen out accurate connected domains that meet the actual green space encroachment standards, that is, those non-vegetation covered areas that meet the preset conditions in size and shape. Spatial statistical analysis methods are used to evaluate the spatial distribution patterns between different connected domains to ensure that the final output of the green space encroachment location map is both accurate and geographically meaningful.

[0082] In the embodiment of the present application, the optional scheme first uses a deep learning target detection network that has been pre-trained and optimized through transfer learning to automatically identify each green space image within the corrected green space boundary, mark all non-vegetation covered areas, and thus generate a preliminary green space encroachment location map. Then, the shape regularization processing is performed on the preliminary recognition results by combining morphological operations with the watershed algorithm, so that the detected non-vegetation covered areas are more in line with the actual situation. Next, the connected domain analysis technology and the distance transformation algorithm are used to further size filter the shape-regularized areas, and only those connected domains that meet the actual green space encroachment standards are retained. Finally, the information of these accurate connected domains is comprehensively processed, and spatial statistical analysis is used to evaluate the relationship between each connected domain to ensure that the final green space encroachment location map output not only accurately reflects the current situation, but also provides a reliable basis for subsequent spatial analysis.

[0083] In a large-scale urban greening management project, the environmental protection department decided to use the above optimization scheme to monitor the changes in multiple parks and public green spaces in the city. They first collected high-resolution remote sensing images over the past few years and corrected the green space boundaries based on the latest GIS data.

[0084] Next, the team selected a deep learning model that had performed well in similar environments and applied transfer learning technology to fine-tune it to more accurately identify local non-vegetation cover features. Subsequently, they used morphological operations and watershed algorithms to regularize the shapes of the initially generated green space encroachment location map to ensure that each suspected encroachment point has a reasonable geometric shape. In order to eliminate small-area anomalies that are false positives, they also implemented connected domain analysis and distance transformation to retain only those areas that meet the actual green space encroachment standards. Finally, through spatial statistical analysis, the team was able to fully understand the spatial relationship between each connected domain and provide decision makers with a detailed report on the green space encroachment phenomenon, including specific locations, scales, and possible impact ranges. This method significantly improves work efficiency and also provides scientific support for the formulation of targeted protection measures.

[0085] In the task of identifying urban green space encroachment, a deep learning target detection network is used in combination with a transfer learning algorithm to optimize model parameters, and each green space image in the multi-temporal image dataset is automatically identified as a non-vegetation covered area. This process requires a mathematical formula to describe how to generate a preliminary green space encroachment location map based on the corrected green space boundary through a pre-trained and optimized target detection model. The design of the formula needs to consider the update of model parameters, the calculation of loss functions, the consistency of feature distribution, and the final classification and prediction.

[0086] Optionally, based on the corrected green space boundary, a pre-trained deep learning target detection network is used, combined with a transfer learning algorithm to optimize model parameters, and each green space image in the multi-temporal image data set is automatically identified for non-vegetation covered areas to obtain a preliminary green space encroachment location map, including: using a deep learning target detection network suitable for urban green space scenes to pre-process each green space image in the multi-temporal image data set to generate prepared image data, wherein the pre-processing includes size adjustment, color space conversion and normalization; obtaining a small sample data set collected locally, using a transfer learning algorithm to fine-tune the pre-trained model, optimizing the last layer or layers of the pre-trained model, adapting the pre-trained model to the data distribution of the new task, optimizing the parameters of the pre-trained model, and obtaining a target detection model optimized by transfer learning;

[0087] The target detection model optimized by transfer learning is applied to perform pixel-level classification and bounding box prediction on the prepared green space image, and non-maximum suppression technology is used to remove redundant detection boxes, retain the most likely candidate areas, and generate an image marked with non-vegetation covered areas; based on the image marked with non-vegetation covered areas, a preliminary green space encroachment location map is formed, and the image marked with non-vegetation covered areas shows the location and scope of the suspected encroachment area.

[0088] More specifically, the embodiment of the present application further provides a formula for model parameters after transfer learning optimization, which is used to calculate the model parameters after transfer learning optimization. The specific calculation formula is as follows:

[0089]

[0090] Among them, P base Represents the basic parameters of the pre-trained model; P opt represents the model parameters after transfer learning optimization; η represents the learning rate, which controls the speed of parameter update; N represents the number of different batches in the small sample data set S collected locally; α i represents the importance weight of each batch; L(S,P) represents the loss function calculated on the small sample data set S; represents the gradient of the loss function L(S,P) with respect to the parameter P; γ represents the update factor of the weight matrix W, which controls the adjustment speed of the feature map weight; S i represents the small sample data of the i-th batch; W represents the feature map weight matrix, which indicates the degree of attention paid by the model to different features; W(S i ) represents the small sample data S in the i-th batch i The feature map weight matrix generated above; W base Represents the weight matrix of the basic feature map of the pre-trained model; D KL(W(S i )||W base ) is the Kullback-Leibler divergence, which is used to measure the difference between the feature map weight distribution in the new task and the feature map weight distribution of the pre-trained model; β represents the regularization term coefficient, which is used to prevent overfitting; R(P base ) represents the regularization term to ensure the generalization ability of the model.

[0091] The overall formula aims to improve the adaptability and accuracy of the pre-trained model in specific urban green space scenarios by introducing transfer learning and optimization techniques. The formula not only covers the update rules of the model parameters, but also takes into account the importance weights of different batches of data, learning rate decay, regularization terms, and the adjustment of the feature map weight matrix to ensure that the model can efficiently learn from local small sample data sets and maintain good generalization capabilities. In addition, by introducing the Kullback-Leibler divergence to measure the consistency of the feature distribution of new and old tasks, it helps the model to smoothly transition to new tasks.

[0092] The following is a brief introduction to the design reasons of each sub-item of the formula:

[0093]

[0094] After the model parameters P are optimized by transfer learning opt In the formula, P opt The optimized model parameters are designed to enable the model to show higher accuracy and adaptability in specific urban green space scenarios through transfer learning and local data fine-tuning.

[0095] P base It refers to introducing the basic parameters of the pre-trained model and using its sufficient training results on large-scale general datasets as the starting point for new tasks, reducing the time and resources required for training from scratch.

[0096] η refers to the introduction of a dynamically decaying learning rate to accelerate the convergence speed in the early training stage, and gradually reduce the learning rate in the later stage to prevent overfitting, ensuring that the model can fine-tune parameters and find the optimal solution.

[0097] N refers to the number of different batches, taking into account that the small sample data set collected locally may contain multiple batches of data, so as to better manage the impact of each batch on the model update.

[0098] S i Refers to the small sample data S for each batch i , calculate the gradients respectively, ensure that the model can make full use of limited local data for effective learning and improve the generalization ability of the model.

[0099] α iRefers to the introduction of importance weight α i , weighted according to the quality score of each batch of data, ensuring that high-quality data has a greater impact on the model, thereby improving the robustness and accuracy of the model.

[0100] It means calculating the loss function with respect to the parameter P base The gradient guides the update direction of the model parameters and ensures that each iteration moves in the direction of reducing the loss.

[0101] γ refers to the introduction of the feature map weight update factor γ, which controls the update speed of the feature map weight matrix, helps the model smoothly transition to new tasks, maintains the consistency of feature distribution, and avoids the model from deviating too much from the original knowledge.

[0102] D KL (W(S i )||W base ) refers to using the Kullback-Leibler divergence to measure the difference between the weight distribution of the feature map in the new task and the weight distribution of the feature map of the pre-trained model, ensuring that the model can flexibly adapt to the needs of new tasks while retaining useful features.

[0103] β refers to the introduction of the regularization term coefficient β, which is used to prevent the model from overfitting and ensure that the model has good generalization ability and can perform well on unseen data.

[0104] R(P base ) refers to adding regularization terms to constrain the complexity of model parameters, avoid overfitting due to excessive model complexity, and improve the stability and reliability of the model.

[0105] The following is a brief introduction to how to obtain the parameters of the formula:

[0106] P base : Download model weights that have been trained on large-scale datasets from public pre-trained model libraries. For example, use the official pre-trained models provided by PyTorch or TensorFlow.

[0107] η (learning rate): The initial learning rate can be set to 0.001 based on experience and calculated using an exponential reduction formula based on the number of iterations. The learning rate can also be adjusted dynamically through a learning rate scheduler. For a specific task, the optimal learning rate range can be found through grid search or random search.

[0108] N (number of batches): Determined by the size of the small sample dataset collected locally. Usually, small sample datasets can be processed in batches according to timestamps, geographic locations, or other features, with each batch containing a certain number of images. Automatic batching can be achieved through dataset partitioning tools.

[0109] S i (Data of batch i): Regularly obtain high-resolution images of urban green spaces through field collection equipment, or download multi-temporal remote sensing images of the required area from a public geospatial data platform. Each batch of data should contain annotation information for supervised learning.

[0110] α i (Importance weight): Based on the image quality score q i , calculated by indicators such as image clarity and annotation accuracy. Image quality can be evaluated using an image quality assessment algorithm and combined with the accuracy of manual annotation to give a comprehensive score. These scores are then normalized into importance weights using the softmax function.

[0111] γ (feature map weight update factor): Initially set to a small value (such as 0.01) and based on the Kullback-Leibler divergence D KL (W(S i )||W base ) changes dynamically. Appropriate initial values ​​and adjustment strategies can be determined through experiments to ensure that the model can smoothly transition to new tasks.

[0112] D KL (W(S i )||W base )(Kullback-Leibler divergence): obtained by calculating the difference between the feature map weight distribution in the new task and the feature map weight distribution of the pre-trained model. It can be directly calculated using built-in functions in the deep learning framework.

[0113] β (regularization coefficient): The initial value is set to a small constant (such as 0.0005) and dynamically adjusted through methods such as cross-validation or Bayesian optimization. You can monitor changes in the mean and standard deviation of the training loss and fine-tune the regularization coefficient based on the performance of the model to prevent overfitting.

[0114] R(P hase )(Regularization term): Choose an appropriate regularization method, such as L2 regularization or Dropout, based on the model structure. You can refer to the literature or best practice guides to choose a regularization technique that is suitable for the task at hand. For deep learning models, you can also use ensemble regularization methods, such as weight decay and batch normalization.

[0115] In order to more accurately reflect the impact of different batches of data on model training, the system needs to quantify the quality of each batch of data. i , which ensures that high-quality data is given greater importance during training, thereby improving the robustness and accuracy of the model. Specifically, the batch importance weight αi The calculation formula is as follows:

[0116]

[0117] Among them, q i is the quality score of the small sample data of the i-th batch, which can be calculated based on indicators such as image clarity and annotation accuracy; λ is an adjustment parameter that controls the degree of weight difference between different batches; N represents the number of different batches in the locally collected small sample data set S;

[0118] In the batch importance weight α i The purpose of this formula is to introduce the softmax function to score the quality of each batch of data. i Normalization is performed so that high-quality data has a higher weight, while the impact of low-quality data is relatively reduced. At the same time, the parameter λ is used to adjust the sensitivity of weight calculation to meet the needs of different tasks. This method not only ensures that the model can learn more general and useful features from high-quality data, but also effectively prevents overfitting and improves the performance of the model in specific application scenarios.

[0119] In order to ensure that the model can converge quickly in the early stage of training and avoid oscillation and overfitting when approaching the optimal solution, the system needs to dynamically adjust the learning rate. By introducing the attenuation mechanism, it can ensure that the learning rate is large in the early stage of training to accelerate the convergence speed, and gradually decreases in the later stage of training to fine-tune the parameters. Specifically, the calculation formula of the learning rate η is as follows:

[0120]

[0121] Among them, η0 is the initial learning rate; t is the current number of iterations; T is the total number of iterations; γ is the decay factor, which controls the speed at which the learning rate changes over time;

[0122] In the calculation formula of the learning rate η, the purpose of this formula is to provide a larger value in the early stage of training to accelerate the convergence speed, and gradually reduce the learning rate as the number of iterations increases, so as to ensure that the model can adjust the parameters more finely in the later stage and converge to the optimal solution stably. The attenuation factor controls the speed at which the learning rate decreases, so that the model can avoid instability or overfitting problems caused by excessive learning rate while maintaining efficient training, thereby improving the generalization ability and final performance of the model.

[0123] In order to ensure that the model can effectively prevent overfitting during training and dynamically adjust the strength of regularization according to data distribution, the system needs to adaptively adjust the regularization term coefficient. Specifically, the calculation formula of the regularization term coefficient β is as follows:

[0124] β=β0·(1+exp(-κ·(μ-σ))) -1

[0125] Among them, β0 is the initial regularization coefficient; μ is the mean of the training loss; σ is the standard deviation of the training loss; κ is the sensitivity parameter, which controls the speed at which the regularization coefficient changes with the loss distribution;

[0126] In the calculation formula of the regularization term coefficient β, the purpose of this formula is to dynamically adjust the strength of regularization according to the statistical characteristics of the training loss (such as the mean and standard deviation), so that the model has strong regularization in the early stage of training to suppress overfitting. As the training progresses and the model performance improves, the regularization coefficient is gradually reduced, allowing the model to better adapt to the complex patterns in the data. This approach ensures that the model maintains good generalization capabilities throughout the training period and can flexibly respond to the training needs of different stages, thereby optimizing the final performance and stability.

[0127] In order to ensure that the model can smoothly transition to new tasks during transfer learning and retain useful pre-trained features, the system needs to dynamically adjust the feature map weights based on the difference in feature distribution between the new and old tasks. Specifically, the calculation formula for the feature map weight update factor γ is as follows:

[0128]

[0129] Among them, γ0 is the initial feature map weight update factor; D KL (W(S i )||W base ) is the Kullback-Leibler divergence, which measures the difference between the weight distribution of the feature map in the new task and the weight distribution of the feature map of the pre-trained model; is the maximum possible Kullback-Leibler divergence value, used for normalization.

[0130] In the calculation formula of the feature map weight update factor γ, the purpose of this formula is to optimize the performance of the model in transfer learning and ensure that it can flexibly adapt to the needs of new tasks while retaining useful pre-trained features. The system introduces the feature map weight update factor γ, which dynamically adjusts the update speed of the feature map weight by measuring the Kullback-Leibler divergence difference between the feature distribution of the new task and the feature distribution of the pre-trained model, ensuring that the model can not only maintain the advantages of the original features, but also flexibly respond to the specific needs of new tasks, thereby improving overall performance and stability.

[0131] Assuming the current number of iterations t = 5000, calculate the learning rate η:

[0132]

[0133] Assume that the quality score q of a batch of data is i =0.8, calculate the importance weight α i :

[0134]

[0135] Here we assume that all batches have the same data quality score.

[0136] Assume that the Kullback-Leibler divergence D of a batch KL (W(S i )||W base )=0.3, calculate the feature map weight update factor γ:

[0137]

[0138] Through the above formula and parameter settings, we have successfully optimized the YOLOv5 model to make it more suitable for urban green space scenarios. As the number of iterations increases, the learning rate gradually decreases, which helps the model to adjust parameters more finely and avoid oscillation near the optimal solution. Importance weight α i Ensuring high-quality data has a greater impact on model training and improves the robustness and stability of the model. The feature map weight update factor γ controls the update speed of the feature map weight matrix, allowing the model to flexibly adapt to the needs of new tasks while retaining useful features. Ultimately, the optimized model can accurately identify non-vegetated areas in green spaces and generate a preliminary green space encroachment location map, providing a scientific basis for subsequent spatial analysis and decision support.

[0139] Optionally, applying the object detection model optimized by transfer learning to perform pixel-level classification and bounding box prediction on the prepared green space image, using non-maximum suppression technology to remove redundant detection boxes, retaining the most likely candidate areas, and generating an image marked with non-vegetation covered areas, includes:

[0140] Using the target detection model optimized by transfer learning, the prepared green space image is classified at the pixel level to obtain a probability map of each pixel belonging to the background or the non-vegetation covered area; based on the probability map of each pixel belonging to the background or the non-vegetation covered area, the regression information output by the target detection model is used to obtain the detected non-vegetation covered area, and the bounding box of the detected non-vegetation covered area is predicted to generate a series of candidate area bounding boxes containing position information and confidence scores, wherein the position information includes center coordinates, width and height;

[0141] Based on the series of candidate region bounding boxes containing location information and confidence scores, a non-maximum suppression technique is applied, the candidate region bounding boxes are sorted according to the confidence scores, and the intersection-and-union ratios between adjacent bounding boxes of the candidate region bounding boxes are compared in sequence. For the bounding boxes whose intersection-and-union ratios are higher than a preset threshold, the one with the highest confidence is retained, and other redundant detection boxes are removed to obtain the retained bounding box; the retained bounding box is drawn onto the periodic multi-temporal image dataset to form an image marked with a non-vegetation covered area.

[0142] In this step, the concepts involved in this solution include deep learning target detection network, transfer learning algorithm, preprocessing, non-maximum suppression technology, pixel-level classification and bounding box prediction. The deep learning target detection network is a neural network that can automatically identify specific objects in an image. It is used for the identification of non-vegetation covered areas of the multi-temporal green space image dataset in this case. The transfer learning algorithm allows the use of a model trained on a field or task and adapts it to a new but related field or task through fine-tuning. Here, it means using a pre-trained model and optimizing it for a small sample dataset collected locally to better adapt to urban green space scenarios.

[0143] Preprocessing refers to a series of transformation operations on the original image data, such as resizing, color space conversion, and normalization, to ensure that the data input to the model meets the requirements and improve the performance of the model. Non-maximum suppression technology is used to select the most likely box from a series of candidate detection boxes and remove redundant overlapping boxes by comparing the intersection-over-union ratio. Pixel-level classification is the process of classifying each pixel in the image, while bounding box prediction generates location information and confidence scores for detected objects.

[0144] In an embodiment of the present application, the scheme first pre-processes each green space image in the multi-phase image data set, including resizing, color space conversion and normalization, to generate prepared image data suitable for model input. Then, based on a small sample data set collected locally, a transfer learning algorithm is used to fine-tune the pre-trained target detection model, especially optimizing the last layer or several layers to make the model parameters more in line with the data distribution of the new task. Next, the optimized target detection model is applied to perform pixel-level classification and bounding box prediction on the prepared image to obtain a probability map of each pixel belonging to the background or non-vegetation covered area, and a candidate area bounding box containing location information and confidence scores is generated accordingly. Finally, non-maximum suppression technology is used to remove redundant detection frames, retain the most likely candidate areas, and form images marked with non-vegetation covered areas. These images show the location and range of suspected encroachment areas, thereby constructing a preliminary green space encroachment location map.

[0145] In a smart city project, in order to monitor the changes in urban park green spaces, the relevant departments decided to use the above solution to analyze the phenomenon of green space encroachment. The team first collected high-resolution satellite images taken in different seasons over the past few years as a multi-temporal image dataset and performed necessary preprocessing. They selected a deep learning target detection model that performed well in similar environments and optimized it with transfer learning using a local small sample dataset, especially adjusting the last layer of the model to adapt to local features. The optimized model was applied to the processed images, performing pixel-level classification and bounding box prediction, and generating a series of possible candidate boxes for non-vegetation covered areas.

[0146] Then, the optimal bounding box was selected through non-maximum suppression technology, forming an image with clear markings of non-vegetation covered areas. These images not only show the specific location and scope of suspected green space encroachment, but also help the team accurately assess the trend of green space changes, providing strong support for the subsequent formulation of protection measures. Ultimately, this method improves the efficiency and accuracy of green space management and helps maintain the green ecological balance of the city.

[0147] Optionally, in step 104, based on the optimized green space encroachment location map, the spatiotemporal data analysis framework is used in combination with environmental parameters to obtain the temporal and spatial distribution characteristics of each factor, the temporal and spatial distribution characteristics of each factor are combined with the change trend of the non-vegetation covered area to predict and process, generate trend prediction results, perform quantitative evaluation processing on the trend prediction results, and generate a degradation severity report, including: using the optimized green space encroachment location map data, integrating the information of the non-vegetation covered area to obtain a data set suitable for the spatiotemporal data analysis framework; collecting and preprocessing the environmental parameters according to the data set suitable for the spatiotemporal data analysis framework to generate a degradation severity report related to the non-vegetation covered area; Multidimensional spatiotemporal data related to changes in covered areas; based on the multidimensional spatiotemporal data related to changes in non-vegetation covered areas, the temporal and spatial distribution characteristics of each factor in the environmental parameters are analyzed and modeled to obtain a relationship model between different environmental parameters and changes in non-vegetation covered areas; using the relationship model between different environmental parameters and changes in non-vegetation covered areas, the change trend of the non-vegetation covered areas is predicted and processed to generate trend prediction results for a period of time in the future; based on the trend prediction results for a period of time in the future, the expansion speed, impact range and ecological impact of the non-vegetation covered areas are quantitatively evaluated and processed to generate a degradation severity report. Optionally, using the relationship model between different environmental parameters and changes in non-vegetation covered areas to predict the change trend of the non-vegetation covered areas and generate trend prediction results for a period of time in the future includes:

[0148] Using the relationship model between the different environmental parameters and the changes in non-vegetation covered areas, a suitable machine learning or statistical model is selected to train the change data of the non-vegetation covered areas, and to construct a prediction model that captures the temporal and spatial change laws; according to the needs of the prediction model that captures the temporal and spatial change laws, the pre-processed multidimensional spatiotemporal data is integrated, and the non-vegetation covered area change data including historical periods and the latest observations are used as input to generate an input data set suitable for prediction;

[0149] Based on the input data set suitable for prediction, different future scenario assumptions are set to simulate the changing trends of different non-vegetation covered areas that may occur in the future, and generate multiple prediction scenarios covering the best case, the most likely case and the worst case. The different future scenario assumptions include different paths of climate change and changes in the intensity of human activities;

[0150] By using the multiple prediction scenarios covering the best case, the most likely case and the worst case, the short-term and long-term change trends of the non-vegetation covered area are predicted and processed to generate a series of trend prediction results at future time points.

[0151] In this step, the concepts involved in this scheme include spatiotemporal data analysis framework, environmental parameters, multidimensional spatiotemporal data, relational models, prediction models and scenario assumptions. The spatiotemporal data analysis framework is an analysis method that integrates time series and spatial distribution information to study the changing patterns of geographical phenomena over time and place. Environmental parameters refer to various external factors that affect the changes in green space, such as climate conditions and the intensity of human activities. Multidimensional spatiotemporal data combines the changes in non-vegetated areas with environmental parameters to form a data set containing multiple dimensions to fully reflect the change process. The relational model refers to the correlation model between different environmental parameters and the changes in non-vegetated areas established by statistical or machine learning methods to reveal the causal relationship between the two. The prediction model is a mathematical model trained with historical data that can capture the laws of temporal and spatial changes and is used to predict future trends. Scenario assumptions are to set different possible scenarios for future development, such as different paths of climate change and changes in the intensity of human activities, to simulate multiple possible results and provide a basis for decision-making.

[0152] In the embodiment of the present application, the scheme first integrates and processes the non-vegetation covered area information in the optimized green space encroachment location map to generate a data set suitable for the spatiotemporal data analysis framework. Then, relevant environmental parameters are collected and preprocessed to construct multidimensional spatiotemporal data related to the changes in non-vegetation covered areas. Next, based on these multidimensional spatiotemporal data, the temporal and spatial distribution characteristics of each environmental parameter are analyzed and modeled to obtain a relationship model between different environmental parameters and the changes in non-vegetation covered areas. Use these relationship models to select appropriate machine learning or statistical models, train the change data of non-vegetation covered areas, and construct a prediction model that captures the laws of temporal and spatial changes. According to the needs of the prediction model, the preprocessed multidimensional spatiotemporal data is integrated to generate an input data set suitable for prediction, and different future scenario assumptions are set to simulate different possible change trends. Finally, by predicting the short-term and long-term change trends of multiple prediction scenarios covering the best case, the most likely case and the worst case, a series of trend prediction results at future time points are generated, and the speed of expansion of non-vegetation covered areas, the scope of influence and the degree of ecological impact are quantitatively evaluated, and finally a degradation severity report is generated.

[0153] In an urban planning project, the environmental protection department decided to use the above scheme to assess the future trend of urban green space changes and their potential impacts. They first integrated the optimized green space encroachment location map in the past few years, extracted the information of non-vegetation covered areas, and formed a complete spatiotemporal data set. Then, the team collected local climate data and human activity data, and pre-processed these data to generate multi-dimensional spatiotemporal data related to the changes in non-vegetation covered areas.

[0154] Then, based on these data, the team established a relationship model between different environmental parameters and the changes in non-vegetation areas, revealing the impact of climate change and human activities on green space changes. Using these relationship models, the team selected random forest regression as a prediction model and trained the change data of non-vegetation areas.

[0155] In order to generate accurate prediction results, the team integrated the pre-processed historical period and the latest observed data on changes in non-vegetated areas, and set three future scenario assumptions: best case, most likely case, and worst case. Based on these scenarios, the team simulated different possible trends in the next 10 years and generated a series of detailed prediction results. Finally, the team quantitatively evaluated the expansion rate, impact range, and ecological impact of non-vegetated areas, compiled a detailed degradation severity report, and provided scientific decision-making support for the municipal government. This method not only improves the accuracy of predictions, but also provides a strong basis for formulating reasonable urban planning and environmental protection strategies.

[0156] Optionally, in step 102, according to the land boundary information recorded in the historical geographic information system and stipulated by law, the dynamic information reflecting the change of the green space over time is vectorized to establish a digital model as a reference for the green space boundary, and the digital model is superimposed and compared with the latest remote sensing image to obtain a corrected green space boundary, including:

[0157] Using the land boundary information recorded in the historical geographic information system and the legally prescribed land boundary information, the dynamic information reflecting the change of the green space over time is vectorized to obtain vector data that accurately represents the change of the green space and the legal boundary of the green space change; based on the vector data that accurately represents the change of the green space and the legal boundary of the green space change, a digital model is established and used as a reference for the green space boundary, converted into a digital format, and a high-precision digital model of the green space change and boundary position is generated;

[0158] Modern remote sensing technology is used to obtain the latest remote sensing images with high resolution, high precision and high timeliness. The coordinate system of the digital model is made consistent with that of the latest remote sensing image through image registration technology, and superposition and comparison processing is performed to obtain the detection result of the green space boundary change. Based on the detection result of the green space boundary change, combined with field surveys or other auxiliary data sources, the dynamic information reflecting the change of the green space over time is adjusted and corrected to generate a corrected green space boundary.

[0159] In this step, the concepts involved in this scheme include vector processing, digital models, legal boundaries of green space changes, remote sensing images and image registration technology. Vector processing refers to converting dynamic information reflecting the changes of green space over time, such as historical geographic information system records and legal land boundary information, into a vector data format composed of geometric elements such as points, lines, and surfaces, so as to facilitate accurate analysis and management by computers. The digital model is a reference benchmark for green space changes and boundaries established by digital means. It can accurately represent the location, shape and size of green space and can be integrated and analyzed with other geographic information data. The legal boundaries of green space changes refer to the green space use boundaries determined according to laws and regulations to ensure the protection status of green space in urban planning. Remote sensing images refer to surface information obtained by sensors on platforms such as satellites or aircraft. They have the characteristics of high resolution, high precision and high timeliness, and can provide the latest surface coverage for green space monitoring. Image registration technology is used to adjust the coordinate systems between images from different sources so that they are aligned with each other for direct comparison and analysis.

[0160] In the embodiment of the present application, the scheme first uses the land boundary information recorded in the historical geographic information system and stipulated by law to vectorize the dynamic information reflecting the changes of green space over time, and generate vector data that accurately represents the changes of green space and its legal boundaries. Then, based on these vector data, a digital model is established as a reference for the green space boundary, which is converted into a digital format to generate a high-precision digital model of the green space change and boundary position. Next, the latest remote sensing images are obtained through modern remote sensing technology, and the image registration technology is used to keep the coordinate system of the digital model consistent with that of the remote sensing image, and a superposition comparison is performed to detect the changes in the green space boundary. Finally, according to the detection results of the green space boundary changes, combined with field surveys or other auxiliary data sources, such as on-site photos or local regulations updates, the dynamic information reflecting the changes of green space over time is adjusted and corrected, thereby generating a more accurate corrected green space boundary, providing a scientific basis for green space management and protection.

[0161] In a green space protection project in a city, the relevant departments adopted the above scheme in order to more accurately understand the changes of green spaces and formulate effective protection strategies. They first collected historical geographic information system records and legal land boundary information from the past few decades, including the original distribution of green spaces, changes over the years, and related legal documents, for vector processing, forming a set of vector data sets that accurately represent the changes of green spaces and their legal boundaries. Then, the team established a digital model to convert the information of green space changes into digital format and constructed a high-precision digital model of green space changes and boundary locations. Then, the team obtained the latest high-resolution remote sensing images through modern remote sensing technology, and used image registration technology to ensure that the coordinate system of the digital model and the remote sensing image matched, and performed overlay comparison to identify the changes in the green space boundaries. Based on the detection results of these changes, the team members conducted field visits, verified the results of remote sensing image analysis, and combined with other auxiliary data sources, such as feedback provided by local residents and recent reports on urban construction activities, to make necessary adjustments and corrections to the information on green space changes. Finally, the team generated a corrected green space boundary report, which not only improved the accuracy of green space boundaries, but also provided an important reference for future green space protection work, contributing to the effective management and sustainable development of urban green spaces.

[0162] Embodiment 2

[0163] Figure 2 The present application provides a schematic diagram of a system for identifying urban green space encroachment and degradation based on computer vision recognition. Figure 2 As shown, the system includes:

[0164] The acquisition module 21 (i.e., the drone imaging detection platform) is used to periodically acquire multi-phase images of urban green spaces, sort the multi-phase images according to timestamps, generate difference maps between the images of each phase, and obtain dynamic information reflecting the changes of urban green spaces over time based on the difference maps between the images of each phase;

[0165] The processing module 22 (i.e., a high-performance server) is used to vectorize the dynamic information reflecting the changes of the green space over time based on the historical geographic information system records and the land boundary information stipulated by the law, so as to establish a digital model as a reference for the green space boundary, and to superimpose and compare the digital model with the latest remote sensing image obtained to obtain a corrected green space boundary. The latest remote sensing image refers to an image obtained by modern remote sensing technology at the most recent time point, which is used to provide the latest surface condition information;

[0166] The recognition module 23 is used to identify the non-vegetation covered area in the multi-temporal image based on the corrected green space boundary by using a deep learning target detection network to generate a preliminary green space encroachment location map, and to perform shape regularization and size filtering processing on the preliminary green space encroachment location map by using morphological operations and connected domain analysis technology to generate an optimized green space encroachment location map, wherein the non-vegetation covered area includes artificial structures and hardened ground in the green space image;

[0167] A generation module 24 is used to obtain the temporal and spatial distribution characteristics of each factor based on the optimized green space encroachment location map, using a spatiotemporal data analysis framework and combining environmental parameters, and to predict the temporal and spatial distribution characteristics of each factor in combination with the changing trend of the non-vegetation covered area to generate a trend prediction result, to quantitatively evaluate the trend prediction result, and to generate a degradation severity report, wherein the environmental parameters include climate data and the intensity of human activities, and each factor refers to various variables or conditions that may affect green space encroachment and degradation.

[0168] The acquisition module is a drone imaging detection platform: the drone imaging detection platform is used to periodically acquire multi-temporal images of urban green spaces. These platforms can be equipped with high-resolution cameras, multispectral or thermal imaging sensors and other equipment. The processing module preferably uses a high-performance server; it is used to store large amounts of historical geographic information system (GIS) data and land boundary information, as well as to perform complex vectorization and model building tasks. The server needs to have sufficient CPU / GPU computing power to support large-scale data processing and machine learning algorithm training. The above-mentioned high-performance server will also implement the functions of the above-mentioned recognition module 23 and generation module 24;

[0169] Figure 2 The urban green space encroachment and degradation identification system based on computer vision recognition can be implemented Figure 1The implementation principle and technical effect of the urban green space encroachment and degradation identification method based on computer vision recognition described in the illustrated embodiment will not be described in detail. The specific manner in which each module and unit performs operations in the urban green space encroachment and degradation identification system based on computer vision recognition in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0170] In one possible design, Figure 2 The urban green space encroachment and degradation identification system based on computer vision recognition in the illustrated embodiment can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0171] In a possible design, the present invention provides a computer storage medium storing a computer program, which, when executed by a computer, implements a method for identifying encroachment and degradation of urban green space based on computer vision recognition.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An urban green space encroachment and degradation identification system based on computer vision recognition, characterized in that: include: An acquisition module is used to periodically acquire multi-phase images of urban green space, sort the multi-phase images according to timestamps, generate a difference map between the images of each phase, and obtain dynamic information reflecting the change of urban green space over time based on the difference map between the images of each phase; A processing module is used to vectorize the dynamic information reflecting the change of green space over time according to the land boundary information recorded in the historical geographic information system and the legal provisions, so as to establish a digital model as a reference for the green space boundary, and to superimpose and compare the digital model with the latest remote sensing image to obtain a corrected green space boundary. The latest remote sensing image refers to an image obtained by modern remote sensing technology at the most recent time point, which is used to provide the latest surface condition information; an identification module, for identifying non-vegetation covered areas in the multi-temporal image based on the corrected green space boundary by using a deep learning target detection network to generate a preliminary green space encroachment location map, and performing shape regularization and size filtering processing on the preliminary green space encroachment location map by using morphological operations and connected domain analysis technology to generate an optimized green space encroachment location map, wherein the non-vegetation covered areas include artificial structures and hardened ground in the green space image; A generation module is used to obtain the time and space distribution characteristics of each factor based on the optimized green space encroachment location map, using a spatiotemporal data analysis framework and combining environmental parameters, predict the time and space distribution characteristics of each factor in combination with the change trend of the non-vegetation covered area, generate trend prediction results, quantitatively evaluate the trend prediction results, and generate a degradation severity report. The environmental parameters include climate data and human activity intensity, and each factor refers to various variables or conditions that may affect green space encroachment and degradation.

2. A method for identifying urban green space encroachment and degradation based on computer vision recognition, characterized in that: include: Periodically collect multi-phase images of urban green space, sort the multi-phase images according to timestamps, generate difference maps between the images of each phase, and obtain dynamic information reflecting the changes of urban green space over time based on the difference maps between the images of each phase; Based on the historical geographic information system records and the land boundary information stipulated by the law, the dynamic information reflecting the changes of the green space over time is vectorized to establish a digital model as a reference for the green space boundary, and the digital model is superimposed and compared with the latest remote sensing image obtained to obtain a corrected green space boundary. The latest remote sensing image refers to an image obtained by modern remote sensing technology at the most recent time point to provide the latest surface condition information; Based on the corrected green space boundary, a deep learning target detection network is used to identify non-vegetation covered areas in the multi-temporal image to generate a preliminary green space encroachment location map, and morphological operations and connected domain analysis techniques are used to perform shape regularization and size filtering on the preliminary green space encroachment location map to generate an optimized green space encroachment location map, wherein the non-vegetation covered areas include artificial structures and hardened ground in the green space image; Based on the optimized green space encroachment location map, the spatiotemporal data analysis framework is used in combination with environmental parameters to obtain the temporal and spatial distribution characteristics of each factor, and the temporal and spatial distribution characteristics of each factor are combined with the changing trend of the non-vegetation covered area to perform forecast processing to generate trend forecast results, and the trend forecast results are quantitatively evaluated to generate a degradation severity report, wherein the environmental parameters include climate data and the intensity of human activities, and each factor refers to various variables or conditions that may affect green space encroachment and degradation.

3. The method according to claim 2, characterized in that Based on the corrected green space boundary, the deep learning target detection network is used to identify the non-vegetation covered area in the multi-temporal image to generate a preliminary green space encroachment location map, and the morphological operation and connected domain analysis technology are used to perform shape regularization and size filtering on the preliminary green space encroachment location map to generate an optimized green space encroachment location map, including: Based on the corrected green space boundary, a pre-trained deep learning target detection network is used in combination with a transfer learning algorithm to optimize model parameters, and each green space image in the multi-temporal image dataset is automatically identified for non-vegetation covered areas to obtain a preliminary green space encroachment location map; According to the preliminary green space encroachment location map, the detected non-vegetation covered area is regularized by using morphological operation technology combined with a watershed algorithm to obtain a shape-regularized area, and the shape-regularized area is filtered by using a connected domain analysis technology combined with a distance transformation algorithm to generate an accurate connected domain that meets the actual green space encroachment standard; The information of the accurate connected domain that meets the actual green space encroachment standard is comprehensively processed, and the spatial relationship between each connected domain in the accurate connected domain is evaluated using a spatial statistical analysis method to output a final green space encroachment location map.

4. The method according to claim 3, characterized in that Based on the corrected green space boundary, a pre-trained deep learning target detection network is used, and the model parameters are optimized in combination with a transfer learning algorithm, and each green space image in the multi-temporal image data set is automatically identified for non-vegetation covered areas to obtain a preliminary green space encroachment location map, including: Using a deep learning target detection network suitable for urban green space scenes, preprocessing each green space image in the multi-temporal image dataset to generate prepared image data, wherein the preprocessing includes resizing, color space conversion, and normalization processing; Obtain a small sample data set collected locally, use a transfer learning algorithm to fine-tune the pre-trained model, optimize the last layer or layers of the pre-trained model, adapt the pre-trained model to the data distribution of the new task, optimize the parameters of the pre-trained model, and obtain a target detection model optimized by transfer learning; Applying the object detection model optimized by transfer learning, performing pixel-level classification and bounding box prediction on the prepared green space image, using non-maximum suppression technology to remove redundant detection boxes, retaining the most likely candidate areas, and generating an image marked with non-vegetation covered areas; Based on the image marked with non-vegetation covered areas, a preliminary green space encroachment location map is formed, and the image marked with non-vegetation covered areas shows the location and scope of the suspected encroachment areas.

5. The method according to claim 4, characterized in that The object detection model optimized by applying the transfer learning is used to perform pixel-level classification and bounding box prediction on the prepared green space image, and a non-maximum suppression technique is used to remove redundant detection boxes, retain the most likely candidate areas, and generate an image marked with non-vegetation covered areas, including: Using the object detection model optimized by transfer learning, the prepared green space image is classified at the pixel level to obtain a probability map of each pixel belonging to the background or non-vegetation covered area; Based on the probability map of each pixel belonging to the background or non-vegetation covered area, the detected non-vegetation covered area is obtained by using the regression information output by the target detection model, and a bounding box prediction is performed on the detected non-vegetation covered area to generate a series of candidate area bounding boxes including position information and confidence scores, wherein the position information includes center coordinates, width and height; Based on the series of candidate region bounding boxes containing position information and confidence scores, applying non-maximum suppression technology, sorting according to the confidence scores, sequentially comparing the intersection-and-union ratios between adjacent bounding boxes of the candidate region bounding boxes, and retaining the one with the highest confidence for the bounding boxes whose intersection-and-union ratios are higher than a preset threshold, and removing other redundant detection boxes to obtain the retained bounding boxes; The retained bounding box is drawn onto the periodic multi-temporal image dataset to form an image with non-vegetation covered areas marked.

6. The method according to claim 5, characterized in that Based on the optimized green space encroachment location map, the spatiotemporal data analysis framework is used in combination with environmental parameters to obtain the temporal and spatial distribution characteristics of each factor, the temporal and spatial distribution characteristics of each factor are combined with the change trend of the non-vegetation covered area to predict and process, generate trend prediction results, perform quantitative evaluation processing on the trend prediction results, and generate a degradation severity report, including: Using the optimized green space encroachment location map data, the information of the non-vegetation covered area is integrated and processed to obtain a data set suitable for a spatiotemporal data analysis framework; Collecting and preprocessing environmental parameters according to the data set suitable for the spatiotemporal data analysis framework to generate multidimensional spatiotemporal data related to changes in non-vegetated areas; Based on the multidimensional spatiotemporal data related to the changes in non-vegetation covered areas, the temporal and spatial distribution characteristics of each factor in the environmental parameters are analyzed and modeled to obtain a relationship model between different environmental parameters and the changes in non-vegetation covered areas; Using the relationship model between the different environmental parameters and the change of the non-vegetation covered area, the change trend of the non-vegetation covered area is predicted and processed to generate trend prediction results for a period of time in the future; Based on the trend prediction results for the future period, a quantitative assessment is conducted on the expansion speed of the non-vegetation covered area, the scope of influence and the degree of ecological impact of the non-vegetation covered area to generate a degradation severity report.

7. The method according to claim 6, characterized in that The method of using the relationship model between the different environmental parameters and the change of the non-vegetation covered area to predict the change trend of the non-vegetation covered area and generate trend prediction results for a period of time in the future includes: Using the relationship model between the different environmental parameters and the changes in the non-vegetation covered area, selecting a suitable machine learning or statistical model, training the change data of the non-vegetation covered area, and constructing a prediction model that captures the temporal and spatial change rules; According to the requirements of the prediction model that captures the temporal and spatial variation patterns, the pre-processed multi-dimensional spatiotemporal data are integrated, and the non-vegetation covered area variation data including the historical period and the latest observations are used as input to generate an input data set suitable for prediction; Based on the input data set suitable for prediction, different future scenario assumptions are set to simulate the changing trends of different non-vegetation covered areas that may occur in the future, and generate multiple prediction scenarios covering the best case, the most likely case and the worst case. The different future scenario assumptions include different paths of climate change and changes in the intensity of human activities; By using the multiple prediction scenarios covering the best case, the most likely case and the worst case, the short-term and long-term change trends of the non-vegetation covered area are predicted and processed to generate a series of trend prediction results at future time points.

8. The method according to claim 2, characterized in that: The dynamic information reflecting the change of green space over time is vectorized based on the land boundary information recorded in the historical geographic information system and the legal provisions, so as to establish a digital model as a reference for the green space boundary, and the digital model is superimposed and compared with the latest remote sensing image to obtain the corrected green space boundary, including: Using historical geographic information system records and legally required land boundary information, vectorize the dynamic information reflecting the changes of green space over time to obtain vector data that accurately represents the changes of green space and the legal boundaries of green space changes; Based on the vector data accurately representing the green space transition and the legal boundaries of the green space transition, a digital model is established and used as a reference for the green space boundary, converted into a digital format, and a high-precision digital model of the green space transition and boundary position is generated; Acquire the latest remote sensing images with high resolution, high precision and high timeliness, make the coordinate system of the digital model consistent with that of the latest remote sensing images through image registration technology, and perform superposition and comparison processing to obtain the detection result of the green space boundary change; Based on the detection result of the green space boundary change, combined with field investigation or other auxiliary data sources, the dynamic information reflecting the change of the green space over time is adjusted and corrected to generate a corrected green space boundary.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a method for identifying encroachment and degradation of urban green space based on computer vision recognition as described in any one of claims 2-8.

10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, the method for identifying encroachment and degradation of urban green space based on computer vision recognition as described in any one of claims 2 to 8 is implemented.

Images