A method for extracting urban green space from multi-source remote sensing images in cooperation

Abstract

This invention relates to the field of remote sensing image processing technology, and in particular to a method for extracting urban green spaces from collaborative multi-source remote sensing imagery. The method includes acquiring multi-source remote sensing image data to construct an urban green space information map, preprocessing the images and extracting features to generate image feature vectors, fusing them through semantic mapping to generate a comprehensive feature vector, training a green space extraction model for small-sample scenarios based on a transfer learning framework, and using real-time imagery to determine green space types and update the map. This invention can improve the accuracy of urban green space extraction, adapt to small-sample scenarios, and achieve dynamic monitoring and information improvement.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and urban green space information extraction technology, specifically a method for extracting urban green space from collaborative multi-source remote sensing images. Background Technology

[0002] Traditional research in the comprehensive assessment of the multifunctionality of urban green spaces has focused primarily on the analysis of single functions, neglecting the synergistic promotion of ecological, social, and economic benefits. Utilizing remote sensing technology to acquire macro-level data such as urban green space coverage and vegetation biomass allows for a large-scale understanding of the ecological status of green spaces and provides a foundation for quantifying their multifunctionality. By establishing regression models with ecological indicators of green spaces as independent variables, combined with variables reflecting social and economic benefits (such as indicators of improved quality of life for residents and data related to regional economic growth), the specific roles of urban green spaces in reducing the urban heat island effect, improving quality of life, and promoting economic development can be revealed. However, existing research still has shortcomings in the synergistic quantitative assessment of the multifunctionality of green spaces, particularly lacking systematic methods for multi-source data integration and dynamic monitoring.

[0003] Currently, some progress has been made in spatial analysis of urban green spaces using GIS technology. Leveraging GIS's powerful spatial data processing capabilities, accurate green space distribution maps can be created, the boundaries and extent of different types of green spaces can be delineated, and coverage rates can be calculated to analyze regional differences. Simultaneously, multi-temporal remote sensing imagery data can be used to dynamically monitor the changing trends of urban green spaces, identifying increases and decreases in green space area and type transformations. Nevertheless, existing methods still have room for improvement in the collaborative utilization of multi-source remote sensing imagery, particularly in terms of adaptability to complex urban environments and multi-functional extraction capabilities.

[0004] The "High-Resolution Urban Green Space Identification Method" (publication number CN118351454B) achieves high-precision identification of green spaces along roadsides using a deep learning model. This method consists of two steps: preliminary identification and refined identification, improving accuracy and transferability. However, it primarily relies on high-resolution remote sensing imagery and fails to effectively integrate LiDAR point cloud data and thermal infrared imagery, limiting its effectiveness in complex terrain or areas with low vegetation cover. Furthermore, deep learning models have high requirements for high-quality labeled data, and the diversity and dynamic changes of urban green space scenes may affect the model's generalization ability, especially when applied across regions.

[0005] The aforementioned problems indicate that existing urban green space extraction methods still have room for improvement in terms of collaborative utilization of multi-source remote sensing data, adaptability to complex scenarios, and model generalization ability. Therefore, this invention aims to construct a more robust and efficient extraction model by integrating multi-source data such as optical imagery, radar data, and thermal infrared imagery to meet the needs of urban ecological monitoring and planning management. Summary of the Invention

[0006] The purpose of this invention is to provide a method for extracting urban green space from collaborative multi-source remote sensing images, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for extracting urban green space from collaborative multi-source remote sensing imagery, the method comprising:

[0008] Acquire multi-source remote sensing image data of urban green spaces, extract basic features from the images and label green space type, and construct an urban green space information map that includes green space type, associated area and historical monitoring records;

[0009] Multi-source remote sensing image data is preprocessed to obtain a standardized image set. A multimodal feature extraction network is used to extract the local detail features and global distribution features of each image in the standardized image set to generate image feature vectors.

[0010] The image feature vector is semantically mapped to the green space type entities, regional relationships and historical monitoring records in the urban green space information map. The image features and information map content are integrated through the reasoning module to generate a comprehensive feature vector.

[0011] A green space extraction model is built based on the transfer learning framework. A small number of labeled comprehensive feature vectors are used as training samples, and unlabeled comprehensive feature vectors are used as test samples. The model parameters are iteratively optimized to train a green space extraction model suitable for small sample scenarios.

[0012] The real-time images of the urban green space to be extracted are input into the extraction model to extract the comprehensive feature vector of the real-time image. The similarity probability between the comprehensive feature vector and each type of green space in the training sample is calculated. If the similarity probability exceeds the preset threshold, the area is determined to belong to the corresponding type of green space. The threshold and urban green space information map are updated regularly based on newly added monitoring data.

[0013] Preferably, the specific method for acquiring multi-source remote sensing image data of urban green spaces, extracting basic features from the images and labeling them with green space type tags, and constructing an urban green space information map containing green space types, associated areas, and historical monitoring records is as follows:

[0014] Raw image data, including optical images, radar data, and thermal infrared images, are collected from multi-source sensors in the urban green space monitoring system.

[0015] Radiometric and geometric corrections are performed on the original image data, and spectral, textural, and temperature features are extracted from the image as basic features.

[0016] Based on historical monitoring reports, the corresponding green space type labels are marked according to the basic characteristics. Green space types include park green space, protective green space, and ancillary green space.

[0017] By using green space type as the core node of the information map, area name and historical monitoring time as node attributes, and the relationship between green space type and corresponding area as the relationship edge of the information map, an urban green space information map is constructed.

[0018] Preferably, the specific method for preprocessing multi-source remote sensing image data to obtain a standardized image set, and then using a multimodal feature extraction network to extract local detail features and global distribution features of each image in the standardized image set to generate image feature vectors is as follows:

[0019] Multi-source remote sensing image data are processed to unify resolution and suppress noise, resulting in a standardized image set with consistent spatial resolution.

[0020] The standardized image set is input into a multimodal feature extraction network. The network includes a shallow feature extraction module and a deep feature extraction module. The shallow feature extraction module extracts local detail features such as edges and corners of the image, while the deep feature extraction module extracts global distribution features such as the overall structure and green space distribution of the image.

[0021] Local detail features and global distribution features are stitched together to generate an image feature vector for each image.

[0022] Preferably, the specific method for semantically mapping the image feature vector with green space type entities, regional relationships, and historical monitoring records in the urban green space information map, and then fusing the image features and information map content through the inference module to generate a comprehensive feature vector is as follows:

[0023] Semantic transformation is performed on the image feature vectors to obtain feature representations consistent with the dimensions of the core nodes of the information graph;

[0024] In the urban green space information map, based on the attribute information of green space type entities and the association strength of relation edges, the probability of green space distribution and the characteristic distribution of historically similar green spaces in the current image area are calculated.

[0025] The semantically transformed image features are weighted and summed with the probability of green space distribution and the distribution of historically similar green space features to generate a comprehensive feature vector that integrates image information and map content.

[0026] Preferably, the specific method for constructing a green space extraction model based on a transfer learning framework, using a small number of labeled comprehensive feature vectors as training samples and unlabeled comprehensive feature vectors as test samples, and training a green space extraction model suitable for small sample scenarios by iteratively optimizing the model parameters is as follows:

[0027] A green space extraction model is constructed, which includes a feature encoder and a classifier. The feature encoder is used to extract the deep representation of the comprehensive feature vector, and the classifier is used to calculate the similarity probability between the comprehensive feature vector and each green space type.

[0028] A small number of comprehensive feature vectors for each type of green space were selected from historical monitoring data as training samples, and the remaining comprehensive feature vectors were used as test samples.

[0029] In the transfer learning process, the training samples are used as the learning task and the test samples are used as the validation task. The model parameters are updated by minimizing the prediction error of the classifier, so that the model can quickly adapt to new small sample green space types.

[0030] Preferably, the step of inputting real-time images of the urban green space to be extracted into the extraction model, extracting the comprehensive feature vector of the real-time images, calculating the similarity probability between the comprehensive feature vector and each type of green space in the training samples, and determining that the area belongs to the corresponding type of green space if the similarity probability exceeds a preset threshold, and periodically updating the threshold and urban green space information map based on newly added monitoring data, is as follows:

[0031] The real-time images of urban green spaces to be extracted are preprocessed, feature extracted, and information fused sequentially to obtain a comprehensive feature vector of the real-time images;

[0032] The comprehensive feature vector of real-time images is input into the classifier of the extraction model to calculate the similarity probability between it and types such as park green space, protective green space and attached green space in the training samples.

[0033] The preset threshold is the average similarity probability of each green space type in the training samples. If the similarity probability of the real-time image is greater than the threshold, the area is determined to belong to the corresponding type of green space.

[0034] New monitoring data is collected monthly, and the comprehensive feature vectors of new green space types are added to the training samples to update the green space type entities and relationship edges in the urban green space information map.

[0035] Preferably, the specific method for acquiring raw image data from multi-source sensors in the urban green space monitoring system, including optical images, radar data, and thermal infrared images, is as follows:

[0036] In the urban green space monitoring system, a high-resolution camera is set up on the optical sensor to collect surface reflectivity images, a synthetic aperture radar is set up on the radar sensor to collect surface structure images, and a long-wave infrared camera is set up on the thermal infrared sensor to collect surface temperature images.

[0037] The acquisition frequency of each sensor is synchronized with the monitoring cycle.

[0038] Preferably, the specific method for constructing an urban green space information map that includes green space types, associated areas, and historical monitoring records is as follows:

[0039] Using green space type as the core node, add a regional attribute field to each green space type node to record the associated regional name, and add a time attribute field to record the first monitoring time and the most recent monitoring time;

[0040] Using the region name as the associated node, establish the association edge between green space type nodes and region nodes, and establish the temporal evolution relationship edge between green space type nodes with different time attributes;

[0041] The green space area, green space coverage rate, and ecological function indicators from historical monitoring reports will be used as additional descriptive information for nodes to improve the content of the urban green space information map.

[0042] Preferably, the specific method for optimizing model parameters through transfer learning and training a green space extraction model suitable for small sample scenarios is as follows:

[0043] The transfer learning process includes an initialization phase and an optimization phase. The initialization phase uses a large number of non-green area images to pre-train the feature encoder, enabling the feature encoder to learn general image representation capabilities.

[0044] In the optimization phase, multiple small sample tasks are selected, each containing training and test samples of different green space types. The parameters of the feature encoder and classifier are jointly optimized by calculating the classification loss of the training samples within the task and the generalization loss of the test samples between tasks.

[0045] When the generalization loss fails to decrease for three consecutive iterations, the transfer learning process is stopped, and the trained green area extraction model is obtained.

[0046] Preferably, the specific method for jointly optimizing the parameters of the feature encoder and classifier by calculating the classification loss of training samples within the task and the generalization loss of test samples between tasks is as follows:

[0047] For the training samples of each few-sample task, the cross-entropy loss between the similarity probability output by the classifier and the real green area label is calculated as the intra-task classification loss; for the test samples of all few-sample tasks, the variance of the prediction accuracy of the model across different tasks is calculated as the inter-task generalization loss.

[0048] The intra-task classification loss and inter-task generalization loss are added together with a weight ratio of 7:3 to obtain the total loss function. The parameters of the feature encoder and classifier are updated by minimizing the total loss function.

[0049] This method for extracting urban green space from collaborative multi-source remote sensing imagery has several beneficial effects, as detailed below:

[0050] 1. By integrating multi-source remote sensing data such as optical imagery, radar data, and thermal infrared imagery, it can comprehensively capture the spectral, texture, and temperature characteristics of urban green spaces, reducing the limitations of single data sources. A multimodal feature extraction network extracts local detail features and global distribution features, and then semantically maps and fuses these features with urban green space information maps to generate a comprehensive feature vector, enriching feature representation and making the extraction results more accurate.

[0051] 2. A green area extraction model is constructed based on the transfer learning framework. A feature encoder is pre-trained using a large number of non-green area images to enable it to have general image representation capabilities. Then, it is trained with a small number of labeled samples, which can achieve good model performance under small sample conditions and reduce the dependence on large-scale labeled data.

[0052] 3. Achieve dynamic monitoring and information improvement: By using real-time image input models to determine green space types, the current status of urban green spaces can be grasped in a timely manner. Regularly updating model thresholds and urban green space information maps based on newly added monitoring data can dynamically track green space changes, continuously improve green space information, and provide continuous and accurate data support for the management, planning, and research of urban green spaces.

[0053] 4. Enhanced adaptability to complex environments: The collaborative utilization of multi-source data and the generation of comprehensive feature vectors enable the method to still identify and extract green space information well in complex urban environments, such as when there are building obstructions or complex terrain, thus improving the applicability of the method. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the method of the present invention.

[0055] Figure 2 This is a schematic diagram of the structure of an information map of urban green spaces.

[0056] Figure 3 This is a schematic diagram of the structure of a multimodal feature extraction network.

[0057] Figure 4 This is a schematic diagram of the training process for a transfer learning framework. Detailed Implementation

[0058] This invention provides a method for extracting urban green space from collaborative multi-source remote sensing imagery, the specific implementation of which is as follows. According to... Figure 1The illustrated workflow begins with acquiring multi-source remote sensing imagery data of urban green spaces. First, raw imagery data, including optical images, radar data, and thermal infrared images, is collected using multi-source sensors within the urban green space monitoring system. During this process, optical sensors utilize high-resolution cameras to acquire surface reflectance images, radar sensors use synthetic aperture radar to acquire surface structure images, and thermal infrared sensors employ long-wave infrared cameras to acquire surface temperature images. The acquisition frequency of each sensor is synchronized with the monitoring cycle to ensure data temporal consistency.

[0059] The acquired raw image data needs to undergo radiometric and geometric correction to eliminate image distortion caused by equipment errors or atmospheric effects. After correction, spectral, textural, and temperature features are extracted from the images as basic features. These features are crucial for subsequent analysis. To label green space types, the basic features need to be classified based on information recorded in historical monitoring reports. Green space types mainly include park green spaces, protective green spaces, and ancillary green spaces. Based on this, a system is constructed... Figure 2 The urban green space information map shown depicts nodes representing green space types and associated regions, and edges representing the relationships and temporal evolution between regions and green space types. Specifically, green space type is the core node, with a regional attribute field recording the name of the associated region and a time attribute field recording the first monitoring time and the most recent monitoring time. Simultaneously, using region names as associated nodes, edges are established to link green space type nodes and region nodes, and further, edges are established to link the temporal evolution relationships between green space type nodes with different time attributes. Furthermore, green space area, green space coverage rate, and ecological function indicators from historical monitoring reports are used as additional descriptive information for the nodes, enriching the content of the information map.

[0060] Next, the multi-source remote sensing image data is preprocessed to obtain a standardized image set. For example... Figure 1 As shown, the preprocessing process includes resolution normalization and noise suppression, aiming to ensure that images from different sources have the same spatial resolution, thus facilitating subsequent analysis. The normalized image set after preprocessing is then input. Figure 3 The multimodal feature extraction network shown consists of a shallow feature extraction module and a deep feature extraction module. The shallow feature extraction module is responsible for extracting local detail features such as edges and corners of the image, while the deep feature extraction module is used to extract global distribution features such as the overall structure of the image and the distribution of green areas. The two feature sets are concatenated to generate an image feature vector for each image. This multimodal feature extraction method can comprehensively capture information in the image, providing high-quality data support for subsequent steps.

[0061] The generated image feature vectors need to be semantically mapped to the content in the urban green space information map. This process is implemented through an inference module, specifically by semantically transforming the image feature vectors to align their dimensions with the core nodes of the information map. Subsequently, based on the attribute information of green space type entities and the association strength of relation edges in the information map, the probability of green space distribution in the current image's corresponding area and the feature distribution of historically similar green spaces are calculated. Finally, the semantically transformed image features are weighted and summed with the green space distribution probability and the feature distribution of historically similar green spaces to generate a comprehensive feature vector that integrates image information and map content. This process fully leverages the complementarity of multi-source data, improving feature representation capabilities.

[0062] Building a green space extraction model based on a transfer learning framework is one of the core components of this invention. For example... Figure 4 As shown, the model consists of two main parts: a feature encoder and a classifier. The feature encoder extracts deep representations of the comprehensive feature vector, while the classifier calculates the similarity probability between the comprehensive feature vector and each green space type. In the transfer learning process, an initialization phase is first performed, using a large number of non-green space images to pre-train the feature encoder, giving it general image representation capabilities. Next, an optimization phase is initiated, selecting multiple small-sample tasks, each containing training and test samples of different green space types. For the training samples of each small-sample task, the cross-entropy loss between the similarity probability output by the classifier and the real green space label is calculated as the intra-task classification loss; while for the test samples of all small-sample tasks, the variance of the model's prediction accuracy across different tasks is calculated as the inter-task generalization loss. The intra-task classification loss and the inter-task generalization loss are added with a weighted ratio of 7:3 to obtain the total loss function. The parameters of the feature encoder and classifier are updated by minimizing the total loss function. When the generalization loss fails to decrease for three consecutive iterations, the transfer learning process stops, resulting in a trained green space extraction model.

[0063] After model training is complete, real-time images of urban green spaces to be extracted can be analyzed. First, the images are preprocessed, feature extracted, and information fused sequentially to obtain a comprehensive feature vector for the real-time images. This vector is then input into the classifier of the extraction model to calculate its similarity probability with types such as park green spaces, protective green spaces, and ancillary green spaces in the training samples. A preset threshold is the average similarity probability of each green space type in the training samples. If the similarity probability of the real-time image is greater than the threshold, the area is determined to belong to the corresponding type of green space. New monitoring data is collected monthly, and the comprehensive feature vectors of the new green space types are added to the training samples. Simultaneously, the green space type entities and relation edges in the urban green space information map are updated. This dynamic update mechanism ensures that the model can adapt to the changing trends of urban green spaces, improving its long-term accuracy.

[0064] In practical applications, this method can be widely used in urban planning and ecological protection. For example, in a green space monitoring project in a certain city, the above method was deployed to successfully achieve automated identification and classification of large areas of urban green space. Specifically, remote sensing image data covering the entire city is first collected using multi-source sensors, and then processed step-by-step according to the aforementioned steps to construct an urban green space information map. Subsequently, a green space extraction model is trained using a transfer learning framework and applied to real-time image analysis to quickly locate newly added or changed green spaces. Furthermore, the model and information map are updated regularly, enabling the system to continuously track the dynamic changes in urban green spaces and provide scientific decision-making support for relevant departments.

[0065] Throughout the implementation process, the connections and collaboration between the various components are crucial. For example, the shallow and deep feature extraction modules of the multimodal feature extraction network form a complete feature extraction chain through hierarchical connections. The former focuses on extracting local detailed features, while the latter emphasizes capturing global distribution features. Together, they generate high-quality image feature vectors. In the transfer learning framework, the feature encoder and classifier achieve joint optimization by sharing parameters, ensuring the model performs well in small sample scenarios. Furthermore, the nodes and relational edges of the urban green space information map are closely linked through attribute fields and association strength, providing a reliable foundation for semantic mapping. The rational design and efficient collaboration of these components jointly support the implementation of the technical solution of this invention.

[0066] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0067] In a green space monitoring project in a large city, the method of this invention was used for automated identification and classification of urban green spaces. First, remote sensing image data covering the entire city was collected by deploying multi-source sensors. Optical sensors used high-resolution cameras to acquire surface reflectance images, radar sensors used synthetic aperture radar to acquire surface structure images, and thermal infrared sensors used long-wave infrared cameras to acquire surface temperature images. The acquisition frequency of each sensor was synchronized with the monitoring cycle to ensure data temporal consistency. The acquired raw image data then underwent radiometric and geometric correction processing to eliminate image distortion caused by equipment errors or atmospheric effects. After correction, spectral, textural, and temperature features were extracted from the images as basic features, and combined with historical monitoring reports to label green space types, including park green spaces, protective green spaces, and ancillary green spaces. Based on this, a system was constructed... Figure 2The urban green space information map shown depicts nodes representing green space types and associated regions, and edges representing the relationships and temporal evolution between regions and green space types. Specifically, green space type is the core node, with a regional attribute field recording the name of the associated region and a time attribute field recording the first monitoring time and the most recent monitoring time. Simultaneously, using region names as associated nodes, edges are established to link green space type nodes and region nodes, and further, edges are established to link the temporal evolution relationships between green space type nodes with different time attributes. Furthermore, green space area, green space coverage rate, and ecological function indicators from historical monitoring reports are used as additional descriptive information for the nodes, enriching the content of the information map.

[0068] Next, the multi-source remote sensing image data is preprocessed to obtain a standardized image set. For example... Figure 1 As shown, the preprocessing process includes resolution normalization and noise suppression, aiming to ensure that images from different sources have the same spatial resolution, thus facilitating subsequent analysis. The normalized image set after preprocessing is then input. Figure 3 The multimodal feature extraction network shown consists of a shallow feature extraction module and a deep feature extraction module. The shallow feature extraction module is responsible for extracting local detail features such as edges and corners of the image, while the deep feature extraction module is used to extract global distribution features such as the overall structure of the image and the distribution of green areas. The two feature sets are concatenated to generate the image feature vector for each image. This multimodal feature extraction method can comprehensively capture information in the image, providing high-quality data support for subsequent steps.

[0069] The generated image feature vectors need to be semantically mapped to the content in the urban green space information map. This process is implemented through an inference module, specifically by semantically transforming the image feature vectors to align their dimensions with the core nodes of the information map. Subsequently, based on the attribute information of green space type entities and the association strength of relation edges in the information map, the probability of green space distribution in the current image's corresponding area and the feature distribution of historically similar green spaces are calculated. Finally, the semantically transformed image features are weighted and summed with the green space distribution probability and the feature distribution of historically similar green spaces to generate a comprehensive feature vector that integrates image information and map content. This process fully leverages the complementarity of multi-source data, improving feature representation capabilities.

[0070] Building a green space extraction model based on a transfer learning framework is one of the core components of this invention. For example... Figure 4As shown, the model consists of two main parts: a feature encoder and a classifier. The feature encoder extracts deep representations of the comprehensive feature vector, while the classifier calculates the similarity probability between the comprehensive feature vector and each green space type. In the transfer learning process, an initialization phase is first performed, using a large number of non-green space images to pre-train the feature encoder, giving it general image representation capabilities. Next, an optimization phase is initiated, selecting multiple small-sample tasks, each containing training and test samples of different green space types. For the training samples of each small-sample task, the cross-entropy loss between the similarity probability output by the classifier and the real green space label is calculated as the intra-task classification loss; while for the test samples of all small-sample tasks, the variance of the model's prediction accuracy across different tasks is calculated as the inter-task generalization loss. The intra-task classification loss and the inter-task generalization loss are added with a weighted ratio of 7:3 to obtain the total loss function. The parameters of the feature encoder and classifier are updated by minimizing the total loss function. When the generalization loss fails to decrease for three consecutive iterations, the transfer learning process stops, resulting in a trained green space extraction model.

[0071] After model training is complete, real-time images of urban green spaces to be extracted can be analyzed. First, the images are preprocessed, feature extracted, and information fused sequentially to obtain a comprehensive feature vector for the real-time images. This vector is then input into the classifier of the extraction model to calculate its similarity probability with types such as park green spaces, protective green spaces, and ancillary green spaces in the training samples. A preset threshold is the average similarity probability of each green space type in the training samples. If the similarity probability of the real-time image is greater than the threshold, the area is determined to belong to the corresponding type of green space. New monitoring data is collected monthly, and the comprehensive feature vectors of the new green space types are added to the training samples. Simultaneously, the green space type entities and relation edges in the urban green space information map are updated. This dynamic update mechanism ensures that the model can adapt to the changing trends of urban green spaces, improving its long-term accuracy.

[0072] Throughout the implementation process, the connections and collaboration between the various components are crucial. For example, the shallow and deep feature extraction modules of the multimodal feature extraction network form a complete feature extraction chain through hierarchical connections. The former focuses on extracting local detailed features, while the latter emphasizes capturing global distribution features. Together, they generate high-quality image feature vectors. In the transfer learning framework, the feature encoder and classifier achieve joint optimization by sharing parameters, ensuring the model performs well in small sample scenarios. Furthermore, the nodes and relational edges of the urban green space information map are closely linked through attribute fields and association strength, providing a reliable foundation for semantic mapping. The rational design and efficient collaboration of these components jointly support the implementation of the technical solution of this invention.

[0073] Through the steps described above, this method successfully achieves automated identification and classification of large areas of urban green space. Regularly updating the model and infographics allows the system to continuously track the dynamic changes in urban green spaces, providing relevant departments with a scientific basis for decision-making.

[0074] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for extracting urban green space from collaborative multi-source remote sensing imagery, characterized in that, include: The method involves acquiring multi-source remote sensing image data of urban green spaces, extracting basic features from the images, labeling green space types, and constructing an urban green space information map that includes green space types, associated areas, and historical monitoring records. S1. Collect raw image data from multi-source sensors in the urban green space monitoring system, including optical images, radar data, and thermal infrared images; S2. Perform radiometric and geometric correction on the original image data, and extract spectral, texture and temperature features from the image as basic features; S3. Label the corresponding green space type based on the historical monitoring report as the basic feature. Green space types include park green space, protective green space, and ancillary green space. S4. Using green space type as the core node of the information map, and area name and historical monitoring time as node attributes, the relationship between green space type and corresponding area as the relationship edge of the information map, to construct an urban green space information map; Multi-source remote sensing image data is preprocessed to obtain a standardized image set. A multimodal feature extraction network is used to extract the local detail features and global distribution features of each image in the standardized image set to generate image feature vectors. The image feature vector is semantically mapped to green space type entities, regional relationships, and historical monitoring records in the urban green space information map. A comprehensive feature vector is generated by fusing image features and information map content through an inference module. The specific method is as follows: S1. Perform semantic transformation on the image feature vectors to obtain feature representations consistent with the dimensions of the core nodes of the information graph; S2. In the urban green space information map, based on the attribute information of green space type entities and the association strength of relation edges, calculate the green space distribution probability and the characteristic distribution of historically similar green spaces in the area corresponding to the current image; S3. The semantically transformed image features are weighted and summed with the green space distribution probability and the historical similar green space feature distribution to generate a comprehensive feature vector that integrates image information and map content; A green space extraction model is built based on the transfer learning framework. A small number of labeled comprehensive feature vectors are used as training samples, and unlabeled comprehensive feature vectors are used as test samples. The model parameters are iteratively optimized to train a green space extraction model suitable for small sample scenarios. The real-time images of the urban green space to be extracted are input into the extraction model to extract the comprehensive feature vector of the real-time image. The similarity probability between the comprehensive feature vector and each type of green space in the training sample is calculated. If the similarity probability exceeds the preset threshold, the area is determined to belong to the corresponding type of green space. The threshold and urban green space information map are updated regularly based on newly added monitoring data.

2. The method for extracting urban green space from collaborative multi-source remote sensing imagery as described in claim 1, characterized in that, The method for preprocessing multi-source remote sensing image data to obtain a standardized image set, and then using a multimodal feature extraction network to extract local detail features and global distribution features of each image in the standardized image set to generate image feature vectors is as follows: Multi-source remote sensing image data are processed to unify resolution and suppress noise, resulting in a standardized image set with consistent spatial resolution. The standardized image set is input into a multimodal feature extraction network. The network includes a shallow feature extraction module and a deep feature extraction module. The shallow feature extraction module extracts the local details of the edges and corners of the image, while the deep feature extraction module extracts the overall structure and global distribution features of the green space in the image. Local detail features and global distribution features are stitched together to generate an image feature vector for each image.

3. The method for extracting urban green space from collaborative multi-source remote sensing imagery as described in claim 1, characterized in that, The specific method for constructing a green space extraction model based on the transfer learning framework, using a small number of labeled comprehensive feature vectors as training samples and unlabeled comprehensive feature vectors as test samples, and training a green space extraction model suitable for small sample scenarios through iterative optimization of model parameters is as follows: A green space extraction model is constructed, which includes a feature encoder and a classifier. The feature encoder is used to extract the deep representation of the comprehensive feature vector, and the classifier is used to calculate the similarity probability between the comprehensive feature vector and each green space type. A small number of comprehensive feature vectors for each type of green space were selected from historical monitoring data as training samples, and the remaining comprehensive feature vectors were used as test samples. In the transfer learning process, the training samples are used as the learning task and the test samples are used as the validation task. The model parameters are updated by minimizing the prediction error of the classifier, so that the model can quickly adapt to new small sample green space types.

4. The method for extracting urban green space from collaborative multi-source remote sensing imagery as described in claim 3, characterized in that, The process involves inputting real-time images of the urban green space to be extracted into the extraction model, extracting the comprehensive feature vector of the real-time images, calculating the similarity probability between the comprehensive feature vector and each type of green space in the training samples, and determining that the area belongs to the corresponding type of green space if the similarity probability exceeds a preset threshold. The specific method for periodically updating the threshold and urban green space information map based on newly added monitoring data is as follows: The real-time images of urban green spaces to be extracted are preprocessed, feature extracted, and information fused sequentially to obtain a comprehensive feature vector of the real-time images; The comprehensive feature vector of real-time images is input into the classifier of the extraction model to calculate the similarity probability between it and the park green space, protective green space and attached green space types in the training samples. The preset threshold is the average similarity probability of each green space type in the training samples. If the similarity probability of the real-time image is greater than the threshold, the area is determined to belong to the corresponding type of green space. New monitoring data is collected monthly, and the comprehensive feature vectors of new green space types are added to the training samples to update the green space type entities and relationship edges in the urban green space information map.

5. The method for extracting urban green space from collaborative multi-source remote sensing imagery as described in claim 4, characterized in that, The specific method for acquiring raw image data from multi-source sensors in the urban green space monitoring system, including optical images, radar data, and thermal infrared images, is as follows: In the urban green space monitoring system, a high-resolution camera is set up on the optical sensor to collect surface reflectivity images, a synthetic aperture radar is set up on the radar sensor to collect surface structure images, and a long-wave infrared camera is set up on the thermal infrared sensor to collect surface temperature images. The acquisition frequency of each sensor is synchronized with the monitoring cycle.

6. The method for extracting urban green space from collaborative multi-source remote sensing imagery as described in claim 5, characterized in that, The specific method for constructing an urban green space information map that includes green space types, associated areas, and historical monitoring records is as follows: Using green space type as the core node, add a regional attribute field to each green space type node to record the associated regional name, and add a time attribute field to record the first monitoring time and the most recent monitoring time; Using the region name as the associated node, establish the association edge between green space type nodes and region nodes, and establish the temporal evolution relationship edge between green space type nodes with different time attributes; The green space area, green space coverage rate, and ecological function indicators from historical monitoring reports will be used as additional descriptive information for nodes to improve the content of the urban green space information map.

7. The method for extracting urban green space from collaborative multi-source remote sensing imagery as described in claim 6, characterized in that, The specific method for constructing a green space extraction model based on the transfer learning framework, optimizing model parameters through transfer learning, and training a green space extraction model suitable for small sample scenarios is as follows: The transfer learning process includes an initialization phase and an optimization phase. The initialization phase uses a large number of non-green area images to pre-train the feature encoder, enabling the feature encoder to learn general image representation capabilities. In the optimization phase, multiple small sample tasks are selected, each containing training and test samples of different green space types. The parameters of the feature encoder and classifier are jointly optimized by calculating the classification loss of the training samples within the task and the generalization loss of the test samples between tasks. When the generalization loss fails to decrease for three consecutive iterations, the transfer learning process is stopped, and the trained green area extraction model is obtained.

8. The method for extracting urban green space from collaborative multi-source remote sensing imagery as described in claim 7, characterized in that, The specific method for jointly optimizing the parameters of the feature encoder and classifier by calculating the classification loss of training samples within the task and the generalization loss of test samples between tasks is as follows: For each training sample of a few-sample task, the cross-entropy loss between the similarity probability output by the classifier and the real green area label is calculated as the intra-task classification loss. For all test samples of small sample tasks, the variance of the model's prediction accuracy across different tasks is calculated as the generalization loss between tasks. The intra-task classification loss and inter-task generalization loss are added together with a weight ratio of 7:3 to obtain the total loss function. The parameters of the feature encoder and classifier are updated by minimizing the total loss function.

Images