Semantic segmentation-based ecological space management and control area construction activity identification method and system

Abstract

The invention provides an ecological space management and control area construction activity identification method and system based on semantic segmentation. The recognition method comprises the following steps: S1, segmenting two time-phase earth satellite remote sensing images uploaded by a user, inputting the segmented images into a pre-trained semantic segmentation model and a pre-trained change detection model, and outputting a semantic segmentation result and a change detection result; s2, performing user-defined filtering on the change detection result according to the semantic segmentation result to obtain an accurate change detection result, and storing the accurate change detection result; and S3, establishing an image pyramid for the two-time-phase geostationary satellite remote sensing images uploaded by the user, dynamically loading the two-time-phase geostationary satellite remote sensing images with corresponding resolutions according to the zoom level, and displaying the two-time-phase geostationary satellite remote sensing images and the accurate change detection result layer converted into the geojson format in a roller shutter mode. According to the method, a change detection filtering mechanism based on semantic segmentation is adopted, the capacity of different monitoring requirements is met, the detection accuracy is improved, and personalized monitoring requirements are allowed.

Description

Technical Field

[0001] The present invention relates to the technical field of ecological space control zone construction, and specifically to an ecological space control zone construction activity identification method based on semantic segmentation and an ecological space control zone construction activity identification method system based on semantic segmentation. Background Art

[0002] The current red line change monitoring methods in ecological space control areas are mainly the following:

[0003] (1) Manual visual method: Manual visual translation of dual-phase remote sensing data within the monitoring range is performed to mark the range of changes.

[0004] (2) Direct image comparison method. Directly calculate and transform the pixel values ​​in the two registered remote sensing images to find the changed areas. Currently, the commonly used direct comparison methods of spectral data include image difference method, image ratio method, vegetation index comparison method, principal component analysis method, spectral feature variation method, false color synthesis method, band replacement method, change vector analysis method, band cross-correlation analysis and hybrid detection method.

[0005] (3) Post-classification comparison method. First, classify the remote sensing data of the two phases that have been registered, and then compare the classification results to obtain change detection information. Therefore, in the context of intelligent environmental protection, using machine recognition methods such as deep learning to fully explore and utilize environmental data such as remote sensing images for ecological environment monitoring is not only a research hotspot in this field, but also a new technology to assist manual interpretation, reduce workload, and improve efficiency.

[0006] However, the existing monitoring methods have the following problems:

[0007] Manual visual inspection is inefficient and prone to omissions or subjective judgments. First, manual visual translation requires a lot of time and human resources. When processing large-scale image data, a lot of manual labor is required, which significantly increases the time cost of the monitoring project. Secondly, hiring and training a sufficient number and quality of personnel to perform visual translation is an expensive task. Companies must bear high labor costs, which is not conducive to maintaining competitiveness. In addition, as the scale of monitoring projects increases, traditional manual visual translation methods become impractical. It is difficult to cope with large-scale data sets and vast geographical areas, thus limiting the scalability of monitoring capabilities. Finally, manual visual translation is prone to omissions, errors, or subjective judgments. This may lead to false alarms or miss important environmental changes, affecting the accuracy of monitoring.

[0008] Pixel-based classification or clustering methods such as SVM and decision trees are also difficult to meet the needs of large-scale remote sensing information monitoring due to problems such as fragmented results and low universality. Summary of the invention

[0009] Based on this, it is necessary to provide an ecological space control area construction activity identification method and system based on semantic segmentation to address the problems of fragmented results and low universality of existing monitoring methods.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] The method for identifying construction activities in ecological space control areas based on semantic segmentation includes the following steps:

[0012] S1. Segment the satellite remote sensing images of the same place at two times uploaded by the user and input them into the pre-trained semantic segmentation model and change detection model, and output the semantic segmentation results and change detection results;

[0013] S2. Customize and filter the change detection results according to the semantic segmentation results to obtain and store accurate change detection results. The specific steps of custom filtering are as follows:

[0014] S21. The segmented images of the semantic segmentation results are weighted by the confidence matrix to obtain the corresponding two-phase semantic segmentation label map;

[0015] S22. Calculate the area converted from non-construction land to construction land according to the two-phase semantic segmentation label map, and obtain a valid area mask;

[0016] S23. Filter out the area within a certain range where plots of different land uses meet within the effective area to obtain an accurate effective area mask;

[0017] S24. Filtering the change detection result through the precise effective area mask to obtain the precise change detection result and converting it into a vectorized format for storage;

[0018] S3. Create an image pyramid using the satellite remote sensing images uploaded by users at two different times, dynamically load the satellite remote sensing images at two different times with the corresponding resolution according to the zoom level, and display the satellite remote sensing images at two different times and the precise change detection result layer converted into geojson format in a rolling shutter manner.

[0019] Furthermore, in step S1, the specific steps of training the semantic segmentation model and the change detection model are as follows:

[0020] Pre-screen the dataset of the required resolution from the public image change dataset and use it as the training set for pixel-level annotation of change detection and semantic segmentation;

[0021] The labeled data set is input into the pre-built semantic segmentation model and change detection model for training to obtain the trained semantic segmentation model and change detection model.

[0022] Furthermore, multi-category semantic labels are assigned to the images in the training set for semantic segmentation, and the multi-category semantic labels include water bodies, roads, trees and buildings; and binary classification labels are assigned to the training set for change detection, and the binary classification labels consist of changes and no changes.

[0023] Furthermore, semantic segmentation models and change detection models are stored in a model library; semantic segmentation models include U-Net, FCN, SegNet, PSPNet or DeepLab; change detection models include STANet, BIT-basedNet or AIE-changedet.

[0024] Furthermore, the semantic segmentation model and the change detection model are determined according to the user selection. If the user does not select a model, a default model is used.

[0025] Furthermore, the model library also includes models uploaded by users. After verification and security testing, the models uploaded by users are updated to the model library according to categories to provide users with selection.

[0026] The ecological space control zone construction activity recognition system based on semantic segmentation includes image acquisition module, model selection module, data set organization module, deep learning module, custom filtering module and rendering module.

[0027] The image acquisition module is used to cut the overlapping satellite remote sensing images uploaded by the user into small images for temporary storage;

[0028] The model selection module is used to determine the semantic segmentation model and the change detection model from the model library according to the user selection or the default selection;

[0029] The dataset organization module is used to filter out datasets of required resolution from public image change datasets, and use them as training sets for pixel-level annotation of change detection and semantic segmentation; the annotated datasets are input into the semantic segmentation model and change detection model for training, and trained semantic segmentation models and change detection models are obtained;

[0030] The deep learning module is used to input the segmented images temporarily stored in the image acquisition module into the trained semantic segmentation model and change detection model, and output the semantic segmentation results and change detection results;

[0031] The custom filtering module is used to calculate the mask of the valid area of ​​the two-phase semantic segmentation labels of the semantic segmentation results, filter the change detection results, obtain accurate change detection results and store them;

[0032] The rendering module is used to build an image pyramid from the satellite remote sensing images uploaded by users at two different times, dynamically load the satellite remote sensing images at two different times with the corresponding resolution according to the zoom level, and display the satellite remote sensing images at two different times and the precise change detection result layer converted into geojson format in a rolling shutter manner.

[0033] Furthermore, the ecological space control zone construction activity identification system based on semantic segmentation also includes a custom update module, which is used to receive the model weight file containing the model name and category uploaded by the user, and update it to the model library after verification and security check for user selection.

[0034] Furthermore, in the rendering module, the satellite remote sensing images of the same place at two times in the tif format uploaded by the user are used to generate corresponding pyramid files in the ovr format to establish an image pyramid.

[0035] Furthermore, in the custom filtering module, Hadamard product, Schur product, Pointwise multiplication or Element-wise multiplication matrix operation method is used to realize the masking function.

[0036] Compared with the prior art, the beneficial effects of the present invention include:

[0037] 1. The present invention explores the prediction performance of multiple different models through two paths: the indirect method of semantic segmentation and the direct method of change detection; adopts a change detection filtering mechanism based on semantic segmentation, uses semantic segmentation technology to customize the filtering of change detection results to meet the ability of different monitoring needs, improves the accuracy of detection, reduces false positives, and allows personalized monitoring needs;

[0038] 2. The present invention reduces rendering time by establishing an image pyramid, thereby achieving faster rendering effects; the satellite images and recognition results before and after the change can be compared in the form of a rolling curtain, which improves the efficiency of users in observing and interacting with recognition results;

[0039] 3. The present invention provides greater model flexibility by allowing users to upload and update models to adapt to the continuous changes in data and the evolution of enterprise needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes and are not intended to limit the scope of protection of the present invention. Among them:

[0041] Figure 1 This is a flow chart of the method for identifying construction activities in ecological space control areas based on semantic segmentation introduced in the present invention;

[0042] Figure 2 Path map for the construction activity identification method of ecological space control areas based on semantic segmentation. DETAILED DESCRIPTION

[0043] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the whole of the present invention or as a limitation or restriction to the technical solution of the present invention.

[0044] Example 1

[0045] See also Figure 1 This embodiment introduces a method for identifying construction activities in an ecological space control area based on semantic segmentation, which specifically includes the following steps:

[0046] S1. Segment the satellite remote sensing images of the same place at two times uploaded by the user and input them into the pre-trained semantic segmentation model and change detection model, and output the semantic segmentation results and change detection results.

[0047] In step S1, the established semantic segmentation model and change detection model are trained through public data sets. A large number of public data sets are collected and compared to find data sets with the resolution that best matches the actual needs, and they are annotated at the pixel level for change detection and semantic segmentation: a binary label of "changed" or "unchanged" is assigned to each pixel of a set of images in the training set for change detection, and a multi-classification semantic label is assigned to each pixel of the image in the training set for semantic segmentation, such as "water body", "road", "tree", "building", etc., and then the semantic segmentation model and change detection model are trained.

[0048] The architecture of the semantic segmentation model includes an encoder, a decoder, and a pixel-level classification layer. The change detection model uses ResNet18 as the backbone network to extract feature maps corresponding to two times, encodes and segments the results through convolution to obtain semantic tokens of the number of feature map channels, and obtains contextual information in the spatial and temporal domains through encoding and decoding.

[0049] The semantic segmentation model and change detection model are stored in the model library; the semantic segmentation model can use U-Net, FCN, SegNet, PSPNet, DeepLab, etc.; the change detection model can use STANet, BIT-basedNet, AIE-changedet, etc. The semantic segmentation model and change detection model are determined according to the user's selection. If the user does not select a model, the default model is used.

[0050] For the model library, a user-defined update mechanism is set up to adapt to the continuous changes in data and the continuous evolution of enterprise needs, and to meet the needs of future data updates and model optimization. It supports users to upload updated models as needed and put them into use.

[0051] S2. Customize and filter the change detection results according to the semantic segmentation results to obtain and store accurate change detection results. The specific steps of custom filtering are as follows:

[0052] S21. The segmented images of the semantic segmentation results are weighted by the confidence matrix to obtain the corresponding two-phase semantic segmentation label map;

[0053] S22. Calculate the area converted from non-construction land to construction land according to the two-phase semantic segmentation label map, and obtain a valid area mask;

[0054] S23. Filter out the area within a certain range where plots of different land uses meet within the effective area to obtain an accurate effective area mask;

[0055] S24. Filter the change detection result through the precise effective area mask to obtain the precise change detection result and convert it into a vectorized format for storage.

[0056] In the field of deep learning, the task of identifying development and construction activities in ecological space control areas is often completed with the help of change detection models, which usually rely on pixel-level image changes, which often introduce many false positives because some image changes may not be related to changes in the ecosystem, such as seasonal changes or other non-critical factors; at the same time, more detailed monitoring requirements may be proposed in actual monitoring tasks, and accordingly, there are different definitions of change and non-change. For example, when monitoring human development activities, it is necessary to identify changes from non-construction land to construction land, while in other scenarios, there may be a need to only identify changes in buildings or cultivated land. In order to improve this problem, a custom change detection filtering mechanism based on semantic segmentation technology, namely the custom filtering mechanism, is introduced.

[0057] In step S2, the custom filtering mechanism supports the user to interactively select the filtering threshold, cutting size and the type of change to be identified. Specifically, the changing and unchanged category combinations are customized from the N_classes^2 category combinations composed of categories. For example, the change from cultivated land to cultivated land and from cultivated land to forest land are defined as unchanged. Then, the mask of the effective area is calculated based on the two-phase semantic segmentation label, and the Hadamard product is performed with the change detection result. The obtained accurate change detection result is converted into a vectorized format for storage. The accurate change detection result can be converted into geojson format for result display. In the process of making the effective area mask, the custom filtering mechanism can not only use the Hadamard product, but also use matrix operation methods such as Schur product, Pointwise multiplication or Element-wise multiplication as alternatives to realize the mask function.

[0058] Through this improvement, the ecological space control area development and construction activity identification system can more accurately detect and identify development and construction activities, improve detection efficiency, reduce false alarms, and support more personalized identification needs.

[0059] S3. Create an image pyramid using the satellite remote sensing images uploaded by users at two different times, dynamically load the satellite remote sensing images at two different times with the corresponding resolution according to the zoom level, and display the satellite remote sensing images at two different times and the precise change detection result layer converted into geojson format in a rolling shutter manner.

[0060] Since the satellite remote sensing images uploaded by users are relatively large, online observation and interaction of recognition results often take too long. In step S3, by establishing an image pyramid, the rendering time is reduced and a better rendering effect is achieved.

[0061] After receiving the tif format remote sensing image uploaded by the user, the corresponding .ovr format pyramid file will be generated. The design idea of ​​this image pyramid is to divide the original image into multiple resolution levels, and each level is saved as an independent image. The low-resolution image is used to zoom in on a larger map view, and the high-resolution image is used to zoom in on a smaller map view. At the same time, the remote sensing images of the previous and next phases and the change detection result layer converted into the geojson format are displayed in a rolling manner. In this way, the map software can dynamically load and display images of the corresponding resolution according to the current zoom level, thereby improving rendering efficiency while ensuring rendering quality.

[0062] This embodiment explores the prediction performance of multiple different models through two paths: the indirect method of semantic segmentation and the direct method of change detection; adopts a change detection filtering mechanism based on semantic segmentation, and uses semantic segmentation technology to customize the filtering of change detection results to meet the ability of different monitoring needs, thereby improving the accuracy of detection, reducing false alarms, and allowing personalized monitoring needs; reduces rendering time by establishing an image pyramid, thereby achieving faster rendering effects; can quickly drag the form of a rolling shutter to compare satellite images and recognition results before and after the change, thereby improving the efficiency of users in observing and interacting with recognition results; allows users to upload and update models to adapt to the continuous changes in data and the evolution of enterprise needs, providing greater model flexibility.

[0063] Example 2

[0064] This embodiment introduces an ecological space control zone construction activity recognition system based on semantic segmentation, which mainly includes an image acquisition module, a model selection module, a data set organization module, a deep learning module, a custom filtering module and a rendering module.

[0065] The image acquisition module is used to cut the overlapping satellite remote sensing images uploaded by the user into small images and temporarily store the two satellite remote sensing images uploaded by the user in tif format.

[0066] The model selection module is used to determine the semantic segmentation model and the change detection model from the model library according to the user selection or the default selection;

[0067] When the user clicks the corresponding button on the system page, a pop-up window of an upload form will be triggered. In this form, the user needs to provide the name of the model, which helps to identify and label the model; secondly, the user needs to select the category of the model so that it can be correctly classified with other models; finally, the user needs to upload their model weight file. The user completes the form and submits it, and the system will receive and process the uploaded file. The model weight file will be saved to a specific folder for use in future analysis. In addition, the backend will perform some verification and security checks to ensure that the uploaded model weight file is valid and secure. After completion, the newly uploaded model will be added to the list of optional models, enabling users to choose to use them in future analysis processes. No additional setup or configuration is required, just simply select the model they uploaded in the model selector to start using it.

[0068] The dataset organization module is used to filter out datasets of the required resolution from public image change datasets, and use them as training sets for pixel-level annotation of change detection and semantic segmentation; the annotated datasets are input into the semantic segmentation model and change detection model for training to obtain trained semantic segmentation models and change detection models.

[0069] The deep learning module is used to input the segmented images temporarily stored by the image acquisition module into the trained semantic segmentation model and change detection model, and output the semantic segmentation results and change detection results.

[0070] The custom filtering module is used to calculate the mask of the effective area of ​​the two-phase semantic segmentation labels of the semantic segmentation results, filter the change detection results, obtain accurate change detection results and store them.

[0071] The rendering module is used to build an image pyramid from the satellite remote sensing images uploaded by users at two different times, dynamically load the satellite remote sensing images at two different times with the corresponding resolution according to the zoom level, and display the satellite remote sensing images at two different times and the precise change detection result layer converted into geojson format in a rolling shutter manner.

[0072] In this system, the front-end is based on the overall architecture and data binding provided by the Vue 3 framework, and uses the interface elements and beautification provided by the Element-UI component library. With the help of the OpenLayers plug-in, the map and remote sensing image display are realized, and the back-end is efficiently interacted with by Axios to form a complete front-end system. The back-end is based on the Python language in the environment configuration of python3.6, pytorch1.6.0, and torchvision0.7.0, and deploys semantic segmentation models such as UNet and change detection models such as BIT-basedNet. It also uses geographic libraries such as gdal and geopandas and graphics libraries such as opencv and pillow to pre-process and integrate geographic and image data to form a complete change recognition system. On this basis, the python-based lightweight Web framework Flask is used for model encapsulation and data transmission, and the Flask-CORS package is used to configure Cors cross-domain to form a complete back-end system.

[0073] like Figure 2 As shown in the figure, the main process of change detection in the system is as follows:

[0074] (1) After the front end sends the upload request, the back end receives and stores the original remote sensing images of the two phases, generates the corresponding OVR tile file and returns the file address for the front end to accelerate the display.

[0075] (2) In the case of custom filtering, after the front-end uploads the two-phase semantic segmentation label map used to filter the change area and sends a category analysis request, the back-end receives and stores the two-phase semantic segmentation label map and counts the label categories therein, and returns a category list. (In the case of default filtering, the back-end directly performs analysis and calculation according to the default options)

[0076] (3) In the case of custom filtering, after the front-end sends the analysis request, the back-end receives the blocksize (size of the segmented image), modelvision (model version used for analysis), threshold (area filter threshold), filter category and other parameters selected by the front-end to start analysis and calculation:

[0077] Ⅰ) Cut the image into overlapping small images and store them temporarily;

[0078] II) Input the change detection model to analyze one by one and output the corresponding change label map (in the case of default filtering, it is necessary to input the semantic segmentation model to analyze one by one and output the corresponding two-phase semantic segmentation label map);

[0079] III) The small images are weighted by the confidence matrix to obtain a complete label image;

[0080] IV) Calculate the area where non-construction land is transformed into construction land based on the semantic segmentation label map of the two phases, and obtain the effective area mask;

[0081] V) filtering out the area within a certain range where plots of different land uses meet, and obtaining a further effective area mask;

[0082] VI) The original change labels are filtered by masks to obtain the final change labels and converted into vectorized format for storage;

[0083] Ⅶ) Return the changed area data in geojson format for front-end display;

[0084] (4) After the front end sends a download request, the back end returns the changed area data in shp vector format or tif raster format according to the selected format.

[0085] For semantic segmentation and change detection tasks, multiple data sets are selected and data enhancement such as noise addition is performed. At the same time, multiple models are selected and trained on multiple data sets in turn, and evaluation indicators such as recall rate on the same test set are compared to select the optimal model.

[0086] To give a specific example, in the semantic segmentation task, two classic models in the field of semantic segmentation, UNet and SegNet, were selected for training on the LoveDA and Guofeng datasets. The UNet model trained on the LoveDA dataset achieved better results. In the change detection task, STANet, BIT-CD, and AIE-changedet were selected for training on remote sensing datasets with different resolutions and classification labels, such as LEVIR, DSIFN, and SenseEarth. Among them, the accuracy of BIT-CD and AIE-changedet on the SenseEarth dataset reached more than 94%.

[0087] In addition, other models can also be used as alternatives to realize the functions of the system. For example, the semantic segmentation model can also choose U-Net, FCN, SegNet, PSPNet, DeepLab v1 / v2 / v3 / v3+, etc., and the change detection model can choose STANet, BIT-basedNet, AIE-changedet, so as to build a new activity recognition system. The system can also realize the effective identification of development and construction activities in the ecological space control area.

[0088] In addition, in the process of custom filtering mechanism, in the process of effective area masking, not only the Hadamard product can be used, but also matrix operation methods such as Schur product, Pointwise multiplication or Element-wise multiplication can be used as alternatives to realize the masking function.

[0089] The system development built the user interface through the front-end JavaScript framework VUE. At the same time, the OpenLayers library is used to create geospatial content on the web page, including displaying maps, geographic data, etc., and integrated in the back-end service. In addition, the system uses the Flask framework to build a suitable web application, and finally realizes the left and right functions of the technical solution. There are similar alternatives for these three parts. For example, the front-end VUE framework can be replaced by React and Angular; the Openlayers library can be replaced by Leaflet and Mapbox; the back-end Flask can also be replaced by Django. Among them, React is also a very popular front-end framework, generally used in large-scale projects and enterprise-level applications. If the system needs a larger-scale upgrade, this framework can be used. Angular is a front-end framework developed and maintained by Google, with strong community support and a rich ecosystem, which can realize many personalized design solutions. For the alternatives to OpenLayers, Leaflet and Mapbox, the former is a lightweight JavaScript library for creating interactive maps on web pages, and the latter is a commercial map service provided by DigitalGlobe, which has a wide range of applications and support. The two have relatively large differences in visualization effects. Django, an alternative to Flask, is a full-featured Python web framework that is widely used to build large web applications. It has powerful built-in functions and third-party libraries that can greatly reduce development time and complexity.

[0090] Therefore, this embodiment has the following technical effects:

[0091] 1) Model selection and comparison optimization: Explore the prediction performance of multiple different models from two paths: the indirect method of semantic segmentation (UNet model, SegNet model) and the direct method of change detection (STANet model, BIT-basedNet model, AIE-changedet model). And optimize the task characteristics under the guidance of model evaluation indicators. Finally, based on the effect comparison, select the model with excellent performance and integrate it into the final system.

[0092] 2) Semantic segmentation-based change detection filtering mechanism: The ability to use semantic segmentation technology to filter change detection results in a customized manner to meet different monitoring needs. This improves detection accuracy, reduces false positives, and allows for personalized monitoring needs.

[0093] 3) Customized monitoring task settings: Users can customize the filter threshold, cut size, and type of changes to be identified according to the needs of specific monitoring tasks. This provides users with flexibility to adapt the system to the personalized needs of different monitoring tasks.

[0094] 4) Accelerated rendering and visualization method for identification results of development and construction activities in ecological space control areas: By establishing an image pyramid to reduce rendering time, faster rendering can be achieved. The satellite images and identification results before and after the change can be quickly compared by dragging a rolling shutter, which improves the efficiency of users in observing and interacting with identification results.

[0095] 5) User-defined update mechanism for deep learning models: Users can upload and update models to adapt to the ever-changing data and the evolution of enterprise needs, providing greater model flexibility. When users click the corresponding button on the system page, a pop-up window of an upload form will be triggered. In this form, users need to provide the name of the model, which helps to identify and label the model; secondly, users need to select the category of the model so that it can be correctly classified with other models; finally, users need to upload their model weight file. The user completes the form and submits it, and the back-end system will receive and process the uploaded file. The model weight file will be saved to a specific folder for use in future analysis. In addition, the back-end will perform some verification and security checks to ensure that the uploaded model weight file is valid and secure. Once completed, the model selector on the front-end will be updated immediately. The newly uploaded model will be added to the list of optional models, enabling users to choose to use them in future analysis processes. No additional settings or configurations are required. Simply select the model they uploaded in the model selector to start using it.

[0096] 6) System integration and front-end and back-end workflow connection: With the help of front-end platforms such as Vue framework and OpenLayers and back-end technology stacks such as Flask, the web pages deployed on the enterprise bastion host can realize functions such as file transfer and interaction, object classification, change detection, model update and result export. The research and application results can effectively assist enterprises in identifying and monitoring development and construction activities in the ecological space control area.

[0097] The technical scope of the present invention is not limited to the contents in the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A method for identifying construction activities in ecological space control areas based on semantic segmentation, characterized in that: It includes the following steps: S1. Segment the satellite remote sensing images of the same place at two times uploaded by the user and input them into the pre-trained semantic segmentation model and change detection model, and output the semantic segmentation results and change detection results; S2. Customize and filter the change detection results according to the semantic segmentation results to obtain and store accurate change detection results. The specific steps of custom filtering are as follows: S21. The segmented images of the semantic segmentation results are weighted by the confidence matrix to obtain the corresponding two-phase semantic segmentation label map; S22. Calculate the area converted from non-construction land to construction land according to the two-phase semantic segmentation label map, and obtain a valid area mask; S23. Filter out the area within a certain range where plots of different land uses meet within the effective area to obtain an accurate effective area mask; S24. Filtering the change detection result through the precise effective area mask to obtain the precise change detection result and converting it into a vectorized format for storage; S3. Create an image pyramid using the satellite remote sensing images uploaded by users at two different times, dynamically load the satellite remote sensing images at two different times with the corresponding resolution according to the zoom level, and display the satellite remote sensing images at two different times and the precise change detection result layer converted into geojson format in a rolling shutter manner.

2. The method for identifying construction activities in ecological space control areas based on semantic segmentation according to claim 1 is characterized in that: In step S1, the specific steps of training the semantic segmentation model and the change detection model are as follows: Pre-screen the dataset of the required resolution from the public image change dataset and use it as the training set for pixel-level annotation of change detection and semantic segmentation; The labeled data set is input into the pre-built semantic segmentation model and change detection model for training to obtain the trained semantic segmentation model and change detection model.

3. The method for identifying construction activities in ecological space control areas based on semantic segmentation according to claim 2 is characterized in that: Multi-category semantic labels are assigned to the images in the training set for semantic segmentation, and the multi-category semantic labels include water bodies, roads, trees, and buildings; binary classification labels are assigned to the training set for change detection, and the binary classification labels consist of changes and no changes.

4. The method for identifying construction activities in ecological space control areas based on semantic segmentation according to claim 1 is characterized in that: The semantic segmentation model and change detection model are stored in the model library; Semantic segmentation models include U-Net, FCN, SegNet, PSPNet or DeepLab; change detection models include STANet, BIT-basedNet or AIE-changedet.

5. The method for identifying construction activities in ecological space control areas based on semantic segmentation according to claim 4 is characterized in that: The semantic segmentation model and change detection model are determined based on user selection. If the user does not select a model, the default model is used.

6. The method for identifying construction activities in ecological space control areas based on semantic segmentation according to claim 5 is characterized in that: The model library also includes models uploaded by users. After verification and security testing, the models uploaded by users are updated to the model library according to categories for user selection.

7. The ecological space control area construction activity recognition system based on semantic segmentation is characterized by: It includes: An image acquisition module is used to cut the overlapping satellite remote sensing images uploaded by the user into small images for temporary storage; A model selection module, which is used to determine a semantic segmentation model and a change detection model from a model library according to a user selection or a default selection; The dataset organization module is used to select datasets with the required resolution from public image change datasets and use them as training sets for pixel-level annotation of change detection and semantic segmentation. The labeled data set is input into the semantic segmentation model and the change detection model for training to obtain the trained semantic segmentation model and the change detection model; A deep learning module is used to input the segmented images temporarily stored in the image acquisition module into the trained semantic segmentation model and change detection model, and output the semantic segmentation results and change detection results; A custom filtering module is used to calculate the mask of the valid area of ​​the two-phase semantic segmentation labels of the semantic segmentation results, filter the change detection results, obtain accurate change detection results and store them; The rendering module is used to create an image pyramid from the satellite remote sensing images uploaded by users at two different times, dynamically load the satellite remote sensing images at two different times with the corresponding resolution according to the zoom level, and display the satellite remote sensing images at two different times and the precise change detection result layer converted into geojson format in a rolling shutter manner.

8. The ecological space control area construction activity identification system based on semantic segmentation according to claim 7 is characterized in that: The ecological space control zone construction activity identification system based on semantic segmentation also includes a custom update module, which is used to receive a model weight file containing a model name and category uploaded by a user, and update it to the model library after verification and security check for user selection.

9. The ecological space control area construction activity identification system based on semantic segmentation according to claim 7 is characterized in that: In the rendering module, the satellite remote sensing images of the same place at two times in the tif format uploaded by the user are used to generate corresponding pyramid files in the ovr format to establish an image pyramid.

10. The ecological space control area construction activity identification system based on semantic segmentation according to claim 7 is characterized in that: In the custom filtering module, Hadamard product, Schur product, Pointwise multiplication or Element-wise multiplication matrix operation method is used to realize the mask function.

Images