Farmland remote sensing image intelligent segmentation system and method based on dynamic feature learning

Abstract

The invention belongs to the technical field of image recognition, and discloses a farmland remote sensing image intelligent segmentation system and method based on dynamic feature learning, and the system carries out the automatic extraction of farmland features based on visible light, multispectrum, synthetic aperture radar (SAR) image data and historical segmentation results. High-precision geometric adaptive segmentation of farmland boundaries is realized through multispectral band recombination, multi-temporal feature fusion, dynamic deformable convolutional network operation and the like. And a standardized farmland distribution diagram and a standardized statistical report containing factors such as field area statistics, crop type distribution and a boundary topological relation are dynamically generated according to a farmland drawing specification and a preset agricultural thematic standard, so that a whole-process intelligent technical support is provided for businesses such as farmland resource supervision, agricultural subsidy accounting and planting structure analysis.

Description

Technical Field

[0001] The present invention belongs to the field of image recognition technology, and in particular relates to a system and method for intelligent segmentation of farmland remote sensing images based on dynamic feature learning. Background Art

[0002] With the rapid development of high-resolution remote sensing satellites and drone technology, the agricultural sector has entered an era of all-weather, multi-scale precision monitoring. Modern remote sensing platforms can acquire multispectral image data with sub-meter spatial resolution, covering more than 10 spectral channels such as visible light, red edge, and near-infrared. They can accurately capture the distribution of farmland boundaries, crop growth characteristics, and changes in soil moisture. These data contain core information such as cultivated land utilization, crop planting structure, and farmland infrastructure, and are of great strategic value in areas such as cultivated land protection and supervision, agricultural insurance loss assessment, and high-standard farmland construction. However, faced with the annual growth of terabyte-level (storage data volume of more than 1TB) agricultural remote sensing data, how to efficiently extract detailed farmland boundary information, especially to deal with the identification of fragmented fields in complex terrain conditions such as terraced fields and hilly and mountainous areas in the south, has become a technical bottleneck restricting the development of smart agriculture.

[0003] In recent years, artificial intelligence technologies, represented by deep learning, have injected new momentum into agricultural remote sensing analysis. Convolutional neural networks, by autonomously learning farmland texture, spectral, and spatiotemporal characteristics, have demonstrated significant advantages in tasks such as crop classification and field segmentation. Typical models, such as U-Net and DeepLabv3+, have achieved an overall farmland recognition accuracy exceeding 85%, significantly surpassing traditional visual interpretation methods. However, existing technologies still face significant challenges: periodic fluctuations in crop spectral characteristics in multi-temporal remote sensing data lead to reduced model generalization capabilities; confusion between the boundaries of small ridges and ditches causes fragmented segmentation results; and unstable image quality in cloudy and rainy areas leads to missegmentation. This is particularly true when dealing with large-scale farmland in the black soil of Northeast China and scattered cultivated land in the hilly areas of Southwest China. A single algorithmic model struggles to balance adaptability to large-scale scenarios with sensitivity to small-scale features.

[0004] Currently, simply achieving intelligent segmentation of farmland areas is no longer sufficient to meet the profound demands of modern, refined agricultural management. In practical applications, there is an urgent need to transform segmentation results into standardized farmland distribution maps and statistical reports that conform to agricultural management standards and include elements such as field area statistics, crop type distribution, and boundary topology relationships. Such results can provide full-process intelligent technical support for operations such as cultivated land resource supervision, agricultural subsidy calculation, and planting structure analysis. Therefore, building a complete technology chain from intelligent interpretation of remote sensing imagery to the automatic generation of agricultural thematic products has become a key breakthrough in promoting the implementation of precision agriculture. Summary of the Invention

[0005] The purpose of the present invention is to provide a system and method for intelligent segmentation of farmland remote sensing images based on dynamic feature learning. The system automatically extracts farmland features based on multi-source remote sensing data (visible light, multispectral and synthetic aperture radar SAR image data) and historical segmentation results. High-precision geometric adaptive segmentation of farmland boundaries is achieved through multispectral band recombination, multi-temporal feature fusion, dynamic deformable convolutional network calculation, etc., and standardized farmland distribution maps and standardized statistical reports containing factors such as field area statistics (accuracy ±0.05 mu), crop type distribution (GB / T 33469 coding) and boundary topological relationships are dynamically generated according to farmland mapping specifications and preset agricultural thematic standards. This provides full-process intelligent technical support for businesses such as cultivated land resource supervision, agricultural subsidy accounting, and planting structure analysis.

[0006] The technical solution adopted by the present invention is an intelligent segmentation system for farmland remote sensing images based on dynamic feature learning, including a multi-source data analysis module, a farmland template management module, a dynamic segmentation generation module, a human-computer interactive optimization module, and a thematic product generation module, wherein:

[0007] The multi-source data analysis module is used to read and analyze multispectral image data, visible light image data, SAR image data and historical segmentation results, and automatically extract farmland features;

[0008] The farmland template management module is used to comprehensively manage and apply the farmland segmentation templates stored in the system;

[0009] The dynamic segmentation generation module selects the farmland segmentation template parameters that match the current scene based on the automatically extracted farmland feature data and the preset farmland segmentation template parameters, and then dynamically generates the initial segmentation results according to the farmland mapping specifications and preset agricultural thematic standards;

[0010] The human-computer interaction optimization module is used to support users to interactively edit and modify the initial segmentation results through the coordinated display of farmland remote sensing images and segmentation boundaries and the real-time update preview of the initial segmentation results;

[0011] The thematic product generation module is used to generate standardized farmland distribution maps and standardized statistical reports.

[0012] Furthermore, the multi-source data parsing module includes a remote sensing image reading and parsing unit, a historical segmentation result reading and parsing unit, and a farmland feature automatic extraction unit; the remote sensing image reading and parsing unit reads and parses multispectral image data, visible light image data, and SAR image data to obtain the spatial resolution, spectral characteristics, and geographic coordinate information of the farmland remote sensing image; the historical segmentation result reading and parsing unit decodes the farmland vector boundary file in Shapefile / GeoJSON format to obtain field polygon coordinates, crop type labels, and area statistical information; the farmland feature automatic extraction unit automatically extracts the following key parameters based on the parsing results: spatial resolution, spectral characteristics, geographic coordinate information, morphological characteristics, temporal characteristics, and texture characteristics.

[0013] Furthermore, the farmland template management module includes a template visualization display management unit, a template parameter application unit, a template import and parsing unit, a template export adaptation unit and a template deletion and cleaning unit; the template visualization display management unit performs layer preview and rendering, name modification and category retrieval on the farmland segmentation templates stored in the system; after the user selects the preset farmland segmentation template, the template parameter application unit calls the corresponding segmentation parameters from the spatial database and synchronously updates them to the core segmentation engine; the template import and parsing unit automatically parses the standardized farmland segmentation template files pre-stored in the spatial database, is compatible with the farmland segmentation template data generated by ArcGIS / QGIS software, and stores it in the spatial database after parsing; the template export adaptation unit encapsulates the farmland segmentation template into a cross-platform format and exports it to the designated GIS analysis platform; the template deletion and cleaning unit is responsible for providing a hierarchical authority management mechanism to authorize users to perform batch deletion of redundant farmland segmentation templates.

[0014] Furthermore, the dynamic segmentation generation module includes a template parameter matching unit and a segmentation result dynamic generation unit; the template parameter matching unit intelligently associates the automatically extracted farmland feature data with preset farmland segmentation template parameters to select farmland segmentation template parameters that match the current scene; the dynamic segmentation result generation unit automatically performs the following steps based on the matched farmland segmentation template parameters in accordance with farmland mapping specifications and preset agricultural thematic standards to dynamically generate an initial segmentation result:

[0015] Multispectral band recombination;

[0016] Multi-temporal feature fusion;

[0017] Adjust the deformation parameters of the deformable convolution kernel to complete the dynamic deformable convolution network operation;

[0018] Generate initial segmentation results and output farmland vector boundaries.

[0019] Furthermore, the human-computer interaction optimization module includes an image and segmentation boundary collaborative display unit, a segmentation result real-time update preview unit and a farmland boundary interactive editing unit; the image and segmentation boundary collaborative display unit adopts WebGL technology to realize the slice-free front-end rendering of farmland remote sensing images, and the vector coordinates of the farmland vector boundary output based on the initial segmentation result are superimposed in real time to display the farmland boundary polygon, supporting users to click on the field to view attribute information; the segmentation result real-time update preview unit realizes instant visualization after segmentation parameter adjustment through a dynamic rendering engine, and synchronously displays the comparison view before and after segmentation boundary optimization; the farmland boundary interactive editing unit provides the following functions: vector point drag correction, boundary topology relationship editing and AI-assisted completion.

[0020] Furthermore, the thematic product generation module includes a Shapefile format farmland distribution map generation and output unit, a GeoJSON format farmland distribution map generation and output unit, and a standardized statistical report generation and output unit; the Shapefile format farmland distribution map generation and output unit encapsulates the final segmentation result and outputs it in a Shapefile format compatible with the ESRI system; the GeoJSON format farmland distribution map generation and output unit converts the farmland vector boundary into a GeoJSON format compatible with the WebGIS system and outputs it; the standardized statistical report generation and output unit is used to generate a statistical report containing field area, perimeter and crop type, and supports export in Excel, CSV and PDF formats.

[0021] The intelligent segmentation method of farmland remote sensing images based on dynamic feature learning is implemented according to the above-mentioned intelligent segmentation system of farmland remote sensing images based on dynamic feature learning, and includes the following steps:

[0022] S1, the multi-source data analysis module reads and analyzes multispectral image data, visible light image data, SAR image data and historical segmentation results, and automatically extracts farmland features;

[0023] S2, the farmland template management module comprehensively manages and applies the farmland segmentation templates stored in the system;

[0024] The specific application of farmland segmentation template is as follows: after the user selects a preset farmland segmentation template, the segmentation parameters of the template are called from the spatial database and loaded into the core segmentation engine, and automatically set as the system default farmland segmentation template; if no selection operation is performed, the most recently used default template is used;

[0025] S3, the segmentation dynamic generation module selects the farmland segmentation template parameters that match the current scenario based on the automatically extracted farmland feature data and the preset farmland segmentation template parameters, and then dynamically generates the initial segmentation results according to the farmland mapping specifications and the preset agricultural thematic standards;

[0026] S4, the human-computer interaction optimization module supports users to interactively edit and modify the initial segmentation results through the coordinated display of farmland remote sensing images and segmentation boundaries, as well as the real-time update preview of the initial segmentation results, and finally generates the final segmentation results;

[0027] S5, the thematic product generation module encapsulates the final segmentation results into standardized farmland distribution maps and standardized statistical reports, and stores them in the specified spatial database path.

[0028] Furthermore, the specific steps of S1 include:

[0029] S11, set the multi-source data input path, read and parse the multispectral image data, visible light image data, SAR image data and historical segmentation result files in GeoJSON / Shapefile format;

[0030] S12, based on the analysis results of S11, automatically extracts the following key parameters: spatial resolution, spectral characteristics, geographic coordinate information, morphological characteristics, temporal characteristics and texture characteristics.

[0031] Furthermore, the specific steps of S4 include:

[0032] S41 uses WebGL technology to implement slice-free front-end rendering of farmland remote sensing images. The vector coordinates of the farmland vector boundary output based on the initial segmentation results are overlaid in real time to display the farmland boundary polygon. When the user clicks on a field area, a property panel pops up to display the field area statistics, crop type distribution, boundary topology relationship and segmentation confidence.

[0033] S42, through the dynamic rendering engine to achieve real-time visualization of segmentation parameter adjustments, and simultaneously display the comparison view before and after the segmentation boundary optimization;

[0034] S43: The initial segmentation results are decomposed into independent editable elements according to the field units. Each field supports vector point dragging and correction, boundary topology editing and AI-assisted completion.

[0035] S44, finally generates the final segmentation result and outputs the farmland vector boundary.

[0036] The beneficial effects of the present invention are:

[0037] (1) The present invention automatically extracts farmland features based on multi-source remote sensing data (visible light, multispectral and synthetic aperture radar SAR image data) and historical segmentation results. It achieves high-precision geometric adaptive segmentation of farmland boundaries through multispectral band recombination, multi-temporal feature fusion, and dynamic deformable convolutional network operations. It also dynamically generates standardized farmland distribution maps and standardized statistical reports containing factors such as field area statistics (accuracy ±0.05 mu), crop type distribution (GB / T 33469 coding), and boundary topological relationships based on farmland mapping specifications and preset agricultural thematic standards. This provides full-process intelligent technical support for businesses such as cultivated land resource supervision, agricultural subsidy accounting, and planting structure analysis.

[0038] (2) The present invention effectively improves the positioning accuracy, temporal adaptability, robustness in complex scenarios, processing efficiency and automation level of farmland segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a diagram showing the composition of the farmland remote sensing image intelligent segmentation system based on dynamic feature learning in the present invention.

[0040] Figure 2 This is a flow chart of the intelligent segmentation method of farmland remote sensing images based on dynamic feature learning of the present invention. DETAILED DESCRIPTION

[0041] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings.

[0042] Intelligent segmentation system of farmland remote sensing images based on dynamic feature learning, such as Figure 1 As shown, it includes multi-source data analysis module, farmland template management module, segmentation dynamic generation module, human-computer interactive optimization module and thematic product generation module.

[0043] 1. Multi-source data analysis module

[0044] The multi-source data analysis module is used to read and analyze multispectral image data, visible light image data, SAR image data and historical segmentation results, and complete automatic extraction of farmland features.

[0045] The multi-source data parsing module includes a remote sensing image reading and parsing unit, a historical segmentation result reading and parsing unit, and a farmland feature automatic extraction unit.

[0046] 1. Remote sensing image reading and analysis unit

[0047] Read and analyze multispectral image data, visible light image data, and SAR image data to obtain the spatial resolution, spectral characteristics, and geographic coordinate information of farmland remote sensing images.

[0048] 2. Historical segmentation result reading and parsing unit

[0049] Decode (read and parse) farmland vector boundary files in Shapefile / GeoJSON format to obtain field polygon coordinates, crop type labels, and area statistics.

[0050] 3. Automatic extraction unit of farmland features

[0051] Based on the above analysis results, the following key parameters are automatically extracted: spatial resolution, spectral characteristics, geographic coordinate information, morphological characteristics, temporal characteristics and texture characteristics.

[0052] Morphological characteristics: field boundary curvature (>0.35), area dispersion (coefficient of variation CV≤15%).

[0053] Temporal characteristics: NDVI growth curve (including key nodes of sowing period / heading period / maturity period).

[0054] Texture features: contrast (>120) and homogeneity (<0.3) calculated based on the Gray Level Coocurrence Matrix (GLCM).

[0055] 2. Farmland Template Management Module

[0056] The farmland template management module is used to comprehensively manage and apply the farmland segmentation templates stored in the system.

[0057] The farmland template management module includes a template visualization display management unit, a template parameter application unit, a template import and parsing unit, a template export adaptation unit, and a template deletion and cleaning unit.

[0058] 1. Template visual display management unit

[0059] The farmland segmentation templates stored in the system can be previewed and rendered in layers (with adjustable transparency range of 30%-80%), renamed (supporting Chinese and English naming), and searched by category (supporting queries by resolution / crop type / region number).

[0060] The farmland segmentation template contains the following structured parameter definitions: parameter name, parameter type, constraints, default parameter values, and optimization rules.

[0061] Optimization rules: Prioritize segmentation according to regular shapes such as rectangles and squares, and reduce irregular corner shapes such as triangles and polygons.

[0062] 2. Template parameter application unit

[0063] After the user selects the preset farmland segmentation template (the farmland segmentation template stored in the system), the corresponding segmentation parameters (morphological constraints (area threshold ≥ 30m 2 ), spectral feature range (NDVI 0.4-0.8) and edge optimization intensity (correction amplitude ≤ 0.5 pixels), and are synchronously updated to the core segmentation engine.

[0064] Core segmentation engine: The background processing component in the dynamic segmentation generation module, whose core function is to cluster pixels with similar features in farmland remote sensing images into meaningful objects.

[0065] 3. Template import parsing unit

[0066] The standardized farmland segmentation template files (TIFF / GeoJSON format) stored in the spatial database were automatically parsed and compatible with the farmland segmentation template data generated by ArcGIS / QGIS software (spatial reference system error ≤ 0.1 pixel). After parsing, they were stored in the spatial database (PostGIS version 3.2).

[0067] 4. Template export adaptation unit

[0068] The farmland segmentation template was encapsulated into a cross-platform format (TIFF+XML metadata / GeoJSON format) and exported to the designated GIS analysis platform (ArcGIS / QGIS software).

[0069] 5. Template deletion and cleaning unit

[0070] Responsible for providing a hierarchical authority management mechanism (administrator / operator dual-level authority), authorized users can perform batch deletion of redundant farmland segmentation templates (synchronously clearing associated cache files and log records).

[0071] 3. Split Dynamic Generation Module

[0072] The dynamic segmentation generation module selects the farmland segmentation template parameters that match the current scene based on the automatically extracted farmland feature data and the preset farmland segmentation template parameters, and then dynamically generates the initial segmentation results according to the farmland mapping specifications and preset agricultural thematic standards.

[0073] The segmentation dynamic generation module includes a template parameter matching unit and a segmentation result dynamic generation unit.

[0074] 1. Template parameter matching unit

[0075] The automatically extracted farmland feature data is intelligently associated with the preset farmland segmentation template parameters, and the farmland segmentation template parameters that match the current scene (farmland remote sensing image to be segmented) are selected (such as the spectral feature effective range 0.4-0.8, the minimum field area ≥30m 2 ).

[0076] 2. Segmentation result dynamic generation unit

[0077] Based on the matched farmland segmentation template parameters, the following steps are automatically performed according to the Farmland Mapping Specification (GB / T 33469-2016) and preset agricultural thematic standards to dynamically generate the initial segmentation results:

[0078] Multispectral band reorganization (preferably red edge / near-infrared bands); multi-temporal feature fusion (analysis of NDWI change rate from sowing period to maturity period); adjustment of deformable convolution kernel deformation parameters to complete dynamic deformable convolution network operation (convolution kernel offset ≤ 3×3 pixels); generation of initial segmentation results, and output of farmland vector boundaries (topological tolerance ≤ 0.1 pixel).

[0079] For example:

[0080] Based on the automatically extracted farmland features, if valid farmland targets are detected, the following farmland mapping specifications and preset agricultural thematic standards are used to dynamically generate segmentation results:

[0081] This time, (satellite name) detected (total number of farmlands N) farmland targets in the observation area:

[0082] The target of farmland No. 1 is (name of target farmland No. 1), the area is (target area of ​​farmland No. 1) mu, the crop type is (target crop type of farmland No. 1), and the growth stage is (target growth stage of farmland No. 1);

[0083] The target of farmland No. 2 is (target name of farmland No. 2), the area is (target area of ​​farmland No. 2) mu, the crop type is (target crop type of farmland No. 2), and the growth stage is (target growth stage of farmland No. 2); ......

[0085] The target farmland (N) is (target farmland name), the area is (target farmland area) mu, the crop type is (target crop type of farmland N), and the growth stage is (target growth stage of farmland N);

[0086] If no valid farmland target is detected, the segmentation result content is the default text set in the farmland mapping specification template:

[0087] No effective farmland information was detected in the observation area this time.

[0088] For other elements (non-farmland features), fill the element information values ​​into the designated locations according to the farmland mapping specification style:

[0089] Feature name: feature information value;

[0090] For example:

[0091] Shooting time: 2024-11-29 13:40:30;

[0092] Satellite code: GF-6;

[0093] Number of imaging bands: 10;

[0094] In summary, the segmentation results are generated dynamically.

[0095] 4. Human-computer interactive optimization module

[0096] The human-computer interaction optimization module is used to support users to interactively edit and modify the initial segmentation results through the coordinated display of farmland remote sensing images and segmentation boundaries and the real-time update preview of the initial segmentation results.

[0097] The human-computer interaction optimization module includes an image and segmentation boundary collaborative display unit, a segmentation result real-time update preview unit, and a farmland boundary interactive editing unit.

[0098] 1. Image and segmentation boundary collaborative display unit

[0099] WebGL technology is used to realize the slice-free front-end rendering of farmland remote sensing images (resolution ≥ 0.5m). The vector coordinates (WGS84 coordinate system) of the farmland vector boundary output based on the initial segmentation results are superimposed and displayed in real time on the farmland boundary polygons. Users can click on the field to view attribute information (field area statistics (accuracy ±0.05 mu), crop type distribution (GB / T33469 encoding), boundary topology relationship, and segmentation confidence (>0.85)).

[0100] WebGL technology: a JavaScript API technology that enables interactive graphics rendering by unleashing the parallel computing power of the GPU.

[0101] 2. Real-time update preview unit of segmentation results

[0102] The dynamic rendering engine enables instant visualization of segmentation parameter adjustments (response delay < 200ms), and simultaneously displays a comparison view before and after segmentation boundary optimization (with an adjustable transparency range of 30%-80%).

[0103] Dynamic rendering engine: The system monitors user operations, obtains the latest segmentation parameters, and passes them to the segmentation algorithm in the dynamic segmentation generation module, driving it to recalculate the segmentation results. It also uses WebGL technology to accelerate pixel shading and graphics rendering to ensure the immediacy of display.

[0104] 3. Interactive editing unit of farmland boundaries

[0105] The following functions are provided: vector point drag correction (supports sub-pixel displacement, accuracy of 0.1 pixel), boundary topology relationship editing (automatically maintains the consistency of adjacent field boundaries) and AI-assisted completion (intelligently repairs missing boundaries based on historical segmentation results, with a completion accuracy rate of >92%).

[0106] 5. Special product generation module

[0107] The thematic product generation module is used to generate standardized farmland distribution maps and standardized statistical reports.

[0108] The thematic product generation module includes a Shapefile format farmland distribution map generation and output unit, a GeoJSON format farmland distribution map generation and output unit, and a standardized statistical report generation and output unit.

[0109] 1. Generate output unit of farmland distribution map in Shapefile format

[0110] The final segmentation result is packaged and output in the Shapefile format compatible with the ESRI system.

[0111] The farmland distribution map in Shapefile format contains field polygon vector boundaries (point / line / area features), crop type attribute fields (coding rules GB / T 33469-2016) and metadata files (.shp / .shx / .dbf / .prj).

[0112] 2. Generate output unit of farmland distribution map in GeoJSON format

[0113] The farmland vector boundaries are converted into GeoJSON format output (coordinate system WGS84) compatible with the WebGIS system, supporting real-time rendering on platforms such as Leaflet / OpenLayers (loading delay <1s).

[0114] 3. Standardized statistical report generation output unit

[0115] Generates statistical reports including field area (unit: mu), perimeter, crop type and other indicators, and supports multiple export formats such as Excel (*.xlsx, including pivot tables), CSV (UTF-8 encoding) and PDF (including visual bar charts / heat maps).

[0116] The intelligent segmentation method of farmland remote sensing images based on dynamic feature learning is realized by the above-mentioned intelligent segmentation system of farmland remote sensing images based on dynamic feature learning, such as Figure 2 As shown, the specific steps include:

[0117] S1, the multi-source data parsing module reads and parses multispectral image data, visible light image data, SAR image data, and historical segmentation results, and automatically extracts farmland features. The specific steps include:

[0118] S11, set up the multi-source data input path, read and parse multispectral image data (including red edge / near infrared bands), visible light image data, SAR image data (HH / HV polarization mode), and historical segmentation result files (farmland vector boundary files) in GeoJSON / Shapefile format;

[0119] S12, based on the analysis results of S11, automatically extracts the following key parameters: spatial resolution, spectral characteristics, geographic coordinate information, morphological characteristics, temporal characteristics and texture characteristics.

[0120] S2, the farmland template management module comprehensively manages and applies the farmland segmentation templates stored in the system;

[0121] The specific application of the farmland segmentation template is as follows: after the user selects the preset farmland segmentation template, the segmentation parameters of the template are called from the spatial database and loaded into the core segmentation engine, and it is automatically set as the system default farmland segmentation template; if the selection operation is not performed, the most recently used default template will be used.

[0122] S3, the segmentation dynamic generation module selects the farmland segmentation template parameters that match the current scene based on the automatically extracted farmland feature data and the preset farmland segmentation template parameters, and then dynamically generates the initial segmentation results according to the farmland mapping specifications and preset agricultural thematic standards.

[0123] S4, the human-computer interaction optimization module supports users to interactively edit and modify the initial segmentation results through the coordinated display of farmland remote sensing images and segmentation boundaries, as well as the real-time update preview of the initial segmentation results, and finally generates the final segmentation results. The specific steps include:

[0124] S41 uses WebGL technology to implement slice-free front-end rendering of farmland remote sensing images. The vector coordinates of the farmland vector boundary output based on the initial segmentation results are overlaid in real time to display the farmland boundary polygon. When the user clicks on a field area, a property panel pops up to display the field area statistics, crop type distribution, boundary topology relationship and segmentation confidence.

[0125] S42, through the dynamic rendering engine to achieve real-time visualization of segmentation parameter adjustments, and simultaneously display the comparison view before and after the segmentation boundary optimization;

[0126] S43: The initial segmentation results are decomposed into independent editable elements according to the field units. Each field supports interactive operations such as dragging and dropping vector points, editing boundary topology relationships, and AI-assisted completion.

[0127] S44, finally generates the final segmentation result and outputs the farmland vector boundary.

[0128] S5, the thematic product generation module encapsulates the final segmentation results into standardized farmland distribution maps and standardized statistical reports, and stores them in the specified spatial database path (PostGIS 3.2 version).

[0129] The contents not described in detail in the specification of the present invention belong to the existing common technologies in this technical field.

Claims

1. Intelligent segmentation system of farmland remote sensing images based on dynamic feature learning, characterized by: It includes multi-source data analysis module, farmland template management module, segmentation dynamic generation module, human-computer interactive optimization module and thematic product generation module, among which: The multi-source data analysis module is used to read and analyze multispectral image data, visible light image data, SAR image data and historical segmentation results, and automatically extract farmland features; The farmland template management module is used to comprehensively manage and apply the farmland segmentation templates stored in the system; The dynamic segmentation generation module selects the farmland segmentation template parameters that match the current scene based on the automatically extracted farmland feature data and the preset farmland segmentation template parameters, and then dynamically generates the initial segmentation results according to the farmland mapping specifications and preset agricultural thematic standards; The human-computer interaction optimization module is used to support users to interactively edit and modify the initial segmentation results through the coordinated display of farmland remote sensing images and segmentation boundaries and the real-time update preview of the initial segmentation results; The thematic product generation module is used to generate standardized farmland distribution maps and standardized statistical reports.

2. The farmland remote sensing image intelligent segmentation system based on dynamic feature learning according to claim 1 is characterized in that: The multi-source data parsing module includes a remote sensing image reading and parsing unit, a historical segmentation result reading and parsing unit, and a farmland feature automatic extraction unit; the remote sensing image reading and parsing unit reads and parses multispectral image data, visible light image data, and SAR image data to obtain the spatial resolution, spectral characteristics, and geographic coordinate information of the farmland remote sensing image; the historical segmentation result reading and parsing unit decodes the farmland vector boundary file in Shapefile / GeoJSON format to obtain field polygon coordinates, crop type labels, and area statistical information; the farmland feature automatic extraction unit automatically extracts the following key parameters based on the parsing results: spatial resolution, spectral characteristics, geographic coordinate information, morphological characteristics, temporal characteristics, and texture characteristics.

3. The farmland remote sensing image intelligent segmentation system based on dynamic feature learning according to claim 1 is characterized in that: The farmland template management module includes a template visualization display management unit, a template parameter application unit, a template import and analysis unit, a template export adaptation unit, and a template deletion and cleaning unit; the template visualization display management unit performs layer preview and rendering, name modification, and category retrieval on the farmland segmentation templates stored in the system; after the user selects a preset farmland segmentation template, the template parameter application unit calls the corresponding segmentation parameters from the spatial database and synchronously updates them to the core segmentation engine; the template import and analysis unit automatically parses the standardized farmland segmentation template files pre-stored in the spatial database, is compatible with the farmland segmentation template data generated by ArcGIS / QGIS software, and stores them in the spatial database after parsing; the template export adaptation unit encapsulates the farmland segmentation template into a cross-platform format and exports it to a designated GIS analysis platform; The template deletion and cleaning unit is responsible for providing a hierarchical authority management mechanism to authorize users to perform batch deletion of redundant farmland segmentation templates.

4. The farmland remote sensing image intelligent segmentation system based on dynamic feature learning according to claim 1 is characterized in that: The dynamic segmentation generation module includes a template parameter matching unit and a segmentation result dynamic generation unit. The template parameter matching unit intelligently associates the automatically extracted farmland feature data with the preset farmland segmentation template parameters and selects the farmland segmentation template parameters that match the current scenario. The dynamic segmentation result generation unit automatically performs the following steps based on the matched farmland segmentation template parameters in accordance with farmland mapping specifications and preset agricultural thematic standards to dynamically generate the initial segmentation results: Multispectral band recombination; Multi-temporal feature fusion; Adjust the deformation parameters of the deformable convolution kernel to complete the dynamic deformable convolution network operation; Generate initial segmentation results and output farmland vector boundaries.

5. The farmland remote sensing image intelligent segmentation system based on dynamic feature learning according to claim 1 is characterized in that: The human-computer interaction optimization module includes an image and segmentation boundary collaborative display unit, a segmentation result real-time update preview unit, and a farmland boundary interactive editing unit; the image and segmentation boundary collaborative display unit uses WebGL technology to realize slice-free front-end rendering of farmland remote sensing images, and the vector coordinates of the farmland vector boundary output based on the initial segmentation result are superimposed in real time to display the farmland boundary polygon, allowing users to click on the field to view attribute information; the segmentation result real-time update preview unit uses a dynamic rendering engine to realize instant visualization after segmentation parameter adjustment, and synchronously displays the comparison view before and after segmentation boundary optimization; the farmland boundary interactive editing unit provides the following functions: vector point drag correction, boundary topology relationship editing, and AI-assisted completion.

6. The farmland remote sensing image intelligent segmentation system based on dynamic feature learning according to claim 1 is characterized in that: The thematic product generation module includes a Shapefile format farmland distribution map generation and output unit, a GeoJSON format farmland distribution map generation and output unit, and a standardized statistical report generation and output unit; the Shapefile format farmland distribution map generation and output unit encapsulates the final segmentation result and outputs it in a Shapefile format compatible with the ESRI system; the GeoJSON format farmland distribution map generation and output unit converts the farmland vector boundary into a GeoJSON format compatible with the WebGIS system and outputs it; the standardized statistical report generation and output unit is used to generate a statistical report containing field area, perimeter and crop type, and supports export in Excel, CSV and PDF formats.

7. A method for intelligent segmentation of farmland remote sensing images based on dynamic feature learning, implemented according to the intelligent segmentation system for farmland remote sensing images based on dynamic feature learning according to any one of claims 1 to 6, characterized in that: The following steps are involved: S1, the multi-source data analysis module reads and analyzes multispectral image data, visible light image data, SAR image data and historical segmentation results, and automatically extracts farmland features; S2, the farmland template management module comprehensively manages and applies the farmland segmentation templates stored in the system; The specific application of farmland segmentation template is as follows: after the user selects the preset farmland segmentation template, the segmentation parameters of the template are called from the spatial database and loaded into the core segmentation engine, and automatically set as the system default farmland segmentation template; If no selected action is performed, the most recently used default template will be used; S3, the segmentation dynamic generation module selects the farmland segmentation template parameters that match the current scenario based on the automatically extracted farmland feature data and the preset farmland segmentation template parameters, and then dynamically generates the initial segmentation results according to the farmland mapping specifications and the preset agricultural thematic standards; S4, the human-computer interaction optimization module supports users to interactively edit and modify the initial segmentation results through the coordinated display of farmland remote sensing images and segmentation boundaries, as well as the real-time update preview of the initial segmentation results, and finally generates the final segmentation results; S5, the thematic product generation module encapsulates the final segmentation results into standardized farmland distribution maps and standardized statistical reports, and stores them in the specified spatial database path.

8. The method for intelligent segmentation of farmland remote sensing images based on dynamic feature learning according to claim 7, characterized in that: The specific steps of S1 include: S11, set the multi-source data input path, read and parse the multispectral image data, visible light image data, SAR image data and historical segmentation result files in GeoJSON / Shapefile format; S12, based on the analysis results of S11, automatically extracts the following key parameters: spatial resolution, spectral characteristics, geographic coordinate information, morphological characteristics, temporal characteristics and texture characteristics.

9. The method for intelligent segmentation of farmland remote sensing images based on dynamic feature learning according to claim 7, characterized in that: The specific steps of S4 include: S41 uses WebGL technology to implement slice-free front-end rendering of farmland remote sensing images. The vector coordinates of the farmland vector boundary output based on the initial segmentation results are overlaid in real time to display the farmland boundary polygon. When the user clicks on a field area, a property panel pops up to display the field area statistics, crop type distribution, boundary topology relationship and segmentation confidence. S42, through the dynamic rendering engine to achieve real-time visualization of segmentation parameter adjustments, and simultaneously display the comparison view before and after the segmentation boundary optimization; S43: The initial segmentation results are decomposed into independent editable elements according to the field units. Each field supports vector point dragging and correction, boundary topology editing and AI-assisted completion. S44, finally generates the final segmentation result and outputs the farmland vector boundary.

Images