Remote sensing image ground feature full-factor interpretation method, system and application fusing subarea control and hierarchical extraction

Abstract

This invention provides a method, system, and application for interpreting all features of remote sensing images by integrating zoning control and hierarchical extraction. By innovatively introducing a zoning control network constructed from real geographic elements such as roads and waterways, the large-scale image interpretation task is decomposed into multiple independent geographic units, effectively avoiding the problem of extraction results contradicting geographical common sense. Simultaneously, this invention employs a hierarchical extraction mechanism, extracting features class by class within each zoning according to visual saliency. Each dedicated model only needs to learn to recognize a single or a few categories, fundamentally improving inter-class discrimination and overall extraction accuracy. Furthermore, by encapsulating all stages—data preprocessing, zoning generation, hierarchical extraction, and result fusion—into visual, programmable tool nodes, and uniformly scheduling and executing them through a workflow engine, this completely changes the shortcomings of existing technologies, such as fragmented processes and reliance on manual coordination, significantly improving the standardization of production and engineering application capabilities.

Description

Technical Field

[0001] This invention relates to the field of observation technology, and in particular to a method, system, and application for interpreting all elements of ground features in remote sensing images by integrating zoning control and hierarchical extraction. Background Technology

[0002] With the rapid development of Earth observation technology, high spatial resolution remote sensing image data has exploded in volume and has been widely applied in numerous fields such as land surveys, urban planning, environmental monitoring, and national defense. Traditional manual visual interpretation methods, due to their low efficiency and high subjectivity, can no longer meet the processing demands of massive amounts of data. Therefore, utilizing computer vision and artificial intelligence technologies to achieve automated and intelligent interpretation of remote sensing images, especially the "full-element extraction" task of simultaneously and accurately segmenting multiple land cover categories such as buildings, roads, water bodies, and farmland from a single image, has become a core direction of geographic information science research and an urgent need for industrial applications.

[0003] Currently, semantic segmentation technology, represented by deep learning, has become the mainstream method for solving remote sensing image interpretation problems. Existing technical solutions mainly fall into two categories: one is an end-to-end full-feature extraction scheme based on a single complex network. This type of scheme uses deep encoder-decoder networks such as U-Net and DeepLabv3+, aiming to allow a single model to learn the complex mapping relationship from the image to all preset land cover categories. However, in pursuit of high accuracy, these models are often designed to be extremely large and complex, resulting in heavy computational load and slow inference speed when processing large-scale, high-resolution images, making it difficult to meet timeliness requirements. Furthermore, a single model needs to learn the discrimination boundaries of dozens of land cover types simultaneously. When training samples are limited, this can easily lead to confusion between different categories with similar spectral and textural features in the feature space, limiting the overall accuracy improvement. Moreover, the extraction results lack macro-geospatial logic, often exhibiting phenomena such as building patches crossing roads, which violate geographical common sense. The other category is an extraction scheme based on multi-model ensemble or cascade. This scheme recognizes the limitations of a single model and attempts to distribute tasks by cascading multiple models, for example, performing coarse classification followed by finer subdivision. However, in such solutions, the models are usually loosely coupled, and the process from data preprocessing and model calling to result fusion heavily relies on manual scripts, resulting in a cumbersome process with poor repeatability and low automation, making it difficult to form a large-scale engineering application capability.

[0004] Therefore, there is an urgent need in this field for a method for interpreting all elements of remote sensing images that can systematically solve the above problems. It is hoped that this method can improve processing efficiency, ensure extraction accuracy, ensure that the results conform to geospatial logic, and achieve full-process automation and systematic integration from data to results. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and application for interpreting all features of remote sensing images by integrating zoning control and hierarchical extraction, so as to solve the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following solution: This invention provides a method for interpreting all features of remote sensing images by integrating zoning control and hierarchical extraction, comprising the following steps: S1. Constructing a regional control network: Based on the input remote sensing image, extract the linear elements that constitute the geospatial framework, and generate a natural geographic regional network composed of closed polygons through topology repair and fusion. S2. Hierarchical intelligent extraction: For each geographic partition, according to the preset land cover category order, the dedicated semantic segmentation model corresponding to the category is called and executed in sequence. After the extraction of each category is completed, the classified mask of the partition is updated so that the extraction of subsequent categories is only performed on the pixel area not covered by the current mask. S3. Results Generation and Output: Spatial fusion and stitching of the feature patches extracted from all geographic zones to generate complete vector interpretation results covering the entire region.

[0007] Preferably, step S1 further includes: S11. Use a pre-trained lightweight model to extract initial linear features of major roads and waterways from remote sensing images; S12. Vectorize the initial linear elements and perform breakpoint connection and topology repair based on graph theory algorithms to form a continuous road network and water system network vector layer; S13. The road network and water system network vector layers are spatially superimposed, and multiple minimum closed polygons are automatically divided according to their intersection relationship, with each polygon serving as an independent geographical partition.

[0008] Preferably, step S1 further includes: when the input contains digital elevation model data, extracting ridge lines and valley lines as terrain feature lines, and spatially superimposing the terrain feature lines with the road network and water system network vector layers to construct the zonal control network.

[0009] Preferably, in step S2, the preset order of land cover categories is arranged according to the visual salience of the land cover in the image from strong to weak, or from easy to difficult.

[0010] Preferably, in step S2, the dedicated semantic segmentation model is stored in a model pool, and each model is a relatively simplified deep learning model that is optimized and trained for a single or a small number of specific land cover categories.

[0011] This invention also provides a remote sensing image land cover full-element interpretation system that integrates zoning control and hierarchical extraction to implement the above-mentioned methods, comprising: The zonal control network generation module is used to extract linear geographic features from input remote sensing images, perform topology repair and network fusion, and generate a zonal control network composed of multiple geographic zones. The hierarchical intelligent extraction module has a built-in model pool containing multiple dedicated semantic segmentation models, as well as a hierarchical extraction executor. The hierarchical extraction executor is used to sequentially load the models in the model pool for prediction within each geographic partition, and to isolate classified areas through mask update technology, guiding subsequent models to process only unclassified areas. The process engine and production line management module are used to encapsulate the data input, partition network generation, hierarchical extraction, and result integration into visual and programmable tool nodes, and schedule and execute the processing flow formed by the user, ultimately outputting the interpretation results.

[0012] Preferably, the process engine and production line management module include: A graphical process designer provides a user interface that allows users to build custom interpretation processes by dragging and dropping and connecting pre-packaged tool nodes; The process interpreter and scheduler parses the user-designed flowchart in the background and strictly calls the instantiated tool nodes to execute tasks in sequence according to the data flow. The task status monitor is used to display the process execution progress, current running nodes, and log information in real time.

[0013] Preferably, the tool nodes include an image reading node, a partition network generation node, a model prediction node corresponding to at least one land cover category, a vector post-processing node, and a result mosaicking and summarizing node, each node having a standardized input and output interface.

[0014] This invention also provides an application of a remote sensing image feature interpretation system that integrates zoning control and hierarchical extraction in change detection. Based on the aforementioned process engine and production line management module, a change detection production line is orchestrated, the production line comprising: The first full-element extraction sub-process is used to process the remote sensing image of the first time phase and output the first vector result; The second full-element extraction sub-process, set in parallel or serial with the first full-element extraction sub-process, is used to process remote sensing images of the second time phase and output the second vector result. The change detection tool has its input end connected to the output end of the first and second full-feature extraction sub-processes. It is used to perform spatial overlay analysis on the first and second vector results to identify and extract changed land feature patches.

[0015] Preferably, the change detection production line further includes a change attribute association tool, which is used to automatically assign semantic attributes indicating the change of land cover type based on the spatial location and category of the changed land cover patch.

[0016] The present invention achieves the following beneficial technical effects compared to the prior art: This invention provides a method, system, and application for interpreting all features of remote sensing images by integrating zoning control and hierarchical extraction. It features high processing efficiency, good interpretation accuracy, strong logical consistency of results, and a high degree of automation throughout the entire process. Specifically, by innovatively introducing a zoning control network constructed from real geographic elements such as roads and waterways, the large-scale image interpretation task is decomposed into multiple independent geographic units. This not only achieves parallel processing of the task and significantly improves computational efficiency, but more importantly, it imposes explicit spatial constraints on the subsequent extraction process, ensuring that the final image patches naturally conform to the geographic patterns of the real world, effectively avoiding the problem of extraction results violating geographical common sense. Simultaneously, this invention employs a hierarchical extraction mechanism of "dedicated models + masking technology," extracting features class by class within each zoning according to visual saliency. Each dedicated model only needs to learn to recognize a single or a few categories, greatly reducing the risk of feature confusion and fundamentally improving inter-class discrimination and overall extraction accuracy. Furthermore, by encapsulating all aspects such as data preprocessing, partition generation, hierarchical extraction, and result fusion into visual and programmable tool nodes, and having them uniformly scheduled and executed by the process engine, this invention constructs a "one-click" automated intelligent interpretation production line from raw data input to final vector result output. This completely changes the shortcomings of existing technology processes being fragmented and reliant on manual coordination, and significantly improves the standardization of production and engineering application capabilities. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The flowchart illustrates the method for interpreting all features of remote sensing images by fusing zoning control and hierarchical extraction, as provided in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The purpose of this invention is to provide a method, system, and application for interpreting all features of remote sensing images by integrating zoning control and hierarchical extraction. The core concept of this invention is to provide a brand-new systematic technical framework of "zoning control, hierarchical extraction, and process-driven" to solve the problems of low efficiency, poor accuracy, and insufficient automation in existing remote sensing image interpretation technologies.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1: Figure 1 A flowchart of a remote sensing image full-feature interpretation method integrating zoning control and hierarchical extraction is shown. This method mainly includes three core steps: zoning control network construction, hierarchical intelligent extraction, and result generation and output. In a specific embodiment, this method can be implemented based on a computer system containing the necessary processor, memory, and user interface for performing the following steps.

[0023] First, in step S1, a zonal control network is constructed. The goal of this step is to divide the large-scale, unstructured raw imagery into a series of relatively independent operational units with clear geographic boundaries, based on the spatial relationships between real geographic entities, thus providing spatial constraints for subsequent extraction. This step is crucial for achieving "zonal control."

[0024] Specifically, in a preferred embodiment, step S1 may further include the following sub-steps. Sub-step S11, initial extraction of linear features. Using a pre-trained lightweight semantic segmentation model or an optimized edge detection algorithm, the initial binary raster image of the main roads and main waterways is quickly extracted from the input high-resolution remote sensing image. The lightweight model here aims to pursue processing speed while ensuring a certain level of accuracy. For example, MobileNet can be used as a simplified version of U-Net as the backbone network. Sub-step S12, vectorization and topology repair. The above raster results are vectorized to generate preliminary linear vector data. Subsequently, a topology repair tool is invoked. This tool, based on graph theory algorithms, automatically identifies the breakpoints of linear features and intelligently connects them within a preset tolerance range to form a continuous road network and waterway network. To further improve the integrity of the data, the tool can also integrate publicly available or third-party open-source road and waterway data to supplement road segments or waterways missing due to image quality or algorithm limitations, thereby obtaining a more continuous and complete road and waterway network vector layer. Sub-step S13: Network Fusion and Partition Construction. The repaired road network and water system network vector layers are spatially overlaid. A network fusion processing tool is invoked. This tool automatically identifies each smallest closed polygon region enclosed by these linear features by calculating their intersection and enclosure relationships, and outputs each such polygon as an independent geographic partition. The collection of all partitions constitutes the "partition control network".

[0025] Furthermore, to further enhance the geographical rationality of the zoning control network, especially in mountainous areas with significant topographic relief, step S1 can also include an optional sub-step S14: topographic feature line extraction. When the input data includes a digital elevation model (DEM), the system activates a topographic analysis tool. This tool calculates topographic factors such as the aspect and curvature of the DEM and uses an automatic ridgeline / valleyline extraction algorithm to generate a topographic feature line vector layer. Subsequently, in sub-step S13, these topographic feature lines are spatially overlaid with the road network and water system network to construct a more refined zoning control network that better conforms to the natural geographical pattern.

[0026] Next, in step S2, hierarchical intelligent extraction is performed. This step is executed independently within each geographic partition generated in step S1, extracting geographic features category by category in order of increasing difficulty and from visible to hidden, and avoiding interference between categories through masking techniques. This step is the core of achieving "hierarchical extraction".

[0027] In one specific embodiment, the system pre-configures a "dedicated model pool" that stores multiple semantic segmentation model files optimized for different land features (such as buildings, water bodies, farmland, forest land, bare land, etc.). These models are all relatively simplified deep learning models. For example, a dedicated, simplified DeepLabv3+ model can be trained for the building extraction task, and a dedicated D-LinkNet model can be trained for the water body extraction task. Each model only needs to be proficient in recognizing one or a few types of land features and does not need to learn the complex discrimination boundaries of all categories.

[0028] The layered extraction process is handled by a "layered extraction executor." The executor first reads the image data of the current partition and creates a blank mask layer of the same size as the image to record the classified areas. Then, the executor iteratively extracts data according to a preset category order. This preset order, such as "buildings → water bodies → farmland → woodland → bare land," is usually arranged according to the visual salience of the features in the image from strong to weak, or from easy to difficult. Users can also flexibly adjust this order through the system interface.

[0029] In the first layer of extraction, the executor loads a "building extraction model" from the model pool, inputs it along with the current partition image into the prediction unit, and the model outputs a building prediction probability map, which is then thresholded to obtain a binarized result. This result is used to temporarily convert into vector patches and to update the mask layer, marking the pixels occupied by buildings as "occupied".

[0030] Before extracting the second layer (e.g., water bodies), the executor invokes a crucial "masking unit." This unit performs logical operations on the current partitioned image and the updated mask layer. Specifically, it sets the pixel values ​​marked as "occupied" in the mask to zero or a special ignore value, thereby generating an image to be processed that "only contains unclassified areas." Then, the executor loads the "water extraction model" and makes predictions only on this masked image. Since classified building areas are physically masked, the water model does not need to consider these areas at all, thus fundamentally avoiding the possibility of misclassifying water bodies as buildings or vice versa. This process iterates until all preset categories have been extracted.

[0031] Within this partition, all extracted vector graphics are overlaid and merged to form the initial interpretation result of all elements in this partition. Its core principle is "incremental" and "exclusion method". Through masking operations, the classified regions are technically isolated, and the misjudgment of the determined regions by the subsequent model is physically prevented, effectively reducing inter-class confusion.

[0032] Finally, in step S3, the results are generated and output. After all partitions have been extracted hierarchically, the "Results Mosaic and Summary Tool" is invoked to spatially fuse and stitch together the vector results of all geographic partitions. This tool automatically handles the edge matching problem of map features at the partition boundaries, ensuring that the final result is a seamless, attribute-rich, and complete vector interpretation result covering the entire region.

[0033] To automate and engineer the above methods, this invention also provides a remote sensing image feature interpretation system that integrates zoning control and hierarchical extraction. In one embodiment, the system mainly includes three core modules: a zoning control network generation module, a hierarchical intelligent extraction module, and a process engine and production line management module.

[0034] The zonal control network generation module performs the function of step S1 described above. It receives raw remote sensing imagery and auxiliary data, and outputs geographically meaningful zonal network data. The hierarchical intelligent extraction module has a built-in model pool containing multiple dedicated semantic segmentation models, as well as a hierarchical extraction executor, which performs the function of step S2 described above.

[0035] The process engine and production line management module are key to achieving full-process automation in this system. It connects all the discrete technical steps—data input, partitioning, hierarchical extraction, and result fusion—in a visual way, forming a customizable, repeatable, and monitorable automated production line. This module specifically includes three sub-units: a graphical process designer, a process interpreter and scheduler, and a task status monitor.

[0036] In its implementation, the system first encapsulates each independent function—including "image reading," "zoning network generation," "building model prediction," "vector post-processing (smoothing, fragment removal)," and "result mosaicking and summarization"—into a "tool node" with standard input and output interfaces. In the graphical workflow designer, users can intuitively define data flow and processing order by dragging and dropping the required tool nodes from the tool library onto the canvas and connecting the output port of one node to the input port of the next node with connecting lines. For example, a "full-feature extraction production line" can be constructed: Input node → Zoning network generation tool → Layered extraction tool → Vector fusion tool → Attribute assignment tool → Output node. Users can also click on each tool node to configure its operating parameters in the pop-up interface, such as specifying a specific list of ground feature extraction sequences for the "layered extraction tool" and setting area thresholds and smoothing coefficients for the "vector post-processing tool." After saving the workflow template, the user can submit it as a production task. The workflow scheduler starts in the background, strictly following the logic of the flowchart, instantiating each tool sequentially, passing intermediate data, and executing it. Meanwhile, the task status monitor interface will highlight the currently executing tool nodes in the form of a flowchart, and display logs, progress bars, and output snapshots of each step in real time, realizing full-process visual monitoring.

[0037] As an important application embodiment of the above system, this invention also provides the application of the interpretation system in change detection. Users can use the workflow engine to orchestrate an automated change detection production line. The specific implementation of this production line is as follows: using a graphical workflow designer, a production line containing two parallel branches is created. One branch processes earlier images, and the other branch processes later images. Each branch contains a complete, independently runnable "full-element extraction sub-flow," that is, each branch calls tools such as zonal control network generation and hierarchical intelligent extraction to output vector results for its respective time phase. At the end of the two branches (i.e., after obtaining their respective vector results), a "change detection tool" node is connected. This tool receives vector data from the two periods, and its internal algorithm first performs spatial overlay analysis. By comparing the spatial location and category of the patches in the two periods, it identifies patches of types such as newly added, disappeared, and range-changed. To further improve the boundary accuracy and confidence of changed patches, the tool can also call an optional sub-function, namely, extracting the original image blocks from the earlier and later periods corresponding to the changed areas, and using a lightweight pixel-level difference analysis model for fine verification. Finally, the production line can also be connected to a "change attribute association tool" node. This tool automatically associates and assigns semantic attributes to the identified change patches from a predefined knowledge base or attribute table based on their spatial location and pre- and post-construction categories, such as "farmland converted to construction land", "forest land logging", and "water area expansion", thereby generating final change monitoring results with rich semantic information.

[0038] In summary, this invention, through its technical framework of "partition control, hierarchical extraction, and process-driven approach," systematically solves the problems of low computational efficiency, severe inter-class confusion, lack of geographic logic in results, and poor process automation in existing technologies, demonstrating significant progress and practicality.

[0039] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0040] It should be noted that the components mentioned in the above embodiments are all general standard parts or components known to those skilled in the art. Their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0041] This invention has illustrated its principles and implementation methods using specific examples. The descriptions of these embodiments are merely illustrative of the method and its core ideas; furthermore, those skilled in the art will recognize that modifications may be made to the specific implementation methods and application scope based on the principles of this invention. Therefore, the content of this specification should not be construed as limiting the invention.

Claims

1. A method for interpreting all features of remote sensing images by integrating zoning control and hierarchical extraction, characterized in that: Includes the following steps: S1. Constructing a regional control network: Based on the input remote sensing image, extract the linear elements that constitute the geospatial framework, and generate a natural geographic regional network composed of closed polygons through topology repair and fusion. S2. Hierarchical intelligent extraction: For each geographic partition, according to the preset land cover category order, the dedicated semantic segmentation model corresponding to the category is called and executed in sequence. After the extraction of each category is completed, the classified mask of the partition is updated so that the extraction of subsequent categories is only performed on the pixel area not covered by the current mask. S3. Results Generation and Output: Spatial fusion and stitching of the feature patches extracted from all geographic zones to generate complete vector interpretation results covering the entire region.

2. The method for interpreting all features of remote sensing images by fusing zoning control and hierarchical extraction according to claim 1, characterized in that, Step S1 further includes: S11. Use a pre-trained lightweight model to extract initial linear features of major roads and waterways from remote sensing images; S12. Vectorize the initial linear elements and perform breakpoint connection and topology repair based on graph theory algorithms to form a continuous road network and water system network vector layer; S13. The road network and water system network vector layers are spatially superimposed, and multiple minimum closed polygons are automatically divided according to their intersection relationship, with each polygon serving as an independent geographical partition.

3. The method for interpreting all features of remote sensing images by fusing zoning control and hierarchical extraction according to claim 2, characterized in that, Step S1 further includes: when the input contains digital elevation model data, extracting ridge lines and valley lines as terrain feature lines, and spatially superimposing the terrain feature lines with the road network and water system network vector layers to construct the zonal control network.

4. The method for interpreting all features of remote sensing images by fusing zoning control and hierarchical extraction according to claim 1, characterized in that, In step S2, the preset order of land cover categories is arranged according to the visual salience of land cover in the image from strong to weak, or from easy to difficult.

5. The method for interpreting all features of remote sensing images by fusing zoning control and hierarchical extraction according to claim 1, characterized in that, In step S2, the dedicated semantic segmentation model is stored in a model pool, and each model is a relatively simplified deep learning model that is optimized and trained for a single or a small number of specific land cover categories.

6. A remote sensing image land feature full-element interpretation system that implements the method of any one of claims 1 to 5, characterized in that, include: The zonal control network generation module is used to extract linear geographic features from input remote sensing images, perform topology repair and network fusion, and generate a zonal control network composed of multiple geographic zones. The hierarchical intelligent extraction module has a built-in model pool containing multiple dedicated semantic segmentation models, as well as a hierarchical extraction executor. The hierarchical extraction executor is used to sequentially load the models in the model pool for prediction within each geographic partition, and to isolate classified areas through mask update technology, guiding subsequent models to process only unclassified areas. The process engine and production line management module are used to encapsulate the data input, partition network generation, hierarchical extraction, and result integration into visual and programmable tool nodes, and schedule and execute the processing flow formed by the user, ultimately outputting the interpretation results.

7. The remote sensing image land cover full-element interpretation system integrating zoning control and hierarchical extraction according to claim 6, characterized in that, The process engine and production line management module include: A graphical process designer provides a user interface that allows users to build custom interpretation processes by dragging and dropping and connecting pre-packaged tool nodes; The process interpreter and scheduler parses the user-designed flowchart in the background and strictly calls the instantiated tool nodes to execute tasks in sequence according to the data flow. The task status monitor is used to display the process execution progress, current running nodes, and log information in real time.

8. The remote sensing image land cover full-element interpretation system integrating zoning control and hierarchical extraction according to claim 7, characterized in that, The tool nodes include an image reading node, a zoning network generation node, a model prediction node corresponding to at least one land cover category, a vector post-processing node, and a result mosaicking and summarizing node. Each node has a standardized input and output interface.

9. The application of the remote sensing image land cover full-element interpretation system integrating zoning control and hierarchical extraction as described in claim 6 in change detection, characterized in that, A change detection production line is orchestrated based on the aforementioned process engine and production line management module. The production line includes: The first full-element extraction sub-process is used to process the remote sensing image of the first time phase and output the first vector result; The second full-element extraction sub-process, set in parallel or serial with the first full-element extraction sub-process, is used to process remote sensing images of the second time phase and output the second vector result. The change detection tool has its input end connected to the output end of the first and second full-feature extraction sub-processes. It is used to perform spatial overlay analysis on the first and second vector results to identify and extract changed land feature patches.

10. The application according to claim 9, characterized in that, The change detection production line also includes a change attribute association tool, which is used to automatically assign semantic attributes indicating the change of land cover type based on the spatial location and category of the changed land cover patch.

Images