A large-format remote sensing image multi-target extraction method and system

Abstract

The application belongs to the technical field of remote sensing earth observation, and particularly relates to a large-format remote sensing image multi-target extraction method and system, which comprises S1 large-format overlapping sliding window segmentation; S2 cross-modal space mapping and geometric prior prompt generation; S3 adaptive segmentation target function based on spatial prior constraint; S4 affine restoration of local pixel coordinates to global coordinates; and S5 overlapping area mask confidence weighted fusion and final result generation. A visual language multi-modal large model is used as a semantic analysis front end, and a segmentation large model is used as a pixel-level decoding back end, the two are connected through a spatial prior prompt, and the whole process is driven by a natural language multi-round dialogue, so that end-to-end high-precision automatic extraction of ground objects is realized.

Description

Technical Field

[0001] This application belongs to the field of remote sensing Earth observation technology, specifically relating to a method and system for extracting multiple targets from large-format remote sensing images. Background Technology

[0002] With the continuous development of remote sensing Earth observation technology, the spatial resolution and coverage width of remote sensing images are constantly improving, and the data volume of a single image can reach several gigabytes. How to quickly and accurately extract target features (such as buildings, roads, water bodies, etc.) from massive high-resolution images has important application value in fields such as urban planning and disaster monitoring.

[0003] In recent years, the Vision-Language Model (VLM) and general segmentation models (such as SegmentAnythingModel3, SAM3) have provided new technical approaches for remote sensing ground feature extraction. However, when directly applying these general models to remote sensing image ground feature extraction, the following technical problems still exist: (1) Feature loss problem caused by limited input resolution of VLM. VLM has an upper limit on input resolution. When processing high-resolution remote sensing images, the original images usually need to be compressed into low-resolution thumbnails before being sent into the network. Since the background of remote sensing images is complex and contains a large number of dense small targets, global downsampling will cause key spatial details and small target information to be lost during compression, affecting the model's detection recall rate for small targets.

[0004] (2) SAM3 is highly dependent on cue and confidence threshold. SAM3 has a certain zero-shot segmentation generalization ability, but in practical applications it needs to rely on manually specified prior cue and confidence threshold. In multi-type land cover extraction scenarios, the model output confidence distribution of different morphological targets (such as large water bodies and scattered buildings) varies significantly. Fixed thresholds are difficult to adapt to complex scenarios and often require repeated manual parameter tuning, which is not conducive to fully automated processing. Summary of the Invention

[0005] To overcome the problems of target loss due to downsampling, coarse VLM boundaries, excessive reliance on manual thresholding in SAM3, and rigid interaction methods in existing technologies for processing large-format remote sensing images, this application provides a method and system for multi-target extraction from large-format remote sensing images. The technical solution is as follows: A method for extracting multiple targets from large-format remote sensing images includes the following steps: S1. Large-format overlapping sliding window segmentation; Based on the preset slice window size and overlap rate, the horizontal and vertical sliding step size is calculated to generate a set of local sub-images that traverse the entire image; each local sub-image records the starting offset coordinates of its upper left pixel in the global image coordinate system during generation, for subsequent coordinate reconstruction. S2. Cross-modal space mapping and geometric prior hint generation; The VLM performs cross-modal semantic parsing on the original resolution sub-image in conjunction with the user's natural language instructions, and outputs a set of target bounding boxes; then the bounding boxes are filtered for validity and converted in format by the geometric mapping function to generate geometric prior hints acceptable to SAM3; S3. Adaptive segmentation objective function based on spatial prior constraints; Based on geometric priors, SAM3 constrains the segmentation search domain to the local pixel range defined by each prior bounding box, adaptively solves for the foreground mask with the best posterior probability in this region, and outputs the corresponding prediction confidence score. S4. Affine reconstruction from local pixel coordinates to global coordinates; Using the initial offset coordinates of each sub-image, perform affine translation transformation on each pixel on the local mask boundary to restore them to the global pixel coordinate system, thus obtaining the global mask representation. S5. Overlapping region mask confidence-weighted fusion and final result generation; For each mask restored to the global coordinate system, a weighted fusion is performed based on the confidence score. The prediction result with the best confidence in the overlapping area is retained, duplicate predictions are eliminated, and a complete global vector feature extraction result is generated after connected component extraction.

[0006] Preferably, step S1, large-format overlapping sliding window segmentation, specifically includes: Let the pixel resolution of the original remote sensing image be... The pixel size of the local slice window is The preset overlap rate of adjacent slices , The pixel step size of the sliding window in the horizontal and vertical directions. , They are respectively: ; ; No. The starting pixel coordinates of the top left corner of each slice window for: ; ; in , The corresponding slice subgraph is denoted as All subgraphs form a set: ; Because adjacent slices have a horizontal width of and vertical direction In overlapping areas, targets that cross slice boundaries can be fully preserved in at least one slice.

[0007] Preferably, step S2 specifically includes: S21. For sets Each local subgraph in This will be compared with the user's natural language input to extract intent. Both are fed into the VLM (Visual Model for Objects) for cross-modal semantic parsing. The VLM acts as a coarse-grained object detector, outputting a set of object bounding boxes in the sub-image pixel coordinate system. ; ; in Indicates the first The bounding box of each target. This refers to the coordinates of the top-left corner of the frame in the sub-image's pixel coordinate system. , These are the pixel width and height of the frame, respectively; S22. Through geometric mapping functions Set the bounding boxes Conversion to SAM3 acceptable geometry hints: ； Its function is to each Perform format conversion and validity filtering to make it meet the input specifications of SAM3 geometry prompts.

[0008] Preferably, step S3 is performed by... Introducing strong local spatial constraints into SAM3 transforms the segmentation task from "global threshold filtering" to "solving the optimal connected component within the prior bounding box"; For subgraph Each bounding box Region of interest Its adaptive segmentation objective function is defined as: ; The meaning is a candidate mask variable, which represents a binary segmentation scheme for a set of pixels (each pixel is labeled as foreground or background); The meaning refers to the posterior probability function output by the SAM3 model, which represents the probability of a candidate mask given a subgraph and geometric cues. This represents the probability value of the actual target region. The higher the value, the better the match between the mask and the target ground features; By limiting the search domain to Within this framework, the system adaptively acquires the land cover region with the best responsivity within each prior bounding box, and SAM3 simultaneously outputs the prediction confidence score corresponding to this local mask. This is for use in subsequent merging of overlapping areas.

[0009] Preferably, the local mask of the SAM3 decoding output Expressed in sub-image pixel coordinates, with a local mask. The subgraph coordinates of any pixel on the boundary are Corresponding slice The starting coordinates of the top left corner in the global pixel coordinate system are: The local coordinates are restored to global pixel coordinates through the following affine translation transformation. : ; Right now , ; right Perform the above transformation on each boundary pixel to obtain its mask representation in the global pixel coordinate system. If the image includes georeferenced information, it will Mapped to geographic coordinates.

[0010] Preferably, since there are overlapping areas between adjacent slices, the same ground feature is predicted independently in multiple slices, resulting in duplicate masks. A pixel-level weighted fusion strategy based on SAM3 prediction confidence score is used to eliminate redundancy.

[0011] Preferably, in step S5, two masks to be fused are assumed to exist in the global coordinate system. and The corresponding SAM3 prediction confidence scores are as follows: and For global coordinates The pixel at that location, the foreground probability value after fusion. Defined as: ; in , indicating mask The predicted foreground value at the corresponding pixel is set to 1 for foreground and 0 for background.

[0012] Preferably, for the fused global probability field Perform connected component extraction, retain connected regions that meet the minimum area threshold, and finally convert them into vector polygons to generate a seamless, target-free global vector extraction result.

[0013] A multi-target extraction system for large-format remote sensing images includes a large-format overlapping sliding window segmentation module, a cross-modal spatial mapping and geometric prior prompt generation module, an adaptive segmentation objective function module based on spatial prior constraints, an affine reconstruction module from local pixel coordinates to global coordinates, and a final result generation module. Large-format overlapping sliding window segmentation module: Calculates the horizontal and vertical sliding step size based on the preset slice window size and overlap rate, and generates a set of local sub-images that traverse the entire image; each local sub-image records the starting offset coordinates of its upper left pixel in the global image coordinate system during generation, for subsequent coordinate reconstruction. The cross-modal spatial mapping and geometric prior hint generation module enables VLM to perform cross-modal semantic parsing on the original resolution sub-image in conjunction with user natural language commands, and outputs a set of target bounding boxes; then, the geometric mapping function performs validity filtering and format conversion on the bounding boxes to generate SAM3 acceptable geometric prior hints; The adaptive segmentation objective function module based on spatial prior constraints: SAM3 constrains the segmentation search domain to the local pixel range defined by each prior box based on geometric prior hints, adaptively solves for the foreground mask with the best posterior probability in this region, and outputs the corresponding prediction confidence score. Affine restoration module from local pixel coordinates to global coordinates: Using the starting offset coordinates of each sub-image recorded by the large-format overlapping sliding window segmentation module, the pixel points on the local mask boundary are transformed one by one into the global pixel coordinate system to obtain the global mask representation. Final result generation module: For each mask restored to the global coordinate system, weighted fusion is performed based on the confidence score, the prediction result with the best confidence in the overlapping area is retained, duplicate predictions are eliminated, and a complete global vector feature extraction result is generated after connected component extraction.

[0014] Preferably, the system performs a low-level topological edge-finding algorithm on the fused global high-precision binary mask image to extract closed vector polygons and generate exportable vector results. The system automatically reads the coordinate information file that accompanies the uploaded remote sensing image, and combines it with the corresponding six-parameter affine transformation model to directly map the pixel plane coordinates of the mask to the real geospatial coordinates. Then, the vector polygons with real projections are packaged with the corresponding natural language semantic tags and directly output as standard file format, forming a complete business closed loop from "language input" to "professional surveying and mapping results output".

[0015] Compared with the prior art, the beneficial effects of this application are as follows: The core idea of ​​this invention is to construct a cascaded architecture of "large model guiding large model": the visual language multimodal large model (VLM) is used as the semantic parsing front end, and the segmentation large model (SAM3) is used as the pixel-level decoding back end. The two are connected through spatial prior prompts (GeometricPrompt), and the whole process is driven by multi-turn dialogue of natural language to achieve end-to-end high-precision automatic extraction of ground features. Attached Figure Description

[0016] Figure 1 For processing flowcharts; Figure 2 This is a logic diagram of the algorithm for fusion of overlapping sliding window slices and target restoration of large-format remote sensing images. Detailed Implementation

[0017] 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.

[0018] (1) Large-format overlapping sliding window segmentation: To avoid the loss of small target information caused by directly compressing large-format remote sensing images into a VLM, this invention introduces an overlapping sliding window slicing mechanism.

[0019] Let the pixel resolution of the original remote sensing image be... The pixel size of the local slice window is The preset overlap rate of adjacent slices The pixel step size of the sliding window in the horizontal and vertical directions. , They are respectively: ; ; No. The starting pixel coordinates of the top left corner of each slice window for: ; ; in , The corresponding slice subgraph is denoted as... All subgraphs form a set: ; Because adjacent slices have a width of (Horizontal direction) and In the (vertical) overlapping area, targets that cross the slice boundary can be completely preserved in at least one slice, thus mechanistically avoiding the problem of small targets being lost due to downsampling.

[0020] (2) Cross-modal space mapping and geometric prior hint generation: For sets Each local subgraph in This will be compared with the user's natural language input to extract intent. Both are fed into the VLM (Visual Model for Imaging) for cross-modal semantic parsing. The VLM, acting as a coarse-grained object detector, outputs a set of object bounding boxes in the sub-image pixel coordinate system. ； ; in Indicates the first The bounding box of each target. This refers to the coordinates of the top-left corner of the frame in the sub-image's pixel coordinate system. , These represent the pixel width and height of the frame, respectively.

[0021] Subsequently, through the geometric mapping function Set the bounding boxes Conversion to SAM3 acceptable geometry hints: ； Its function is to each Perform format conversion and validity filtering (such as removing boxes with low confidence or abnormal size) to make them meet the input specifications of SAM3 geometry hints.

[0022] (3) Adaptive segmentation objective function based on spatial prior constraints: Traditional SAM3 uses a fixed global confidence threshold. Foreground truncation is applied to the predicted probability map, i.e.: ; In remote sensing scenarios involving the extraction of multiple types of land features, the confidence distribution of model responses for different target shapes varies significantly, making it difficult to adapt to fixed thresholds.

[0023] This invention is achieved through Introducing strong local spatial constraints into SAM3 transforms the segmentation task from "global threshold filtering" to "solving for the optimal connected component within the prior bounding box". For subgraphs... Each bounding box Region of interest (Right now The adaptive segmentation objective function for the set of pixels within the range is defined as: ; This formula limits the search domain to Within this framework, the system adaptively acquires the terrain feature region with the best responsivity within each prior bounding box, eliminating the need for manually specifying a uniform global confidence threshold. SAM3 simultaneously outputs the prediction confidence score corresponding to this local mask. This is for use in subsequent merging of overlapping areas.

[0024] (4) Affine reconstruction from local pixel coordinates to global coordinates: SAM3 decoder output partial mask Expressed in subimage pixel coordinates. Let the subimage coordinates of any pixel on the mask boundary be... Corresponding slice The starting coordinates of the top left corner in the global pixel coordinate system are: The local coordinates are restored to global pixel coordinates through the following affine translation transformation. : ; Right now , .

[0025] right Perform the above transformation on each boundary pixel to obtain its mask representation in the global pixel coordinate system. If the imagery includes georeferenced information (GeoTransform), it can be further... Mapped to geographic coordinates.

[0026] (5) Overlapping region mask confidence-weighted fusion and final result generation: Because adjacent slices overlap, the same ground feature may be predicted independently in multiple slices, resulting in duplicate masks. This invention employs a pixel-level weighted fusion strategy based on SAM3 prediction confidence scores to eliminate redundancy.

[0027] Suppose there are two masks to be fused in the global coordinate system. and The corresponding SAM3 prediction confidence scores are as follows: and For global coordinates The pixel at that location, the foreground probability value after fusion. Defined as: ; in , indicating mask The predicted foreground value at the corresponding pixel is set to 1 for foreground and 0 for background. This max operation preserves prediction results with higher confidence and effectively suppresses boundary inconsistencies caused by repeated predictions in overlapping areas.

[0028] For the fused global probability field Perform connected component extraction, retain connected regions that meet the minimum area threshold (which can be set according to task requirements), and finally convert them into vector polygons to generate a seamless, target-free global vector extraction result.

[0029] A multi-target extraction system for large-format remote sensing images includes a large-format overlapping sliding window segmentation module, a cross-modal spatial mapping and geometric prior prompt generation module, an adaptive segmentation objective function module based on spatial prior constraints, an affine reconstruction module from local pixel coordinates to global coordinates, and a final result generation module. Large-format overlapping sliding window segmentation module: Calculates the horizontal and vertical sliding step size based on the preset slice window size and overlap rate, and generates a set of local sub-images that traverse the entire image; each local sub-image records the starting offset coordinates of its upper left pixel in the global image coordinate system during generation, for subsequent coordinate reconstruction. The cross-modal spatial mapping and geometric prior hint generation module enables VLM to perform cross-modal semantic parsing on the original resolution sub-image in conjunction with user natural language commands, and outputs a set of target bounding boxes; then, the geometric mapping function performs validity filtering and format conversion on the bounding boxes to generate SAM3 acceptable geometric prior hints; The adaptive segmentation objective function module based on spatial prior constraints: SAM3 constrains the segmentation search domain to the local pixel range defined by each prior box based on geometric prior hints, adaptively solves for the foreground mask with the best posterior probability in this region, and outputs the corresponding prediction confidence score. Affine restoration module from local pixel coordinates to global coordinates: Using the starting offset coordinates of each sub-image recorded by the large-format overlapping sliding window segmentation module, the pixel points on the local mask boundary are transformed one by one into the global pixel coordinate system to obtain the global mask representation. Final result generation module: For each mask restored to the global coordinate system, weighted fusion is performed based on the confidence score, the prediction result with the best confidence in the overlapping area is retained, duplicate predictions are eliminated, and a complete global vector feature extraction result is generated after connected component extraction.

[0030] The system performs a low-level topological edge-finding algorithm on the fused global high-precision binary mask image to extract closed vector polygons and generate exportable vector results. The system automatically reads the coordinate information file that accompanies the uploaded remote sensing image and, combined with the corresponding six-parameter affine transformation model, directly maps the pixel plane coordinates of the mask to the real geospatial coordinates. Then, the vector polygons with the real projection are packaged with the corresponding natural language semantic tags and directly output as a standard file format, forming a complete business closed loop from "language input" to "professional surveying and mapping results output".

[0031] Standard file formats include, but are not limited to, Shapefile (.shp), GeoJSON (.geojson), GeoPackage (.gpkg), and KML / KMZ (.kml / .kmz) formats.

[0032] Example: This embodiment uses a typical remote sensing business scenario (such as extracting "all buildings with red roofs" from a remote sensing image of a certain area) as an example to illustrate in detail the complete end-to-end operation process of the system: Step 1: System Initialization and Dynamic Configuration of the Large Model Interface. Due to the significant computational resource consumption of the Visual Language Model (VLM) and the Segmentation Model (SAM3), this system adopts a separate engine deployment architecture. At the interactive front end, the user first dynamically inputs the authentication information (including APIKey, BaseURL, and the specific large model model selected) of the remote or local VLM through the configuration panel. SAM3 then serves as the locally resident pixel-level decoding and inference engine. This step ensures loose coupling and high scalability across multiple models.

[0033] Step 2: Large-Format Image Upload and Adaptive Overlapping Sliding Window Tile The user uploads a high-resolution remote sensing image, several gigabytes in size, to the system. The system backend uses memory-mapping technology to read the image on demand, avoiding physical memory overflow. To address the feature loss caused by directly scaling the entire image in traditional VLM, the system performs an overlapping sliding window tile algorithm on the image: Let the tile resolution be... Pixels, overlap rate The step size is set to 10%~20%. The system calculates the step size based on the sliding formula and generates a sequence of local sub-maps that traverse the entire image. All local sub-maps synchronously save their origin offset coordinates in the global image. .

[0034] Step 3: Multi-turn dialogue based on natural language and spatial prior extraction using VLM. The user inputs a natural language command on the interactive interface, such as: "Please help me extract all buildings with red roofs in the image." The system packages each local sub-image with the text command and sends it to the VLM, which has been configured with authentication information. Based on its strong visual and common-sense reasoning capabilities, the VLM filters out irrelevant features (such as gray buildings, red objects that are not buildings, etc.) and accurately outputs the bounding box coordinates (BoundingBoxHints) of the target in each local sub-image. If the user is not satisfied with the initial results, they can continue to impose constraints on the VLM through multi-turn dialogue using natural language (such as adding a request like "exclude houses with chimneys"). The VLM will dynamically update the output bounding boxes in real time.

[0035] Step 4: SAM3 Zero-Shot Local Fine-Tuning and Threshold-Free High-Precision Decoding. For each output sub-image with bounding box priors, the system uses its corresponding bounding box as a strong geometric prompt input to the underlying SAM3 model. Since VLM has already performed rigorous semantic-level filtering and coarse localization, SAM3 no longer needs to rely on manual adjustments to the global confidence threshold to distinguish different land features. SAM3 only needs to calculate the maximum probability connected component within its bounding box local region to generate a local high-precision mask accurate to the pixel level.

[0036] Step 5: Global Inverse Coordinate Mapping and Overlapping Area Fusion Deduplication. After obtaining all local masks, the system uses the global origin offset coordinates retained in Step 2 to losslessly map each local mask back to the original panoramic coordinate system through a translation mapping matrix. To address the potential for repeated predictions of the same feature in overlapping sliding window areas, the system triggers a confidence-based fusion deduplication algorithm: comparing the SAM3 prediction scores of two overlapping pixels, retaining the pixel with the higher score, or performing non-maximum suppression (NMS) by calculating the intersection-over-union (IoU) ratio. This step completely eliminates grid truncation artifacts, ensuring seamless continuity of entities across the entire image.

[0037] Step 6: Export the fused global high-precision binary mask image from the high-precision vector results. Closed vector polygons will be extracted using a low-level topological edge-finding algorithm. The system automatically reads the coordinate information file (such as a .tfw projection file) accompanying the uploaded remote sensing image and, combined with the corresponding six-parameter affine transformation model, directly maps the pixel planar coordinates of the mask to the actual geospatial coordinates. Subsequently, the system packages the vector polygons with the actual projection and the corresponding natural language semantic tags, directly outputting them as a standard Shapefile (.shp) file, forming a complete business loop from "language input" to "professional surveying and mapping results output."

[0038] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for multi-target extraction from large-format remote sensing images, characterized in that, Includes the following steps: S1. Large-format overlapping sliding window segmentation; Based on the preset slice window size and overlap rate, the horizontal and vertical sliding step size is calculated to generate a set of local sub-images that traverse the entire image; each local sub-image records the starting offset coordinates of its upper left pixel in the global image coordinate system during generation, for subsequent coordinate reconstruction. S2. Cross-modal space mapping and geometric prior hint generation; The VLM performs cross-modal semantic parsing on the original resolution sub-image in conjunction with the user's natural language instructions, and outputs a set of target bounding boxes; then the bounding boxes are filtered for validity and converted in format by the geometric mapping function to generate geometric prior hints acceptable to SAM3; S3. Adaptive segmentation objective function based on spatial prior constraints; Based on geometric priors, SAM3 constrains the segmentation search domain to the local pixel range defined by each prior bounding box, adaptively solves for the foreground mask with the best posterior probability in this region, and outputs the corresponding prediction confidence score. S4. Affine reconstruction from local pixel coordinates to global coordinates; Using the initial offset coordinates of each sub-image, perform affine translation transformation on each pixel on the local mask boundary to restore them to the global pixel coordinate system, thus obtaining the global mask representation. S5. Overlapping region mask confidence-weighted fusion and final result generation; For each mask restored to the global coordinate system, a weighted fusion is performed based on the confidence score. The prediction result with the best confidence in the overlapping area is retained, duplicate predictions are eliminated, and a complete global vector feature extraction result is generated after connected component extraction.

2. The method for multi-target extraction from large-format remote sensing images according to claim 1, characterized in that, Step S1, large-format overlapping sliding window segmentation, is as follows: Let the pixel resolution of the original remote sensing image be... The pixel size of the local slice window is The preset overlap rate of adjacent slices , The pixel step size of the sliding window in the horizontal and vertical directions. , They are respectively: ； ； No. The starting pixel coordinates of the top left corner of each slice window for: ； ； in , The corresponding slice subgraph is denoted as All subgraphs form a set: ； Because adjacent slices have a horizontal width of and vertical direction In overlapping areas, targets that cross slice boundaries can be fully preserved in at least one slice.

3. The method for multi-target extraction from large-format remote sensing images according to claim 2, characterized in that, Step S2 is as follows: S21. For sets Each local subgraph in This will be compared with the user's natural language input to extract intent. Both are fed into the VLM (Visual Model for Objects) for cross-modal semantic parsing. The VLM acts as a coarse-grained object detector, outputting a set of object bounding boxes in the sub-image pixel coordinate system. ； ； in Indicates the first The bounding box of each target. This refers to the coordinates of the top-left corner of the frame in the sub-image's pixel coordinate system. , These are the pixel width and height of the frame, respectively; S22. Through geometric mapping functions Set of bounding boxes Conversion to SAM3 acceptable geometry hints: ； Its function is to each Perform format conversion and validity filtering to make it meet the input specifications of SAM3 geometry prompts.

4. The method for multi-target extraction from large-format remote sensing images according to claim 3, characterized in that, Step S3 through Introducing strong local spatial constraints into SAM3 transforms the segmentation task from "global threshold filtering" to "solving the optimal connected component within the prior bounding box"; For subgraph Each bounding box Region of interest Its adaptive segmentation objective function is defined as: ； These are candidate mask variables; It is the posterior probability function output by the SAM3 model; By limiting the search domain to Within this framework, the system adaptively acquires the land cover region with the best responsivity within each prior bounding box, and SAM3 simultaneously outputs the prediction confidence score corresponding to this local mask. This is for use in subsequent merging of overlapping areas.

5. The method for multi-target extraction from large-format remote sensing images according to claim 4, characterized in that, SAM3 decoder output partial mask Expressed in sub-image pixel coordinates, with a local mask. The subgraph coordinates of any pixel on the boundary are Corresponding slice The starting coordinates of the top left corner in the global pixel coordinate system are: The local coordinates are restored to global pixel coordinates through the following affine translation transformation. : ； Right now , ; right Perform the above transformation on each boundary pixel to obtain its mask representation in the global pixel coordinate system. If the image includes georeferenced information, it will Mapped to geographic coordinates.

6. The method for multi-target extraction from large-format remote sensing images according to claim 4, characterized in that, Because adjacent slices have overlapping areas, the same ground feature is predicted independently in multiple slices, resulting in duplicate masks. A pixel-level weighted fusion strategy based on SAM3 prediction confidence scores is used to eliminate redundancy.

7. The method for multi-target extraction from large-format remote sensing images according to claim 6, characterized in that, Step S5: Assume there are two masks to be fused in the global coordinate system. and The corresponding SAM3 prediction confidence scores are as follows: and For global coordinates The pixel at that location, the foreground probability value after fusion. Defined as: ； in Indicates mask The predicted foreground value at the corresponding pixel is set to 1 for foreground and 0 for background.

8. The method for multi-target extraction from large-format remote sensing images according to claim 7, characterized in that, For the fused global probability field Perform connected component extraction, retain connected regions that meet the minimum area threshold, and finally convert them into vector polygons to generate a seamless, target-free global vector extraction result.

9. A multi-target extraction system for large-format remote sensing images, characterized in that, It includes a large-format overlapping sliding window segmentation module, a cross-modal space mapping and geometric prior prompt generation module, an adaptive segmentation objective function module based on spatial prior constraints, an affine reconstruction module from local pixel coordinates to global coordinates, and a final result generation module; Large-format overlapping sliding window segmentation module: Calculates the horizontal and vertical sliding step size based on the preset slice window size and overlap rate, and generates a set of local sub-images that traverse the entire image; each local sub-image records the starting offset coordinates of its upper left pixel in the global image coordinate system during generation, for subsequent coordinate reconstruction. The cross-modal spatial mapping and geometric prior hint generation module enables VLM to perform cross-modal semantic parsing on the original resolution sub-image in conjunction with user natural language commands, and outputs a set of target bounding boxes; then, the geometric mapping function performs validity filtering and format conversion on the bounding boxes to generate SAM3 acceptable geometric prior hints; The adaptive segmentation objective function module based on spatial prior constraints: SAM3 constrains the segmentation search domain to the local pixel range defined by each prior box based on geometric prior hints, adaptively solves for the foreground mask with the best posterior probability in this region, and outputs the corresponding prediction confidence score. Affine transformation module from local pixel coordinates to global coordinates: Using the starting offset coordinates of each sub-image recorded by the large-format overlapping sliding window segmentation module, the pixel points on the local mask boundary are transformed one by one into the global pixel coordinate system to obtain the global mask representation. Final result generation module: For each mask restored to the global coordinate system, weighted fusion is performed based on the confidence score, the prediction result with the best confidence in the overlapping area is retained, duplicate predictions are eliminated, and a complete global vector feature extraction result is generated after connected component extraction.

10. A multi-target extraction system for large-format remote sensing images according to claim 9, characterized in that, The system performs a low-level topological edge-finding algorithm on the fused global high-precision binary mask image to extract closed vector polygons and generate exportable vector results. The system automatically reads the coordinate information file that accompanies the uploaded remote sensing image and, in conjunction with the corresponding six-parameter affine transformation model, directly maps the pixel plane coordinates of the mask to the real geospatial coordinates. Subsequently, the vector polygons with the real projection are packaged with the corresponding natural language semantic tags and directly output as a standard file format, forming a complete business closed loop from "language input" to "professional surveying and mapping results output".

Images