A remote sensing land use map making method based on SAM3

Abstract

The application discloses a remote sensing land use map making method based on SAM3, and relates to the field of remote sensing image intelligent interpretation, and comprises the following steps: acquiring high-resolution remote sensing images of a research area, constructing a prompt word set containing multiple first-level prompt words and corresponding second-level prompt words according to land use / land cover classification requirements of the research area, performing multi-scale block processing of the input images through round-by-round incremental grid division, performing zero-sample segmentation on each sub-block through each prompt word in the second-level prompt word set, performing spatial reorganization and integration on the segmentation results to obtain a three-dimensional response tensor, counting the cumulative response times of any pixel point in the tensor to the corresponding categories of the first-level prompt words in all rounds to form a pixel-level multi-round response matrix, introducing a competition judgment mechanism based on category recognition confidence weighting to realize the judgment of the unique category attribution of each pixel, and thus generating a land cover map with high precision, high stability and strong interpretability.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation technology of remote sensing images, and in particular to a method for creating remote sensing land use maps based on SAM3. Background Technology

[0002] With the rapid development of remote sensing technology, high-resolution satellite and aerial imagery have been widely used in land use / land cover monitoring, urban planning, agricultural management, and ecological environment assessment. Land cover information, as a crucial piece of fundamental geographic information, directly impacts the analytical results and decision-making reliability of related operational systems through its accuracy and update efficiency.

[0003] Traditional land cover mapping methods for remote sensing images mainly rely on manual interpretation or rule-driven classification algorithms based on low-level features such as spectral and textural characteristics. These methods typically require high levels of human intervention, have low processing efficiency, and are poorly adaptable to different regions and imaging conditions, making them difficult to meet the practical application requirements of large-scale, high-frequency updates.

[0004] In recent years, deep learning methods, represented by Convolutional Neural Networks (CNNs), have made some progress in remote sensing image land cover identification and classification tasks. These methods automatically extract image features through end-to-end supervised learning, achieving high classification accuracy on specific datasets. However, supervised learning models such as CNNs typically rely on a large number of high-precision, pixel-level labeled samples for training. The labeling process for high-resolution remote sensing images is highly specialized, costly, and time-consuming. Especially in scenarios with complex land cover types, blurred boundaries, and significant scale differences, high-quality labeled data is difficult to obtain, which to some extent limits the promotion and application of deep learning methods based on the CNN framework in practical LULC mapping.

[0005] With the development of Vision Foundation Models (VFMs), models pre-trained on large-scale natural images have demonstrated strong generalization capabilities under zero-shot or few-shot conditions. SegmentAnything Model 3 (SAM3) can segment target regions through prompts without additional training samples, providing a new technical approach for the automated identification of ground features in remote sensing images.

[0006] However, when applying large visual models such as SAM3 directly to land use / land cover mapping of high-resolution remote sensing images in existing technologies, the following shortcomings still exist: (1) Semantic boundary shifting: In zero-sample scenarios, when using semantically similar cue words (such as "trees" and "forest", "water bodies" and "lakes") to segment the same land feature, the mask boundaries generated by SAM3 often show significant differences. This phenomenon of non-overlapping response areas under synonyms causes the land feature boundaries to fluctuate drastically with the fine-tuning of the cue word description, making it difficult to determine a unified mapping boundary. (2) Lack of multi-scale consistency: In a global perspective, the model tends to output contiguous and holistic segmentation results based on macroscopic semantic cues, such as contiguous forests and continuous built-up areas; while in a local perspective, the model can usually utilize richer detailed information to more fully identify and locate scattered targets and small patches of land features more precisely. (3) Lack of mutual exclusion discrimination: Remote sensing mapping requires unique pixel classification, but SAM3 lacks mutual exclusion constraints between categories. Even semantically dissimilar non-synonymous categories (such as "road" and "concrete building," "bare land" and "desertified grassland") often activate identical image regions due to similar texture features. This spatial overlap conflict across multiple categories makes directly generating pixel-level unique classification maps extremely challenging.

[0007] Therefore, how to improve SAM3's multi-class discrimination capability, scale adaptability, and pixel-level classification stability in high-resolution remote sensing image land cover mapping while maintaining its zero-sample advantage has become an urgent technical problem to be solved. Summary of the Invention

[0008] To address the problems encountered when applying existing large-scale visual models such as SAM3 directly to the production of land use / land cover maps from high-resolution remote sensing images under zero-sample conditions, this invention provides a method for producing remote sensing land use maps based on SAM3. This method constructs a multi-level cue word system and introduces a multi-round, multi-scale cue word detection strategy to fully statistically analyze the responses of complex features in remote sensing images. Furthermore, for pixel-level multi-class confusion areas, a weighted competitive judgment mechanism is proposed to achieve a unique category assignment for each pixel, thereby generating a high-precision, highly stable, and highly interpretable land cover map.

[0009] This invention adopts the following technical solution: a method for creating remote sensing land use maps based on SAM3, comprising the following steps:

[0010] S1. Remote sensing image acquisition and preprocessing: Acquire high-resolution remote sensing images of the study area and convert them into image data containing only RGB information, which will be used as input data for subsequent analysis.

[0011] S2. Construction of Level 1 Prompt Word System: Based on the land use / land cover classification requirements of the study area, a prompt word set containing multiple level 1 prompt words and corresponding level 2 prompt words is constructed. The level 1 prompt words are used to represent macroscopic land cover categories, and the level 2 prompt words are used to describe specific land cover morphological characteristics.

[0012] S3. SAM3 Zero-Shot Segmentation Processing Based on Multi-Round Multi-Scale Grid Partitioning: The input image data is processed by multi-scale block partitioning with progressively increasing grids. Each sub-block is segmented using zero-shot segmentation of each cue word in the constructed set of secondary cue words. The segmentation results are spatially reorganized and integrated to obtain a three-dimensional response tensor, which records the response of each pixel to each secondary cue word at different spatial scales and corresponding rounds.

[0013] S4. Pixel-level multi-round response statistics: Based on step S3, the segmentation results of all rounds are accumulated and statistically analyzed. For any pixel in the three-dimensional response tensor, the cumulative number of responses to each first-level prompt word corresponding to each category is counted in all rounds to form a pixel-level multi-round response matrix.

[0014] S5. Confusion Region Identification and Weighted Competition Decision: A competition decision mechanism based on category recognition confidence weighting is introduced to determine the unique category of each pixel;

[0015] S6. Land cover map generation: Perform the determination process of steps S3 to S5 point by point on the input image pixels to generate a unique land cover category identifier for each pixel, forming the final land cover classification result map.

[0016] As an alternative, in step S3, in order to achieve effective detection of ground features at different scales and fully stimulate the feature recognition capabilities of the Visual Large Model (SAM3) at different spatial granularities, a multi-scale block processing with progressively increasing grid division is adopted to perform zero-sample segmentation on the input remote sensing image.

[0017] The multi-scale block processing of the progressively increasing grid division includes the following sub-steps:

[0018] S3.1, Multi-round incremental mesh generation:

[0019] Assuming the input remote sensing image size is H×W and the total number of processing rounds is R, in the r-th round of processing, the original image is uniformly divided into r×r rectangular sub-blocks in the horizontal and vertical directions;

[0020] Adjacent sub-blocks do not overlap, and all sub-blocks collectively cover the entire original image area. The size of each sub-block is:

[0021] ;

[0022] Where H and W represent the length and width of the original image data, and r represents the round.

[0023] To ensure that each sub-block image contains sufficient information for effective recognition by the large visual model, a progressive multi-scale processing is performed, from global coarse-grained analysis to local fine-grained focusing. As the round r increases, the sub-block size gradually decreases. A minimum side length threshold for rectangular sub-blocks is set. If the calculated sub-block size is less than the threshold, the mesh generation is terminated.

[0024] S3.2, Zero-sample segmentation of sub-block images:

[0025] In each round of processing, all generated sub-block images are sequentially input into the SAM3 model. For each sub-block, zero-sample segmentation is performed using each cue word in the set of secondary cue words constructed in step S2.

[0026] For the current sub-block and the current prompt word, the SAM3 model outputs a binary segmentation mask, where a mask value of 1 indicates that the current pixel responds to the current prompt word, and a mask value of 0 indicates that the current pixel does not respond to the current prompt word;

[0027] S3.3 Spatial Reorganization of Segmentation Results:

[0028] To facilitate subsequent pixel-level statistics across the entire image, the segmentation results of all sub-blocks in each round are stitched together according to their spatial positions in the original image, and recombined into a full-image segmentation response map with the same size H×W as the original image. The value of each pixel in this response map indicates whether it was detected in the current round and under the current prompt word.

[0029] S3.4 Integration of results from multiple rounds and multiple prompts:

[0030] Repeat steps S3.1 to S3.3 until all R rounds of processing are completed. For each secondary prompt word, obtain R full-image response maps, which form a three-dimensional response tensor with dimensions H×W×R.

[0031] This tensor system records the response of each pixel to each specific prompt word at different spatial scales and corresponding rounds, providing multi-scale and multi-round response evidence for subsequent pixel-level statistics and competitive decision-making.

[0032] As a preferred option, in step S4, the cumulative number of responses reflects the frequency with which the pixel is identified by the same category in blocks at different scales and locations. The higher the frequency, the more stable and consistent the response of the corresponding pixel under the corresponding category, which can be used as preliminary evidence that the pixel belongs to the category.

[0033] Set of primary prompt words Represented as:

[0034] ;

[0035] Where M represents the total number of categories, Indicates the first Category 1 prompt words, ;

[0036] In the In the second-level prompt word detection, the category c corresponding to each pixel (i, j) is defined. k The response is :

[0037] ;

[0038] Sum the results of all R rounds to obtain the cumulative response count for each pixel to each category:

[0039] .

[0040] The cumulative number of responses for all pixels A three-dimensional, pixel-level, multi-round response matrix with dimensions H×W×M is constructed. In this matrix, each element reflects the total frequency with which the corresponding pixel is identified by a certain category in the multi-round, multi-scale analysis, serving as the preliminary evidence strength for classifying the pixel as belonging to the corresponding category.

[0041] As a preferred approach, in step S5, since the distinguishability of different land cover categories varies in high-resolution remote sensing images, some categories (such as water bodies and main roads) typically exhibit relatively stable responses in multiple rounds of segmentation, while other categories (such as vegetation and bare land, and different building types) are prone to response fluctuations due to similar spectra and textures. Directly relying on the cumulative number of responses for majority decision-making may lead to insufficient reliability in decisions regarding easily confused categories.

[0042] To address the aforementioned issues, this step introduces a competitive decision-making mechanism based on category recognition confidence weighting, specifically including:

[0043] S5.1, Confusing pixel identification: Based on the pixel-level multi-round response matrix, identify pixels with at least two categories whose cumulative response count is greater than zero, and mark them as confusing pixels to be judged;

[0044] S5.2 Weight setting and basis: Assign a weight coefficient to each land cover category to form a weight set.

[0045] The weighting coefficients comprehensively reflect the identification stability and scene importance of the category, and their settings can be determined based on the visualization analysis of intermediate results of multiple rounds of response and knowledge of remote sensing interpretation.

[0046] For example, categories with high consistency and clear boundaries can be assigned higher weights, while categories that are easily disturbed and have large fluctuations in response can be assigned relatively lower weights.

[0047] S5.3 Weighted Score Calculation: For each pixel to be confused, calculate its weighted score for each response category:

[0048] ;

[0049] in, For category The weight, For the category of this pixel pair The cumulative number of responses;

[0050] S5.4 Competitive Decision: Compare the weighted scores of all response categories of a pixel, and determine the category with the highest score as the final land cover category for that pixel:

[0051] ;

[0052] S5.5 Direct classification of non-confused pixels: For pixels that respond to only a single category, they are directly classified into that category;

[0053] S5.6 If a pixel is not detected by any category in all rounds, it is marked as unclassified.

[0054] The present invention also provides: an electronic device, comprising:

[0055] One or more processors;

[0056] A storage device on which one or more programs are stored;

[0057] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-described methods for creating remote sensing land use maps based on SAM3.

[0058] The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in any of the above-mentioned methods for creating remote sensing land use maps based on SAM3.

[0059] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0060] 1. Improve the stability of multi-class segmentation: The method of this invention significantly reduces the randomness of single-round segmentation results through multi-round prompt word detection and pixel-level response statistics, thereby improving the stability and robustness of ground feature recognition.

[0061] 2. Effectively eliminates confused areas: To address the common problem of multi-class confusion in remote sensing images, the method of this invention introduces a weighted competition determination mechanism to achieve pixel-level unique class decision-making and avoid the problem of multiple class assignments for the same pixel.

[0062] 3. No training samples required, highly adaptable: The method of this invention is based on the zero-sample segmentation capability of SAM3, requiring no additional labeled data or model training, and is applicable to high-resolution remote sensing images of different regions and types.

[0063] 4. High interpretability of results: The pixel category determination of this invention is based on the number of rounds of response and weight calculation. The decision process is clear and easy to analyze and reproduce, which meets the requirements of engineering application and implementation. Attached Figure Description

[0064] Figure 1 This is a flowchart of the SAM3-based remote sensing land use map production method of the present invention.

[0065] Figure 2 This is a flowchart of the multi-level prompt word recognition process of the present invention;

[0066] Figure 3 This is a flowchart of the weighted competition process of the present invention;

[0067] Figure 4 This is a satellite image of the study area in an embodiment of the present invention;

[0068] Figure 5 These are schematic diagrams illustrating various recognition results in embodiments of the present invention;

[0069] Figure 6 This is a diagram illustrating the confusion region identified by the model in an embodiment of the present invention.

[0070] Figure 7 This is the final land cover map according to an embodiment of the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0072] In one embodiment of the present invention, a method for producing remote sensing land use maps based on SAM3 is provided. By performing multi-class PK (Pairwise Knockout) analysis of remote sensing images, pixel-level category competition relationships are quantified. By combining multi-scale processing strategies with the zero-shot recognition capability of SAM3, a complete analysis process from image input to competition result output is realized.

[0073] This embodiment selects the downstream area of ​​the Charshen Reservoir in Ulanhot City, Hinggan League, Inner Mongolia as the area to be studied. The method of this invention is described in detail below, with the overall process as follows: Figure 1 As shown, it includes the following steps:

[0074] 1. Study Area and Data Description

[0075] Satellite imagery of the study area, such as Figure 4 As shown, the EPSG:3857 projection coordinate system is used, and the coverage area is approximately 30.66 km². 2 The spatial resolution is 0.6 meters. The image size is 10496×8192 pixels, containing three bands (RGB), spanning approximately 6.27 km east to west and extending approximately 4.89 km north to south.

[0076] 2. Confirmation word construction

[0077] Based on image features, the primary cue words are set into five main categories: traffic, building, vegetation, water, and bare soil. Each primary cue word category corresponds to several secondary cue words to enhance SAM3's ability to perceive specific land features, as shown in Table 1.

[0078] Table 1: Prompt Word Settings

[0079]

[0080] 3. A zero-shot segmentation method for SAM3 based on multi-round, multi-scale grid generation

[0081] Because high-resolution remote sensing images typically have extremely wide swaths and significant differences in ground feature scale, directly segmenting the entire image or a single scale can easily lead to problems such as missed detection of small-scale features, incomplete segmentation of large-scale features, and local semantic misjudgments. Therefore, this embodiment employs a multi-scale block processing strategy with progressively increasing grid division to fully stimulate the visual large-scale model's response to ground feature features at different spatial scales. The specific process is as follows: Figure 2 As shown.

[0082] (1) Multi-round incremental grid partitioning strategy

[0083] The remote sensing image is input through the input module, with a size of H×W, and the entire processing is divided into R rounds. In this embodiment, R=15 is set.

[0084] In the r-th round of processing, the original remote sensing image is uniformly divided into r×r images, and the space of each sub-block is:

[0085] ;

[0086] Among them, adjacent sub-blocks do not overlap in space and together cover the entire image area.

[0087] To ensure that each sub-block image contains enough information for the large visual model to recognize effectively, this embodiment sets the minimum side length of the sub-block size to be no less than 512 pixels. If the sub-block size calculated according to the division formula is less than this threshold, the subsequent division at a finer scale will be terminated.

[0088] In this way, as the round number r increases, the size of the sub-block gradually decreases, thereby achieving a progressively refined analysis from the global scale to the local detail scale.

[0089] (2) Zero-sample segmentation based on prompt words

[0090] The classification module performs classification processing. In each round of processing, all sub-block images generated in that round are sequentially input into SAM3, and zero-sample segmentation is performed in combination with the set of secondary cue words constructed in step 2.

[0091] For each secondary cue word, SAM3 outputs a binary segmentation mask to represent the spatial response region of the cue word in the current sub-block, thus obtaining the recognition result for each category.

[0092] (3) Cumulative output of multi-round segmentation results

[0093] After completing the segmentation of all sub-blocks in the r-th round, the output module stitches together the segmentation masks of all sub-blocks obtained in this round according to their spatial positions in the original image, restoring a segmentation response map with the same size as the original image, and obtaining the category of a single pixel.

[0094] Repeat the above processing flow until all R=15 rounds of segmentation operations are completed, ultimately obtaining a set of pixel-level segmentation response results covering multiple scales, multiple rounds, and multiple cue words. The various recognition results after multiple rounds of processing are as follows: Figure 5 As shown.

[0095] 4. Pixel-level multi-round response statistics

[0096] In the process of land cover identification in remote sensing images, the results of single-round SAM3 cue word segmentation may have local errors or random fluctuations, especially in areas with blurred land cover boundaries and similar textures. A single-round response is difficult to reliably reflect the true land cover category of the pixel.

[0097] Therefore, this embodiment proposes a pixel-level multi-round response statistical method to perform cumulative quantitative analysis on the segmentation results of multi-round, multi-scale prompt words, thereby providing a reliable basis for subsequent confusion region determination and weighted competitive decision-making.

[0098] Let the image size be H×W, the number of prompt words in one category be M=5, and the category set C be:

[0099] ;

[0100] In the Rth round of secondary prompt word detection, the category c corresponding to each pixel (i, j) is defined. k The response is:

[0101] ;

[0102] Sum the results of all R rounds to obtain the cumulative response count for each pixel to each category:

[0103] ;

[0104] In this embodiment, the identified confusing regions are as follows: Figure 6 As shown.

[0105] 5. Determination of weighted competition in confused regions

[0106] In actual land cover identification using remote sensing imagery, multi-round cue word segmentation may cause some pixels to respond to multiple categories simultaneously, resulting in multi-class conflict. Directly classifying such confused areas can lead to unclear pixel attribution or unstable classification results.

[0107] To address this issue, this embodiment proposes a pixel-level decision-making strategy based on weighted competition, the process of which is as follows: Figure 3 As shown, this achieves a unique category determination for each pixel.

[0108] To reflect the differences in the recognition reliability of different land cover categories in remote sensing images and their relative importance in the land cover classification system, a weighting coefficient needs to be assigned to each category.

[0109] The weights are determined based on the following: After completing the pixel-level multi-round response statistics in step S4, the following can be generated: Figure 5 The images shown represent the preliminary identification results for each category. Visual analysis of these intermediate results reveals differences in the confidence levels for identification across different categories.

[0110] For example, in this embodiment, the categories "water bodies," "transportation facilities," and "built areas" typically have clear boundaries and unique spectral characteristics, exhibiting stable responses and low misclassification rates in multi-round segmentation, thus receiving higher weights (e.g., 2.5, 2.0). "Bare land" and "vegetation" suffer from spectral confusion in some areas, resulting in relatively low identification confidence, therefore receiving lower or medium weights (e.g., 1.0, 1.5), represented as follows:

[0111] ;

[0112] Pixel (i, j) for category c k The weighted score is:

[0113] ;

[0114] Then, by taking into account the number of responses in multiple rounds and the category priority through weighted scoring, the relative support of each pixel for each category is quantified.

[0115] For pixels with conflicting values, the category with the highest weighted score is selected as the final classification:

[0116] ;

[0117] If a pixel is not detected by any category, it is marked as unclassified.

[0118] 6. Generation of the final land cover map

[0119] By applying the weighted competition strategy described above to process each pixel of the entire image, a unique class assignment result is obtained. Generate high-confidence land cover maps, such as Figure 7 As shown.

[0120] As can be seen, the method of the present invention effectively reduces the confusion area, ensures that each pixel belongs to only one land cover category, and improves the stability and interpretability of classification.

[0121] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for creating remote sensing land use maps based on SAM3, characterized in that, Includes the following steps: S1. Remote sensing image acquisition and preprocessing: Acquire high-resolution remote sensing images of the study area and convert them into input image data containing only RGB information; S2. Construction of Level 1 Prompt Word System: Based on the land use / land cover classification requirements of the study area, a prompt word set containing multiple level 1 prompt words and corresponding level 2 prompt words is constructed. The level 1 prompt words are used to represent macroscopic land cover categories, and the level 2 prompt words are used to describe specific land cover morphological characteristics. S3. SAM3 zero-shot segmentation based on multi-round multi-scale grid partitioning: The input image data is processed by multi-scale block partitioning with progressively increasing grids. Each sub-block is segmented using zero-shot segmentation of each cue word in the set of secondary cue words. The segmentation results are spatially reorganized and integrated to obtain a three-dimensional response tensor, which records the response of each pixel to each secondary cue word at different spatial scales and corresponding rounds. The multi-scale block processing of the progressively increasing grid division includes the following sub-steps: S3.1 Multi-round incremental grid generation: The input remote sensing image size is H×W, the total number of processing rounds is R, and the grid generation is... In the round processing, the original image is uniformly divided into horizontal and vertical sections. A rectangular sub-block; S3.2 Zero-sample segmentation of sub-block images: In each round of processing, all generated sub-block images are sequentially input into the SAM3 model. For each sub-block, zero-sample segmentation is performed using each cue word in the secondary cue word set constructed in step S2. For the current sub-block and the current cue word, the SAM3 model outputs a binary segmentation mask. A mask value of 1 indicates that the current pixel responds to the current cue word, and a mask value of 0 indicates that the current pixel does not respond to the current cue word. S3.3 Spatial Reassembly of Segmentation Results: The segmentation results of all sub-blocks in each round are stitched together according to their spatial positions in the original image and reassembled into a full-image segmentation response map with the same size H×W as the original image. The value of each pixel in the map indicates whether it is detected in the current round and under the current prompt word. S3.4 Integration of results from multiple rounds and multiple prompts: Repeat steps S3.1 to S3.3 until all R rounds of processing are completed. For each secondary prompt, R full-image response maps are obtained, forming a three-dimensional response tensor with dimensions H×W×R. S4. Pixel-level multi-round response statistics: Accumulate and statistically analyze the segmentation results of all rounds. For any pixel in the three-dimensional response tensor, count the cumulative number of responses to each primary prompt word category in all rounds to form a pixel-level multi-round response matrix. The cumulative number of responses is calculated as follows: Set of primary prompt words Represented as: ； in, Represented by the total number of categories. Indicates the first Category 1 prompt words, ; In the In the second-level prompt word detection, the category c corresponding to each pixel (i, j) is... k The response is : ； The results of all R rounds of detection are summed to obtain the cumulative number of responses for each pixel to each category. : ； S5. Confusion Region Identification and Weighted Competition Decision: A competition decision mechanism based on category recognition confidence weighting is introduced to determine the unique category of each pixel; S6. Land cover map generation: Perform the judgment process of steps S3 to S5 point by point on the input pixels to generate a unique land cover category identifier for each pixel, forming the final land cover classification result map.

2. The method for creating remote sensing land use maps based on SAM3 according to claim 1, characterized in that, The multi-round incremental grid division ensures that adjacent sub-blocks do not overlap, and all sub-blocks collectively cover the entire original image area. The size of each sub-block is: ； Where H and W represent the length and width of the original image data, and r represents the round.

3. The method for creating remote sensing land use maps based on SAM3 according to claim 1, characterized in that, The multi-round incremental mesh generation also includes progressive multi-scale processing from global coarse-grained analysis to local fine-grained focusing; According to the round As the size of the sub-blocks increases, the size of the sub-blocks gradually decreases. A minimum side length threshold for rectangular sub-blocks is set. If the calculated sub-block size is less than the threshold, the mesh generation is terminated.

4. The method for creating remote sensing land use maps based on SAM3 according to claim 2, characterized in that, The cumulative response count reflects the frequency with which a pixel is identified by the same category in blocks at different scales and locations. The higher the frequency, the more stable and consistent the response of the corresponding pixel under the corresponding category. The cumulative number of responses for all pixels A three-dimensional pixel-level multi-round response matrix is ​​constructed, with dimensions H×W×M. Each element in the matrix reflects the total frequency of the corresponding pixel being identified by the corresponding category in the multi-round multi-scale analysis, serving as the preliminary evidence strength for the pixel's category attribution.

5. The method for creating remote sensing land use maps based on SAM3 according to claim 4, characterized in that, The competition determination mechanism based on category recognition confidence weighting specifically includes: S5.1, Confusing pixel identification: Based on the pixel-level multi-round response matrix, identify pixels with at least two categories whose cumulative response count is greater than zero, and mark them as confusing pixels to be judged; S5.2 Weight Setting and Basis: Assign weight coefficients to each land cover category to form a weight set; S5.3 Weighted Score Calculation: For each pixel to be confused, calculate its weighted score for each response category. : ； in, For category The weight, For the category of this pixel pair The cumulative number of responses; S5.4 Competitive Decision: Compare the weighted scores of all response categories of a pixel, and determine the category with the highest score as the final land cover category for that pixel: ； S5.5 Direct classification of non-confused pixels: For pixels that respond to only a single category, they are directly classified into that category; S5.6 If a pixel is not detected by any category in all rounds, it is marked as unclassified.

6. The method for creating remote sensing land use maps based on SAM3 according to claim 5, characterized in that, The weighting coefficients are used to reflect the recognition stability and scene importance of each category. They are determined based on the visualization analysis of intermediate results from multiple rounds of responses and knowledge from the remote sensing interpretation domain. Categories with higher recognition consistency and clearer boundaries are assigned higher weights.

7. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the SAM3-based remote sensing land use map production method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps in the method for creating a remote sensing land use map based on SAM3 as described in any one of claims 1 to 6.

Images