A high-precision land block segmentation method, device and medium based on satellite images

Abstract

This invention discloses a high-precision land parcel segmentation method, device, and medium based on satellite imagery. The method includes the following steps: after inputting remote sensing imagery, semantic segmentation based on MAE-UNet is used to generate a farmland probability map; instance segmentation based on FastSAM is used to generate a land parcel confidence map; and boundary extraction based on DiffusionEdge is used to generate an edge probability map. Subsequently, the edge probability map is aligned with the land parcel confidence map, and the land parcel confidence map is corrected using the land parcel confidence map. The corrected farmland probability map, boundary probability map, and aligned land parcel confidence map are then weighted and fused to obtain a fused probability map. The final segmentation map is obtained after normalization and binarization thresholding. This invention generates multi-scale semantic probability maps through a self-supervised pre-trained model and combines instance segmentation and boundary optimization algorithms to achieve dynamic fusion of multi-source results. It can accurately segment complex terrain and high-density small targets, achieving high-precision boundary extraction and efficient automated processing.

Description

Technical Field

[0001] This invention relates to the field of land parcel segmentation technology based on satellite imagery, and in particular to a high-precision land parcel segmentation method, device and medium based on satellite imagery. Background Technology

[0002] In the field of land resource management and remote sensing interpretation, accurate extraction of farmland information is one of the core tasks of intelligent remote sensing interpretation. However, traditional methods still face significant challenges when processing remote sensing images in complex scenarios. The primary problem is the high cost of manual annotation, which not only heavily relies on expert experience but also requires processing massive amounts of remote sensing data, resulting in overall low efficiency. Secondly, although multispectral satellite imagery (such as Sentinel-2) can support the extraction of farmland plots to some extent due to its rich spectral information, it is still difficult to accurately identify small plots less than 10 mu (approximately 1.65 acres), and errors are prone to occur in interpreting farmland with fragmented edges and complex shapes. Furthermore, the efficiency of multi-source data fusion is low; the plot boundaries output by different deep learning models (such as mainstream segmentation frameworks) exhibit significant geometric deviations, and there is currently a lack of a unified collaborative optimization mechanism to effectively integrate multiple results. Existing solutions have the following limitations:

[0003] First, instance segmentation models are poorly adapted to complex scenes. Mainstream models (such as Mask R-CNN) require fully supervised training for specific categories, which limits their performance in low-resolution remote sensing images. Candidate regions generated by the anchor box mechanism often lose detailed information. At the same time, in areas with undulating terrain or mixed cropping, the binarization results output by the mask branch are difficult to accurately distinguish the boundaries of adjacent fields, leading to blurred edges and missegmentation.

[0004] Second, semantic segmentation models generally have low segmentation accuracy for small targets. The traditional U-Net architecture is limited by the local receptive field of convolution operations, making it difficult to effectively capture cross-regional field boundaries, resulting in boundary breaks or discontinuities in the segmentation results. Its skip connection mechanism only fuses multi-scale features through simple splicing, failing to fully integrate high-level semantics and low-level details, resulting in the loss of key features. In addition, the model has limited discrimination ability in complex scenes (such as at the boundaries of adjacent fields), making it difficult to achieve accurate differentiation.

[0005] Third, SAM (SegmentAnythingModel) exhibits significant robustness limitations in end-to-end segmentation tasks. Its segmentation results heavily rely on human cues (such as points, boxes, and text), but in complex farmland remote sensing scenes, the quality of these cues is easily affected by interference, the selected points tend to deviate from the true edges, and it is extremely sensitive to changes in image quality, still requiring manual verification. Furthermore, the model has poor adaptability to small targets and fragmented, edged fields: although the global attention mechanism can capture large-scale semantic features, it lacks sufficient response to small fields smaller than 10 acres or farmland fragmented by roads and irrigation ditches, exhibiting limited detail resolution. Experiments show that SAM's average accuracy (APs) in small target segmentation tasks is significantly lower than dedicated segmentation models. Simultaneously, this model is highly dependent on the distribution of pre-training data, facing significant domain differences when transferred to agricultural remote sensing. Its training set mainly consists of natural scene objects (such as people, vehicles, and animals), which have significant semantic differences from farmland scenes (such as field ridges, crops, and ditches), requiring extensive fine-tuning with labeled samples, resulting in high computational and labeling costs.

[0006] Therefore, a high-precision land parcel segmentation method based on satellite imagery is needed to overcome the limitations of existing segmentation methods. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a high-precision land parcel segmentation method, device, and medium based on satellite imagery, thereby solving the problems of existing land parcel segmentation methods in farmland remote sensing scenarios, such as reliance on manual prompts, poor adaptability to small targets and fragmented edge fields, and dependence on pre-training data distribution.

[0008] This invention provides a method, device, and medium for high-precision land parcel segmentation based on satellite imagery.

[0009] First aspect: A high-precision land parcel segmentation method based on satellite imagery, including:

[0010] S1. Input remote sensing image;

[0011] S2. Perform semantic segmentation based on MAE-UNet to generate a semantically segmented farmland probability map;

[0012] S3. Perform instance segmentation based on FastSAM and generate a confidence map of the segmented land parcels;

[0013] S4. Extract boundaries based on DiffusionEdge and generate an edge probability map;

[0014] S5. Align the land parcel confidence map based on the edge probability map, and correct the cultivated land probability map based on the land parcel confidence map;

[0015] S6. Perform weighted fusion of the corrected farmland probability map, boundary probability map and aligned plot confidence map to obtain the fused probability map;

[0016] S7. Normalize and binarize the fusion probability map to obtain the final plot segmentation map.

[0017] In one embodiment of the present invention, in S2, MAE-UNet uses a pre-trained MAE model as the encoder and UNet as the decoder. A multi-head self-attention mechanism is introduced in UNet to dynamically focus on the key features of farmland boundaries and farmland texture.

[0018] In one embodiment of the present invention, the DiffusionEdge in S3 includes,

[0019] Increasing noise is introduced to learn the fuzzy distribution of farmland boundaries during backpropagation and generate an edge probability map.

[0020] Based on the edge probability map, an adaptive threshold segmentation strategy is adopted to dynamically adjust the boundary extraction accuracy according to the local gradient intensity and suppress vegetation texture noise.

[0021] By combining a multi-level pyramid network (FPN), the macro and micro boundaries of cultivated land and plots are synergistically optimized to maintain the integrity of features at different scales.

[0022] In one embodiment of the present invention, step S5, which aligns the land parcel confidence map based on the edge probability map, includes:

[0023] The confidence map of the plot segmented by FastSAM is aligned with the marginal probability map extracted by DiffusionEdge using Hausdorff distance constraints to eliminate plot boundary breaks or misalignments. The formula is as follows:

[0024]

[0025] in, Let d(a,B) be the Hausdorff distance, d(a,B) be the shortest distance from point a to set B, and d(b,A) be the shortest distance from point b to set A. If the value is greater than 3 pixels, then the boundary of the plot confidence map is fine-tuned using the DiffusionEdge edge probability map boundary.

[0026] In one embodiment of the present invention, step S5, which corrects the arable land probability map based on the land parcel confidence map, includes:

[0027] Based on spectral consistency, the class labels of the plot confidence map segmented by FastSAM instances and the farmland probability map segmented by MAE-UNet RGB three-channel semantic segmentation are corrected to correct the class label confusion in conflict areas.

[0028] In one embodiment of the present invention, the spectral consistency correction is performed by calculating the similarity of the RGB three-channel histograms in the conflict region and then performing spectral consistency correction based on the similarity, as expressed by the formula:

[0029]

[0030] in, This is a histogram of the probability map of arable land. Create a confidence histogram for the land parcel.

[0031] In one embodiment of the present invention, the weighted fusion of the corrected farmland probability map, the boundary probability map, and the aligned plot confidence map is expressed by the following formula:

[0032]

[0033] Where α, β, and γ are weight values. To correct the arable land probability map, This is a marginal probability map. To align the confidence map of the land parcels.

[0034] In one embodiment of the present invention, the normalization and binarization thresholding of the fusion probability map in step S7 includes:

[0035] Sigmoid normalization, expressed by the formula:

[0036]

[0037] In the binarization thresholding process, the threshold is set to 0.5.

[0038] Second aspect: An electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method provided in the first aspect.

[0039] Third aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0040] The beneficial effects of this invention are:

[0041] 1. This invention generates multi-scale semantic probability maps through a self-supervised pre-trained model, combines an instance segmentation model with a boundary optimization algorithm, drives the dynamic fusion of multi-source results, and achieves accurate segmentation of high-density small targets. It is applicable to fields such as natural resource surveys, land use monitoring, smart agriculture, and urban planning, and is suitable for complex terrains (such as mountain terraces and low-lying wetlands) and high-density small targets in remote sensing imagery. To meet the automatic labeling requirements of 3 acres, high-precision boundary extraction and efficient automated processing are achieved through multi-scale collaborative segmentation and boundary optimization algorithms.

[0042] 2. This invention enhances the generalization ability of the model through MAE self-supervised pre-training, and effectively solves the problems of missed detection and boundary ambiguity of traditional algorithms for small-scale, fragmented farmland by combining cross-scale feature fusion and DiffusionEdge boundary optimization mechanism.

[0043] 3. This invention combines FastSAM lightweight instance segmentation with a multi-source result collaborative optimization strategy, significantly reducing computational costs. Single-image processing time is shortened, while GPU memory usage is reduced, balancing high precision and real-time requirements. It can meet the efficient processing needs of engineering scenarios such as surveys of tens of thousands of acres of farmland and large-scale land use monitoring.

[0044] 4. This invention significantly improves the robustness of the model in complex interference scenarios such as low light, shadow interference, and vegetation cover by using spectral feature consistency correction, dynamic threshold segmentation, and gradient enhancement algorithms. It also reduces the false detection rate and improves the boundary recall rate for small-scale farmland. Attached Figure Description

[0045] Figure 1 This is a schematic flowchart of the method of the present invention;

[0046] Figure 2 This is a flowchart illustrating the principle of the method of the present invention;

[0047] Figure 3 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0048] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0049] Existing solutions suffer from several problems: instance segmentation models have poor adaptability to complex scenes and limited performance in low-resolution remote sensing images; semantic segmentation models have low segmentation accuracy for small targets, making it difficult to capture cross-regional field boundaries and prone to boundary breakage; and SAM models lack robustness and adaptability, with end-to-end segmentation heavily reliant on manual prompts, and the quality of prompts in farmland scenes is easily affected by interference and sensitive to image quality, requiring manual verification.

[0050] Example 1: To address the above problems, this example discloses a high-precision land parcel segmentation method based on satellite imagery, such as... Figure 1 and Figure 2 As shown, the method includes:

[0051] S1. Input remote sensing image.

[0052] Preprocessing of remote sensing images includes radiometric correction, geometric correction, and atmospheric correction. Radiometric correction eliminates errors inherent in the sensor, ensuring the accuracy of image radiometric values; geometric correction corrects geometric distortions caused by factors such as satellite attitude and terrain undulations, resulting in higher positional accuracy; atmospheric correction removes the influence of the atmosphere on electromagnetic wave propagation, restoring the true reflectivity of ground features.

[0053] The quality and usability of the pre-processed remote sensing images are significantly improved, providing a more reliable data foundation for subsequent land parcel segmentation.

[0054] S2. Perform semantic segmentation based on MAE-UNet to generate a semantically segmented farmland probability map.

[0055] In MAE-UNet, the MAE (Masked Auto Encoder) module is used as the encoder. MAE adopts self-supervised pre-training and is self-supervised pre-trained on medium and high resolution (0.5m~2m) RGB remote sensing images. It learns the general morphological features of farmland (such as farmland outline and terrace texture) through the task of reconstructing masked image patches, thereby improving the model's generalization ability to unlabeled data.

[0056] MAE self-supervised pre-training includes a pre-training phase and a fine-tuning phase.

[0057] During the pre-training phase, a self-supervised learning framework was used to pre-train the ViT-Base encoder on 100,000 unlabeled 512×512 satellite images with a mask ratio of 70%, focusing on learning farmland morphological features (such as field edges and texture distribution).

[0058] During the fine-tuning phase, the underlying parameters of ViT are frozen, and a cross-scale feature interaction module (Cross-Scale Attention) is introduced into the multi-scale segmentation network to fuse the attention of 16×16 and 32×32 scale feature maps, thereby improving the sensitivity to small-scale farmland.

[0059] In the UNet module, its multi-layer Transformer feature extraction capability is retained to enhance the global semantic understanding of farmland plots. As the decoder part, UNet adopts an attention mechanism to enhance the multi-head self-attention module, which dynamically focuses on the key features of farmland boundaries and farmland texture, suppresses background noise interference (such as vegetation texture and shadow), and fuses high and low layer features through a cross-attention mechanism to improve edge sharpness and segmentation accuracy.

[0060] S3. Perform instance segmentation based on FastSAM and generate a confidence map of the segmented land parcels.

[0061] During the FastSAM training phase, a boundary-aware loss function is added to improve the continuity and geometric accuracy of field edges.

[0062] This loss function calculates the pixel-level difference between the predicted and actual boundaries, and, combined with edge gradient direction consistency constraints, applies higher penalty weights to areas of breakage and misalignment in the field outlines, prompting the model to learn boundary features that better match the real terrain. Simultaneously, to address the weak distinguishability between farmland and background (such as roads and water bodies) in remote sensing imagery, a spectral feature enhancement module is introduced in the FastSAM feature extraction stage. This module fuses the band information and texture features of the RGB imagery, strengthening the intra-class consistency and inter-class differences of farmland plots.

[0063] In the inference phase, the preprocessed remote sensing image is input. FastSAM first extracts multi-scale feature maps through a lightweight ViT backbone network and uses a dynamic anchor box generation mechanism to adapt candidate regions of different sizes, avoiding the problem of missed detection of small targets caused by the fixed size of traditional anchor boxes. Subsequently, a mask generation head performs fine segmentation of the candidate regions, outputting a binary mask and corresponding confidence score for each plot. The confidence score comprehensively considers pixel classification accuracy, boundary smoothness, and region integrity; a higher score indicates stronger reliability of the segmentation result. Finally, the confidence scores of all plots are mapped to grayscale images to generate the plot confidence map P3 for instance segmentation, providing instance-level reliability for subsequent multi-source result fusion.

[0064] S4. Extract boundaries based on DiffusionEdge and generate an edge probability map.

[0065] A high-precision boundary extraction branch is constructed by introducing the DiffusionEdge model to optimize the probability diffusion model. The fuzzy distribution of farmland boundaries is learned through incremental noise introduction and backpropagation process to generate sharp and continuous edge maps. The DiffusionEdge model takes preprocessed remote sensing images as input and learns the probability distribution characteristics of farmland boundaries through iterative noise introduction and back denoising process.

[0066] During the training phase, Gaussian noise is randomly added to the original images to enable the model to learn the ability to recover clear boundaries from noisy images. At the same time, constraint loss is designed in combination with the geometric characteristics of farmland boundaries to reduce the generation of false boundaries caused by interference factors such as vegetation texture and shadows.

[0067] During inference, DiffusionEdge outputs a preliminary boundary probability map, which is then optimized using a dynamic threshold segmentation algorithm. This algorithm adaptively adjusts the threshold based on the gradient intensity differences in different regions of the image (e.g., lowering the threshold in high-contrast boundary areas such as terraces to preserve details, and raising the threshold in low-contrast areas covered by vegetation to suppress noise).

[0068] Meanwhile, the DiffusionEdge model combines a multi-level pyramid network (FPN) to perform multi-scale fusion of boundary features, stitching together boundary features at 1 / 8, 1 / 4, and 1 / 2 resolutions. This preserves the integrity of macroscopic boundaries such as field outlines while refining the clarity of microscopic boundaries such as field ridges and cracks. The resulting boundary probability map P2 shows that higher pixel values ​​indicate a greater probability that the location represents a true farmland boundary, providing accurate boundary references for subsequent geometric alignment.

[0069] S5. Align the land parcel confidence map based on the edge probability map, and correct the arable land probability map based on the land parcel confidence map.

[0070] Based on the edge probability map, the land parcel confidence map is aligned in geometric space. Using the boundary probability map P2 generated by DiffusionEdge as a benchmark, the Hausdorff distance between the FastSAM instance segmentation mask boundary and the benchmark boundary is calculated. This distance is used to measure the maximum difference between the two boundary sets, and the calculation formula is as follows:

[0071] Where A is the FastSAM segmentation boundary, B is the DiffusionEdge baseline boundary, and d(a,B) is the shortest distance from point a to set B. If the calculated Hausdorff distance is greater than 3 pixels, it indicates a significant offset in the boundary. In this case, the coordinate information of the baseline boundary is used to fine-tune the edge pixel positions of the FastSAM mask. Local pixel redistribution is used to correct boundary breaks and misalignments, ensuring the geometric consistency between the instance segmentation result and the true boundary.

[0072] Based on the land parcel confidence map, the arable land probability map is corrected using spectral feature consistency. The semantic probability map P1 generated by MAE-UNet is compared with the geometrically aligned FastSAM instance segmentation results to locate the regions where the category judgments conflict between the two results (such as regions that are semantically segmented as arable land but instance segmented as non-arable land).

[0073] For conflict regions, their RGB three-channel histogram features are extracted, and the similarity between the semantic segmentation region histogram H1 and the instance segmentation region histogram H2 is calculated. The formula is as follows:

[0074]

[0075] For example, if the similarity S < 0.6, it indicates that the regional spectral features are more consistent with the semantic segmentation results, and the category determination of P1 is adopted first; if S ≥ 0.6, the edge information of the boundary probability map P2 is combined for further verification, and the corrected semantic probability map P1' is finally output.

[0076] S6. Perform weighted fusion of the corrected farmland probability map, boundary probability map, and aligned plot confidence map to obtain the fused probability map.

[0077] First, prepare the input data for fusion, clarify the three types of input data for weighted fusion, and ensure that the data format and spatial resolution are consistent.

[0078] The corrected farmland probability map P1' is generated from the MAE-UNet semantic segmentation module after spectral feature consistency correction. The pixel value range is [0,1], reflecting the probability that the location belongs to farmland. The resolution is consistent with the original remote sensing image (0.5m~2m).

[0079] The boundary probability map P2 is generated from the DiffusionEdge boundary extraction module after dynamic thresholding and multi-scale edge enhancement. The pixel value range is [0,1], reflecting the probability that the location is a farmland boundary. The resolution matches that of P1'.

[0080] The aligned parcel confidence map P3, derived from the FastSAM instance segmentation module, was generated after geometric spatial alignment and Hausdorff distance constraint fine-tuning. Its pixel value range is [0,1], reflecting the reliability of the parcel region in the instance segmentation results. Its resolution is similar to the previous two types. Figure 1 To.

[0081] Then, based on the roles of the three types of input data, fixed weighting coefficients are set to prioritize the accuracy of semantic categories and boundary precision. The corrected farmland probability map P1' has a weight of α=0.4, focusing on providing the overall semantic judgment basis for farmland areas and ensuring the integrity of plot boundaries. The boundary probability map P2 has a weight of β=0.4, focusing on providing accurate plot edge location information and resolving boundary breakage and ambiguity issues. The aligned plot confidence map P3 has a weight of γ=0.2, focusing on verifying the reliability of instance segmentation results and assisting in correcting local misjudged areas.

[0082] Then, a weighted fusion calculation is performed, using a linear weighted summation formula to fuse the three types of probability maps pixel by pixel. The formula is as follows:

[0083]

[0084] in, To obtain the fusion probability map, substitute the corresponding P1', P2, P3 values ​​and weight coefficients to calculate the fusion probability value at that position, which ranges from [0,1].

[0085] S7. Normalize and binarize the fusion probability map to obtain the final plot segmentation map.

[0086] To ensure that the numerical distribution of the fused probability map meets the requirements of subsequent binarization processing, the following steps are taken: Perform sigmoid normalization to further compress and optimize the pixel value range. The formula is as follows:

[0087]

[0088] Where x is the original probability value after fusion, and σ(x) is the normalized probability value, ensuring that the output result is reasonably distributed in the interval [0,1], and avoiding extreme values ​​from interfering with subsequent segmentation.

[0089] Binarization thresholding is performed using a pixel-by-pixel traversal method with a threshold of 0.5. The threshold is applied to the normalized fused probability map for classification, generating a binary segmentation map. The binarization rule is as follows:

[0090]

[0091] Where B(x) is the pixel value at x in the binarized segmentation image, where 1 represents farmland and 0 represents background.

[0092] After the above processing, the final farmland plot segmentation map is generated. This map is presented in the form of a binarized image, clearly distinguishing farmland plots from the background area, while possessing high-precision boundary contours and complete plot shapes. It can be directly used in practical application scenarios such as land use statistics, agricultural resource surveys, and smart agriculture plot rights confirmation.

[0093] Application examples:

[0094] Input: A satellite image containing 200 fields (0.5m resolution), with target fields ranging in size from 5 to 20 mu (approximately 0.33-0.33 hectares), and the terrain including plain farmland and mountain terraces.

[0095] The processing steps include:

[0096] Semantic segmentation: MAE-UNet generates a semantic probability map, initially detecting 215 candidate regions.

[0097] Instance segmentation: FastSAM generates 230 instance masks, but there is a boundary breakage problem.

[0098] Boundary optimization: DiffusionEdge extracts 208 high-precision edges, and corrects 12 offset boundaries through geometric alignment.

[0099] Spectral correction: Eliminates 8 misjudged areas (such as water bodies being misjudged as farmland).

[0100] Merged output: After weighted merging, 198 reliable plots are retained, and the final counting accuracy reaches 99.0% (198 / 200).

[0101] Efficiency improvement: Single image processing time has been reduced from 12 minutes to 45 seconds, a 16-fold improvement.

[0102] Accuracy improvement: On the test set, small targets ( The accuracy rate of the method (3 mu) reached 82.7%, which is 12.3 percentage points higher than the traditional method.

[0103] Resource optimization: The two-stage framework reduces GPU memory usage by 47%.

[0104] The present invention also provides an electronic device, Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 3 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory, for example, to execute the following method:

[0105] S1. Input remote sensing image;

[0106] S2. Perform semantic segmentation based on MAE-UNet to generate a semantically segmented farmland probability map;

[0107] S3. Perform instance segmentation based on FastSAM and generate a confidence map of the segmented land parcels;

[0108] S4. Extract boundaries based on DiffusionEdge and generate an edge probability map;

[0109] S5. Align the land parcel confidence map based on the edge probability map, and correct the cultivated land probability map based on the land parcel confidence map;

[0110] S6. Perform weighted fusion of the corrected farmland probability map, boundary probability map and aligned plot confidence map to obtain the fused probability map;

[0111] S7. Normalize and binarize the fusion probability map to obtain the final plot segmentation map.

[0112] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0114] S1. Input remote sensing image;

[0115] S2. Perform semantic segmentation based on MAE-UNet to generate a semantically segmented farmland probability map;

[0116] S3. Perform instance segmentation based on FastSAM and generate a confidence map of the segmented land parcels;

[0117] S4. Extract boundaries based on DiffusionEdge and generate an edge probability map;

[0118] S5. Align the land parcel confidence map based on the edge probability map, and correct the cultivated land probability map based on the land parcel confidence map;

[0119] S6. Perform weighted fusion of the corrected farmland probability map, boundary probability map and aligned plot confidence map to obtain the fused probability map;

[0120] S7. Normalize and binarize the fusion probability map to obtain the final plot segmentation map.

[0121] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0122] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-precision land parcel segmentation method based on satellite imagery, characterized in that, include: S1. Input remote sensing image; S2. Perform semantic segmentation based on MAE-UNet to generate a semantically segmented farmland probability map; S3. Perform instance segmentation based on FastSAM and generate a confidence map of the segmented land parcels; S4. Extract boundaries based on DiffusionEdge and generate an edge probability map; S5. Align the land parcel confidence map based on the edge probability map, and correct the cultivated land probability map based on the land parcel confidence map; S6. Perform weighted fusion of the corrected farmland probability map, boundary probability map and aligned plot confidence map to obtain the fused probability map; S7. Normalize and binarize the fusion probability map to obtain the final plot segmentation map; In S2, MAE-UNet uses a pre-trained MAE model as the encoder and UNet as the decoder. A multi-head self-attention mechanism is introduced into UNet to dynamically focus on the key features of farmland boundaries and farmland texture. The S5 section, which aligns the land parcel confidence map based on the edge probability map, includes: The confidence map of the plot segmented by FastSAM is aligned with the marginal probability map extracted by DiffusionEdge using Hausdorff distance constraints to eliminate plot boundary breaks or misalignments. The formula is as follows: ； in, Let d(a,B) be the Hausdorff distance, d(a,B) be the shortest distance from point a to set B, and d(b,A) be the shortest distance from point b to set A. If the value is greater than 3 pixels, then the boundary of the plot confidence map is fine-tuned using the DiffusionEdge edge probability map boundary. The S5 step of correcting the farmland probability map based on the land parcel confidence map includes: Based on spectral consistency, the class labels of the plot confidence map segmented by FastSAM instances and the farmland probability map segmented by MAE-UNet RGB three-channel semantic segmentation are corrected to correct the class label confusion in conflict areas.

2. The method according to claim 1, characterized in that, The DiffusionEdge in S3 includes: Increasing noise is introduced to learn the fuzzy distribution of farmland boundaries during backpropagation and generate an edge probability map. Based on the edge probability map, an adaptive threshold segmentation strategy is adopted to dynamically adjust the boundary extraction accuracy according to the local gradient intensity and suppress vegetation texture noise. By combining a multi-level pyramid network, the macro and micro boundaries of arable land and plots are synergistically optimized to maintain the integrity of features at different scales.

3. The method according to claim 1, characterized in that, The spectral consistency correction involves calculating the similarity of the RGB three-channel histograms in conflict regions and performing spectral consistency correction based on this similarity. The formula is as follows: ； in, This is a histogram of the probability map of arable land. Create a confidence histogram for the land parcel.

4. The method according to claim 1, characterized in that, The weighted fusion of the corrected farmland probability map, boundary probability map, and aligned plot confidence map is expressed by the following formula: ； Where α, β, and γ are weight values. To correct the arable land probability map, This is a marginal probability map. To align the confidence map of the land parcels.

5. The method according to claim 1, characterized in that, The normalization and binarization thresholding of the fusion probability map in S7 includes: Sigmoid normalization, expressed by the formula: ； In the binarization thresholding process, the threshold is set to 0.

5.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 5.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.

Images