Remote sensing image segmentation method and system based on land utilization

Abstract

The invention relates to a remote sensing image segmentation method and system based on land utilization, and the method comprises the steps: collecting continuous periodic remote sensing images of the same region, marking a land type and a seasonal factor vector, and determining whether to carry out zoning or not according to a latitude span; then identifying the boundary of the miniature land parcel, calculating the boundary identification degree, delimiting a protection area, respectively calculating the seasonal factor vectors of the non-protection area and the protection area, and fusing the seasonal factor vectors into a global seasonal factor; secondly, preliminarily segmenting the image by adopting a fuzzy C-means clustering model in combination with a global seasonal factor, and finally carrying out image re-segmentation and verifying the accuracy in combination with a plot seasonal feature and boundary recognition degree correction model; according to the method, through operations such as fine seasonal factor vector construction, latitude zoning, protection area delimiting and multi-feature fusion, the shadow area land type classification accuracy is effectively improved, the influence of seasonal variation and boundary blur in a complex earth surface scene can be reduced, and the robustness and adaptability of the model are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image segmentation in the technical field of image processing, and in particular to a remote sensing image segmentation method and system based on land use. BACKGROUND

[0002] In the field of remote sensing image processing, it is crucial to accurately segment images to obtain land use information. Traditional remote sensing image segmentation methods often only perform simple classification based on the spectral features of the images, and cannot fully consider the seasonal changes of the land and the boundary features of the land parcels. For example, the image features of the same land type in different seasons may have large differences, which will cause the traditional method to missegment. Moreover, the recognition and processing of land parcel boundaries are not fine enough, which cannot effectively distinguish different land parcels, affecting the accuracy of segmentation and making it difficult to meet the demand for precise analysis based on land use.

[0003] With the development of remote sensing technology, it has become possible to obtain continuous period remote sensing images of the same area, which provides a data basis for more accurate image segmentation. However, existing segmentation techniques do not fully utilize the seasonal information and land parcel boundary information in these continuous period images. At the same time, when processing large-scale remote sensing images, the influence of latitude span on image features is not considered, and there is a lack of targeted strategies such as zonation processing. SUMMARY

[0004] The present application provides a remote sensing image segmentation method that can integrate continuous period image information, consider latitude factors, effectively utilize land parcel boundary features, and combine seasonal factors to improve the accuracy and practicality of segmentation, addressing the technical problems in the prior art.

[0005] The technical solution of the present application to solve the above technical problems is as follows: a remote sensing image segmentation method based on land use, comprising: S1, collecting continuous period remote sensing images of the same area, labeling and recording the real land type and period seasonal factor vector of each region; S2, determining whether to zonate according to the latitude span of the real-time remote sensing image; S3, identifying the micro-land parcel boundary after zonation processing, calculating the boundary recognition degree and delineating the protection zone; S4, calculating the non-protection zone single seasonal factor vector and the protection zone adjusted seasonal factor vector, and fusing them into a global seasonal factor; S5, using a fuzzy C-means clustering model combined with the global seasonal factor to preliminarily cluster and segment the image into multiple layers of pixel points; S6, combining the micro-land parcel seasonal characteristics and boundary recognition degree to correct the fuzzy C-means clustering model; S7, image re-segmentation is performed on the target region based on the corrected fuzzy C-means clustering model to verify the segmentation accuracy.

[0006] Preferably, S2 includes determining whether to divide the bands according to the latitude span of the real-time remote sensing image. The real-time remote sensing image to be segmented is acquired, the shooting date and geographical position are recorded, the latitude span of the image coverage area is calculated, if the latitude span is greater than a pre-set threshold, the image is divided into bands according to latitude, and a single-season factor is calculated for each band.

[0007] Preferably, in S4, the non-protected area single-season factor vector and the protected area adjusted seasonal factor vector are calculated, including: According to the micro-protected area boundary, the image is divided into protected area and non-protected area two types of regions, the color feature, vegetation shape and texture feature of all pixels in the non-protected area are extracted, and the matrix form is taken as the unified single-season factor vector of the non-protected area. The micro-plot boundary is identified in each protected area, the plot is divided into high-recognition- degree plot and low-recognition-degree plot, the monthly RGB, NDVI and GLCM contrast features of each high-recognition-degree plot are extracted, the local seasonal factor vector of the plot is calculated, and the local seasonal factor vector is weighted according to the boundary recognition score of the plot: , wherein is an adjustment coefficient, represents the local seasonal factor value of a single micro-plot at time point t; For the low-recognition-degree plot, a spatially adjacent plot with similar seasonal features is found in the surrounding high-recognition-degree plots, the local seasonal factor vector of the adjacent plot is used for interpolation filling of the vector of the current plot, if there is no suitable adjacent plot, the single-season factor vector of the non-protected area is used as the default value; the adjusted seasonal factor vectors of all micro-plots in the protected area are area-weighted averaged to generate the adjusted seasonal factor vector of the protected area: , wherein is the area of the i th plot, and M is the total number of plots.

[0008] Preferably, S3 includes: Based on the micro-plot sub-image after the banding processing, the spectral difference of the ridge is enhanced through the multi-spectral image, and the fuzzy boundary is corrected combined with the laser radar terrain data; For micro-terrain interference, filtering processing is implemented according to slope classification, and the boundary shape is dynamically adjusted by using curvature; A buffer zone is set at the latitude banding junction, the fusion weight is calculated according to the distance of the pixel to the center of the banding, and the boundary angle difference and the height mutation value difference of adjacent bands are constrained; ​Real-time response to the dynamic redrawing of the boundary of the agricultural machinery track, and output multi-level precision segmentation results according to the application scene.

[0009] Preferably, the filtering process according to the slope grading is implemented, and the boundary form is dynamically adjusted by using the curvature, including: The slope value of each pixel in the image is calculated, and the filtering method is selected according to the slope range; and the boundary sharpening intensity is dynamically adjusted according to the slope value: The filtered image is sharpened; Based on the boundary confidence map, the initial boundary line is extracted, and the curvature radius of each boundary line is calculated to distinguish the concave and convex boundaries; The boundary position is corrected by a dynamic adjustment formula Corrected_Boundary=initial boundary x curvature compensation coefficient; wherein the curvature radius of the concave boundary is negative, and a contraction compensation coefficient is adopted; the curvature radius of the convex boundary is positive, and an expansion compensation coefficient is adopted.

[0010] Preferably, the boundary angle difference and the elevation mutation value difference of adjacent zones are constrained, including: A buffer zone is established at the intersection of the latitude zoning, and the fusion weight is calculated according to the distance of the pixel to the zoning center: The pixel weight near the zoning center is high, and the edge pixel weight is low, the boundary feature value in the buffer zone is weighted and averaged, and the parameter jump at the zoning intersection is eliminated; the boundary angle difference and the elevation mutation value difference of adjacent zones are forced to be constrained, and if the constraint condition is not met, the laser radar terrain data is used as the leading, and the boundary parameters of the conflict area are recalculated.

[0011] Preferably, the filtering process according to the slope grading includes: The shadow coverage area is predicted by the sun azimuth and elevation model, and the slope boundary confidence under the shadow interference is dynamically corrected; Based on the rock spectrum feature library, the mimic interference area is identified, and the lithology misjudgment is suppressed by adjusting the terrain and optical feature weight; Multi-angle image fusion and distortion coefficient modeling are used to eliminate the reflection distortion caused by steep terrain, and a reliable boundary candidate set resistant to distortion interference is generated; Conflict arbitration is combined with the terrain, optical and lithological features to guarantee the boundary accuracy in different slope intervals.

[0012] Preferably, the slope boundary confidence under the shadow interference is dynamically corrected, including: The sun vector is obtained by calculating the sun azimuth and elevation based on the astronomical algorithm; a high-precision digital elevation model DEM matching the spatial range of the image is loaded; the slope and slope direction of each pixel are calculated through the DEM to generate a slope map and a slope direction map; According to the solar azimuth and elevation angle, combined with DEM data, the shadow projection direction and length of each terrain point are calculated; based on the solar elevation angle and terrain slope, the shadow coverage distance is calculated through L = h / tan(Elevation), wherein h is the terrain elevation difference, and Elevation is the elevation angle; for each pixel, the proportion of pixels covered by the shadow in its neighborhood is counted to generate a shadow coverage rate map; the pixels with a shadow coverage rate exceeding a threshold value are taken as the key objects for confidence correction; A boundary confidence map is extracted from the boundary detection result, and the value range thereof is [0, 1]. The boundary confidence of the shadow coverage area is dynamically attenuated, and the formula is as follows: Wherein k is an attenuation coefficient.

[0013] Preferably, the generation of the reliable boundary candidate set against distortion interference comprises the following steps: For each corrected image, an initial boundary is extracted using an edge detection algorithm to generate three boundary confidence maps. The boundary line segments that appear in the three boundary maps at the same time, that is, the intersection of the front view and the rear view boundary, are taken as high-reliability candidate boundaries. For the boundary in the nadir image that is not covered by the intersection but has high confidence, if it is located in a low-distortion area, it is included in the candidate set. Non-maximum suppression and connected component analysis are performed on the candidate boundaries to remove fragmented boundaries, and a continuous anti-distortion boundary candidate set is generated.

[0014] The application also provides a land-use-based remote sensing image segmentation system, which comprises: A collection and labeling module is configured to collect remote sensing images of the same region in consecutive periods, label and record the real land type and periodical seasonal factor vector of each region; A zonification processing module is configured to determine whether to perform zonification according to the latitude span of the real-time remote sensing image; An identification and demarcation module is configured to identify the micro-plot boundary after zonification processing, calculate the boundary identification degree, and demarcate a protection area; A calculation and fusion module is configured to calculate the non-protection-area single-season factor vector and the protection-area adjusted seasonal factor vector, and fuse them into a global seasonal factor; A clustering and segmentation module is configured to use a fuzzy C-means clustering model in combination with the global seasonal factor to preliminarily cluster and segment the image to divide multiple layers of pixel points; A clustering model correction module is configured to correct the fuzzy C-means clustering model in combination with the micro-plot seasonal characteristics and the boundary identification degree; A model correction module is configured to correct the fuzzy C-means clustering model in combination with the micro-plot seasonal characteristics and the boundary identification degree; A precision verification module is configured to perform image re-segmentation on a target region based on the corrected fuzzy C-means clustering model, and verify the segmentation accuracy.

[0015] The beneficial effects of the present application are: By collecting multi-temporal remote sensing images to construct a fine seasonal factor vector, combining with latitude zoning processing to adapt to seasonal differences in different regions, using boundary recognition degree to delineate protected areas to reduce segmentation errors, and fusing protected area and non-protected area features to generate global seasonal factor driven fuzzy C-means clustering, the method effectively improves the accuracy of land type classification in shadow areas, especially suitable for complex ground scenes, can significantly reduce the influence of seasonal changes and boundary ambiguity on the segmentation results, and at the same time, through multi-feature fusion and dynamic adjustment mechanism, the robustness and adaptability of the model are enhanced.

[0016] By fusing multi-spectral images and laser radar data to strengthen the ridge boundary features, combining with the terrain self-adaptive compensation mechanism to effectively solve the distortion problem caused by the fuzzy boundary of mountain farmland and terrain interference; using slope grading filtering and curvature dynamic adjustment technology, the boundary segmentation accuracy under complex terrain is significantly improved; cross-zoning buffer fusion and continuity constraint ensure the smooth transition of the boundary, while the real-time response of the agricultural machinery track and the multi-level precision output mechanism take into account the dynamic adaptability and different scene requirements.

[0017] By extracting multi-source features and marking conflicts, constructing terrain-optical, lithology-optical arbitration rules and comprehensive arbitration model, the arbitration score of conflict pixels can be accurately calculated. According to the accuracy grading standard and processing strategy, the boundary accuracy level can be effectively distinguished, the high reliable result is directly output, the medium and low reliable result is processed in a targeted manner, the boundary extraction accuracy is improved, and the manual intervention is reduced, which provides reliable data support for geological analysis and the like. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flowchart of a remote sensing image segmentation method based on land use according to the present application; Figure 2 A structural block diagram of a remote sensing image segmentation system based on land use according to the present application. DETAILED DESCRIPTION

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

[0020] In the description of the present application, the terms "first", "second" are only for descriptive purpose, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0021] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can recognize that the present application can be implemented without using these specific details. In other examples, well-known structures and processes will not be described in detail in order to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope consistent with the principles and characteristics disclosed in the present application.

[0022] Embodiment 1 Figure 1 is a flowchart of a land use-based remote sensing image segmentation method according to an embodiment of the present application, comprising the following steps: S1, collect remote sensing images of the same area in consecutive periods, label and record the real land type and seasonal factor vector of each area.

[0023] Specifically, remote sensing images of the same area in consecutive 12 months are collected to ensure that the images cover different seasons and weather conditions. Each image needs to be labeled with the real land type, such as "summer farmland", "snow", "autumn forest" and the like. Quantify the seasonal change of color, calculate the monthly average RGB value of each land type. Extract the normalized vegetation index NDVI, and establish the seasonal change curve of each land type reflecting the vegetation growth state. Calculate the contrast through the gray level co-occurrence matrix GLCM, and quantify the change of texture features with seasons. Based on the above features, define the seasonal factor vector of 12 months, including color offset (ΔRGB), active threshold (used to judge whether the land type is in active state), texture weight (reflecting the contribution of texture features to segmentation).

[0024] Among them, the color offset (ΔRGB) is to quantify the fluctuation of the RGB value of the land type with the season. For each land type of the monthly image, the RGB mean value of all pixels is calculated to obtain the RGB sequence of 12 months: where N is the number of pixels of the land type, and t is the month. The difference between the monthly RGB and the reference RGB is calculated, and normalized to [0, 1]: where and are the maximum / minimum values of the annual RGB. The representation of the RGB threshold is denoted as .

[0025] The active threshold (NDVI threshold) is used to determine whether the land type is in the active state of vegetation. For each land type, the normalized difference vegetation index (NDVI) is calculated: where NIR is the near-infrared band, and R is the red band. Set the NDVI threshold (e.g., 0.3), if the NDVI of a certain month exceeds the threshold, it is marked as active (1), otherwise it is marked as inactive (0). Record the active state of each month, or calculate the average NDVI of the active months as the threshold: where M is the number of active months. The representation of the active threshold is denoted as .

[0026] The texture weight GLCM contrast is used to quantify the change of texture features with seasons, and to give weight to segmentation. For each land type, the contrast feature of the gray level co-occurrence matrix (GLCM) is calculated: where P(i, j) is the joint probability of gray levels i and j in the GLCM. Normalize the contrast to [0, 1] to reflect the contribution of texture to segmentation: The representation of the texture weight GLCM is denoted as .

[0027] Combine the above three dimensions into a 12-dimensional vector, and expand it into a matrix form in time sequence: Input the seasonal factor vector into the random forest model, combine with the static features to improve the classification accuracy, and identify abnormal changes by comparing the seasonal factor vectors of different years.

[0028] S2, according to the latitude span of the real-time remote sensing image to determine whether to be zoned.

[0029] Specifically, obtain the real-time remote sensing image to be segmented, record the shooting date and geographical location, calculate the latitude span of the image coverage area (maximum latitude Lat_max = max(Lat_UL, Lat_LR), minimum latitude Lat_min = min(Lat_UL, Lat_LR), ​), if the latitude span is greater than a pre-set threshold (defined as 10° in this application, which can be adjusted according to actual situation, if ΔLat> threshold, execute zonation; otherwise, directly calculate the global seasonal factor), the image is zoned according to latitude, and a single seasonal factor is calculated for each zone to consider the difference in seasonal changes in different latitude regions.

[0030] wherein the zonation is calculated according to the dimension span, and the number of zones is N = ΔLat / threshold, for example, when the span is 25°, N = 3 (about 8.3° per zone). The latitude interval is divided according to equal interval: wherein k = 1, 2. If Lat_min = 20° N, Lat_max = 45° N, and threshold = 10°, then there are 3 zones: Band1 = [20°, 28.3°], Band2 = [28.3°, 36.7°], and Band3 = [36.7°, 45°]. The original image is cropped according to the zonation boundary using GIS tools to generate N sub-images, and the original shooting date and the information of the latitude zone to which each sub-image belongs are retained. The sub-images in each latitude zone k are matched, the difference between the monthly RGB mean value in the zone and the reference is calculated, the NDVI threshold is adjusted according to the vegetation type in the zone (for example, the threshold can be lower in high latitude regions) and the texture contrast in the zone is normalized, and an independent seasonal factor is generated for each zone according to the S1 method. For each sub-image, the corresponding seasonal factor vector is used for segmentation or classification, and the results of each zone are spliced into a complete image to process the boundary overlap area. The segmentation result map after zonation, the seasonal factor vector of each zone, and the zonation boundary file are obtained. Whether the zonation boundary is reasonable is checked, the accuracy of zonation and global processing is compared, and the effectiveness of zonation is verified.

[0031] S3, identifying the micro-plot boundary after zonation processing, calculating the boundary recognition degree, and delineating the protection area.

[0032] Specifically, the sub-images after zonation are matched with the corresponding seasonal factor vector, and a segmentation algorithm driven by the seasonal factor vector is used to generate an initial plot polygon. An edge detection algorithm is applied to each polygon to extract the boundary curve, the length-area ratio (the ratio of the boundary length to the polygon area is used to reflect the boundary complexity), the spectral consistency (the spectral similarity of the pixels on both sides of the boundary), and the texture heterogeneity (the difference in contrast / entropy based on GLCM on both sides of the boundary) are calculated for each boundary curve, and the above features are fused by weighting (the weights can be determined by training): , the boundaries with a recognition degree higher than a threshold (the threshold is set according to the actual situation, which is not limited in this application) are retained, the noise or fuzzy boundaries are filtered, and the high-recognition-degree boundaries are expanded outward by a certain distance (5-10 meters) to form a protection area, so as to avoid the influence of segmentation errors on the actual plot. If the image is zoned, the protection area results of each sub-image need to be combined to process the overlapping area (take the union set).

[0033] S4, calculate the non-protected area single-season factor vector and the protected area adjusted seasonal factor vector, and fuse them into a global seasonal factor.

[0034] The protected area adjusted seasonal factor vector is calculated according to the seasonal characteristics and boundary recognition degree of the internal micro plots.

[0035] Specifically, according to the micro protected area boundary generated in S3, the image is divided into protected area and non-protected area two types of regions, for all pixels in the non-protected area, the color feature (monthly RGB mean sequence, calculate the color offset ΔRGB), vegetation state (monthly NDVI sequence, determine the active critical value NDVI threshold) and texture feature (monthly GLCM contrast sequence, normalized to texture weight w) are extracted according to the method of S1, and the above features are taken as the unified single-season factor vector of the non-protected area according to the matrix form of S1 .

[0036] In each protected area, according to the micro plot boundary identified in S3, the plot is divided into high recognition degree plot (recognition degree > threshold) and low recognition degree plot (recognition degree ≤ threshold), for each high recognition degree plot, its monthly RGB, NDVI, GLCM contrast features are extracted (the method is the same as S1), and the local seasonal factor vector (12 dimensions ) of the plot is calculated, according to the boundary recognition degree (Recognition_Score) of the plot, the local seasonal factor vector is weighted: , wherein is the adjustment coefficient (the range defined by the application is 0.1~0.5, which controls the weighting amplitude), represents the local seasonal factor value of a single micro plot at time point t.

[0037] For low recognition degree plots, find spatially adjacent and seasonally similar plots (such as NDVI sequence correlation > 0.8) in the surrounding high recognition degree plots, and use the local seasonal factor vector of the adjacent plot to interpolate and fill the vector of the current plot, if there is no suitable neighborhood plot, use the single-season factor vector of the non-protected area as the default value. For the adjusted seasonal factor vector of all micro plots in the protected area, the adjusted seasonal factor vector of the protected area is generated by area weighted average: , wherein is the area of the i-th plot, and M is the total number of plots.

[0038] According to the area proportion of the non-protected area and the protected area, the weights and ( ) of the two are determined, and for each time point t, the global seasonal factor vector is: The fused vector is time smoothed to eliminate mutations caused by regional differences.

[0039] S5, a fuzzy C-means clustering model is used in combination with a global seasonal factor to preliminarily cluster and segment the image to divide multiple layers of pixel points.

[0040] In this application, the number of clusters c is set to 3 (shadow, non-shadow, and possible shadow).

[0041] Specifically, the color image is converted into a grayscale image, the number of clusters is determined according to actual requirements, the fuzziness of the clustering result is controlled, and the value is between [1.5, 2.5]. The maximum number of iterations of the algorithm is set to prevent infinite loops. The objective function of FCM is defined according to the principle of minimizing the intra-class variance and maximizing the inter-class variance, the membership matrix U is randomly initialized, and the sum of the membership degrees of each sample point to all cluster centers is ensured to be 1. Different weights are given to different cluster centers according to the global seasonal factor, reflecting the influence of the season on the clustering result. The global seasonal factor is introduced into the objective function, so that the clustering result is more in line with the seasonal characteristics.

[0042] According to the current cluster center and the membership matrix, the membership of each sample point to each cluster center is calculated, the coordinates of each cluster center are updated according to the membership matrix, and it is checked whether the objective function value reaches the convergence condition (such as the change is less than a threshold value or reaches the maximum number of iterations). If it does not converge, the iteration calculation is continued.

[0043] The final membership matrix U is output, where each element represents the membership of a sample point to a cluster center, and each sample point is assigned to the category corresponding to the cluster center with the largest membership according to the membership matrix.

[0044] The shadow category in the clustering result is binarized to obtain a binary image of the shadow region, and a connected component labeling algorithm (two-pass scanning method) is used to label the shadow region in the binary image. According to the actual size of the shadow region, an area threshold is set to filter out small connected components (which may be noise or misclassification). The remaining connected components are analyzed for shape features (aspect ratio, circularity) to further select suspected shadow regions.

[0045] The suspected shadow regions are hierarchically divided according to their gray value or membership value to obtain multiple layers of pixel points. For example, the shadow region can be divided into core shadow, edge shadow, and transition shadow. Pixel points are extracted from each level for subsequent applications such as shadow detection, shadow compensation, or shadow removal.

[0046] S6, the fuzzy C-means clustering model is corrected in combination with the seasonal characteristics of the micro-land and the boundary recognition degree.

[0047] Specifically, for each layer of shadow, the average width of the shadow along the direction perpendicular to the boundary (the shadow area width) is calculated, and the gray mean value μ and the standard deviation of each layer of shadow are extracted. Based on historical data, a shadow feature template (width range, gray distribution) of different land types (such as vegetation, water body, building, and bare land) is established. For each layer of shadow, the matching degree with each type of land type template (through Gaussian mixture model or Mahalanobis distance) is calculated to obtain the probability that the suspected shadow area corresponds to each type of land type. The clustering center or membership degree calculation method of FCM is adjusted according to the land type probability to improve the classification accuracy.

[0048] S7, based on the corrected fuzzy C-means clustering model, the target area is re-segmented, and the segmentation accuracy is verified.

[0049] The technical solutions in the embodiments of the present application have at least the following technical effects or advantages: By collecting multi-temporal remote sensing images to construct a fine seasonal factor vector, combining with latitude zoning processing to adapt to seasonal differences in different regions, using boundary recognition degree to delineate the protected area to reduce segmentation error, and fusing the features of the protected area and the non-protected area to generate a global seasonal factor driven fuzzy C-means clustering. This method effectively improves the accuracy of land type classification in shadow areas, especially suitable for complex ground scenes, can significantly reduce the influence of seasonal changes and boundary ambiguity on the segmentation results, and at the same time, through multi-feature fusion and dynamic adjustment mechanism, the robustness and adaptability of the model are enhanced.

[0050] Embodiment 2 In embodiment 1, when performing micro-plot boundary segmentation of farmland based on remote sensing images, it mainly relies on a single optical image feature (such as spectral difference) or a fixed parameter boundary extraction algorithm. However, in complex terrain conditions such as mountainous areas and slope areas, the ridge boundary is easily affected by terrain undulations, micro-topographic disturbances, and crop growth differences, making it difficult for a single feature or fixed parameter model to accurately capture the boundary features. Especially when the terrain mutation degree and crop coverage type difference between different plots are large, the unified model is prone to misjudgment or cracking in the detection of ambiguous boundary regions, and has weak adaptability to complex terrain areas (such as slope ridges). In order to further improve the accuracy and robustness of boundary segmentation, it is necessary to consider multi-source data fusion of optical images and laser radar terrain data and dynamic adjustment strategies that adapt to terrain. Embodiment 2 introduces improvement methods such as terrain feature enhancement, curvature compensation, and cross-zoning continuity constraint, effectively solving the problem of insufficient generalization ability of boundary detection in complex terrain, and achieving more accurate farmland plot boundary segmentation.

[0051] In some embodiments, in step S3, identifying the micro-plot boundary after zoning processing further includes: S31, based on the micro-plot sub-image after zoning processing, the spectral difference of the ridge is enhanced by multi-spectral image, and the fuzzy boundary is corrected combined with laser radar terrain data.

[0052] Specifically, the micro-plot image after zoning processing is enhanced by waveband fusion to enhance the spectral difference between the ridge and the crop, the normalized difference index (NDXI, a combination of near-infrared and red edge waveband) is calculated to highlight the linear features of the ridge, the histogram of oriented gradients (HOG) algorithm is applied to count the distribution of gradient direction in the local area, and the direction of the ridge and the edge directionality are extracted, the linearity score of each pixel is calculated based on the HOG feature vector, and a linear feature enhancement map is generated. The normalized difference index NDXI value and the linearity score are weighted and fused (the weight can be adjusted according to the experiment, for example, NDXI accounts for 0.7 and linearity accounts for 0.3); the fusion result is normalized (Min-Max standardization), a boundary probability heat map is generated, the pixel value range is [0, 1], and the value closer to 1 indicates a higher boundary possibility.

[0053] The laser radar point cloud data of the same area is acquired to generate a high-precision digital elevation model (DEM), and the point cloud classification (distinguishing ground points and non-ground points) is completed. The elevation difference between each pixel and its neighborhood is calculated, the points with vertical height difference exceeding the threshold (> 15 cm) are screened, and the elevation mutation points of the ridge area are extracted; the change of slope direction around the pixel is counted, the area with slope direction change less than the angle threshold (< 10°) is retained, the slope direction consistency index is calculated to ensure the reliability of the terrain features.

[0054] The formula for calculating the terrain mutation coefficient is: The formula for calculating the fusion of optical and terrain features is: A corrected boundary confidence map is generated to strengthen the boundary detection result in complex terrain areas (such as slope ridges).

[0055] S32, for micro-terrain interference, according to the slope classification, the filter processing is implemented, and the boundary shape is dynamically adjusted by using curvature.

[0056] Specifically, the slope value of each pixel in the image is calculated, and the filter method is selected according to the slope range: Gaussian filter is used for low slope (< 5°) to smooth noise and retain the overall boundary trend; bilateral filter is used for medium slope (5-15°) to retain the edge and avoid excessive smoothing; guided filter is applied for high slope (> 15°) to enhance the linear features and compensate for the boundary distortion caused by terrain undulation. The boundary sharpening intensity is dynamically adjusted according to the slope value: The filtered image is sharpened to improve the boundary definition.

[0057] The initial boundary line is extracted based on the boundary confidence map, and the curvature radius of each boundary line is calculated to distinguish the concave (such as a ditch) boundary from the convex (such as a ridge) boundary. The boundary position is corrected by dynamically adjusting the formula: Corrected_Boundary = Initial Boundary x Curvature Compensation Coefficient.

[0058] wherein the curvature radius of the concave boundary is negative, and a contraction compensation coefficient (0.8) is adopted; the curvature radius of the convex boundary is positive, and an expansion compensation coefficient (1.2) is adopted.

[0059] S33, a buffer zone is set at the latitude zoning boundary, a fusion weight is calculated according to the distance of a pixel to the zoning center, and the boundary angle difference and the elevation mutation value difference of adjacent zoning are constrained.

[0060] Specifically, a buffer zone (width 200m) is established at the latitude zoning boundary, and a fusion weight is calculated according to the distance of a pixel to the zoning center: The pixel weight near the zoning center is high (W approaches 1), and the edge pixel weight is low (W approaches 0). The boundary feature values in the buffer zone are weighted and averaged to eliminate the parameter jump at the zoning boundary. The boundary angle difference and the elevation mutation value difference of adjacent zoning are forcibly constrained (the boundary angle difference of adjacent zoning is less than a threshold (<15°); the elevation mutation value difference of adjacent zoning is less than a threshold (<20cm)). If the constraint condition is not met, the laser radar terrain data is used as the dominant, and the boundary parameters of the conflict area are recalculated.

[0061] S34, the boundary is dynamically redrawn in real time in response to the agricultural machine track, and multi-level precision segmentation results are output according to application scenarios.

[0062] Specifically, the agricultural machine track data is accessed in real time. If it is detected that the agricultural machine crosses the historical boundary, the boundary redrawing mode is activated, a temporary boundary is generated based on the agricultural machine track point, and the original segmentation result is covered. The boundary database is dynamically updated to ensure that the segmentation result reflects the latest cultivation state. According to the application scenario, boundary results of different precisions are output. High precision (L1) is used for farmland right registration, and the boundary error is controlled within a very small range; medium precision (L2) is used for precision irrigation planning, and the precision and calculation efficiency are balanced; low precision (L3) is used for regional planting area statistics, and a larger error is allowed to improve the processing speed.

[0063] The technical solutions in the above embodiments of the application have at least the following technical effects or advantages: The boundary features of the ridge are strengthened by fusing multispectral images and LiDAR data, and the distortion problem caused by blurred boundaries of mountainous farmland and terrain interference is effectively solved by combining a terrain self-adaptive compensation mechanism; the boundary segmentation accuracy under complex terrain is significantly improved by using slope grading filtering and curvature dynamic adjustment technology; cross-zoning buffer fusion and continuity constraint ensure smooth transition of the boundary, while real-time response of the agricultural machinery track and multi-level precision output mechanism take into account dynamic adaptability and different scene requirements.

[0064] Embodiment 3 In embodiment 2, although a series of technical means such as zoning processing, multispectral image enhancement, LiDAR terrain data correction, and filtering processing for micro-terrain interference effectively improve the accuracy of ridge boundary detection, and fusion processing is performed at the latitude zoning boundary to reduce parameter jumps, real-time response to agricultural machinery track dynamic redrawing of the boundary and output of multi-level precision segmentation results according to application scenarios are also realized. However, in actual complex farmland environments, the terrain, crop type, and lighting conditions in different regions differ greatly, making it difficult for a single parameter or fixed weight fusion method to achieve optimal boundary detection results in all regions, and there are still some errors in boundary detection under some special terrain or complex lighting conditions, and the generalization ability needs to be further improved.

[0065] Therefore, the embodiments of the present application perform certain optimization on the basis of the above-mentioned embodiments.

[0066] In some embodiments, at step S32, the filtering processing is performed according to slope grading, and further includes: S321, predicting the shadow coverage area by a solar azimuth and elevation model, and dynamically correcting the slope boundary confidence under shadow interference.

[0067] Specifically, based on the astronomical algorithm (SPA algorithm), the solar azimuth (Azimuth, 0° for north, increasing clockwise) and the elevation angle (Elevation, 0° for the horizon, 90° for the zenith) are calculated to obtain the solar vector (including azimuth and elevation). A high-precision digital elevation model (DEM) matching the spatial range of the image is loaded, and the resolution thereof needs to be better than 1 meter. The slope (Steepness) and aspect (Aspect) of each pixel are calculated by the DEM to generate a slope map and an aspect map. The slope is used to quantify the steepness of the terrain, and the aspect is used to determine the relative relationship between the light incidence direction and the terrain.

[0068] According to the solar azimuth and elevation angle, combined with DEM data, the shadow projection direction and length of each terrain point are calculated. The shadow projection direction is determined by the solar azimuth and slope direction (for example, when the solar azimuth is 180° (due south), the north slope will form a shadow). Based on the solar elevation angle and terrain slope, the shadow coverage distance is calculated by trigonometric function (such as L = h / tan(Elevation), where h is the terrain elevation difference). For each pixel, the proportion of pixels covered by the shadow in its neighborhood (such as a 3x3 window) is counted, and a shadow coverage rate map (range 0%~100%) is generated. The pixels with a shadow coverage rate exceeding a threshold (such as 60%) are used as the focus objects for subsequent confidence correction.

[0069] The boundary confidence map is extracted from the preliminary boundary detection result (initial segmentation based on spectrum or texture), with a value range of [0, 1], and the larger the value, the higher the boundary reliability. The boundary confidence of the shadow coverage area is dynamically attenuated, and the formula is: where k is the attenuation coefficient (the value is 2 in this application, and the specific value can be adjusted according to actual application), and the higher the shadow coverage rate, the more significant the confidence attenuation.

[0070] S322, identify the mimic interference area based on the rock spectral feature library, and suppress the lithology misjudgment by adjusting the terrain and optical feature weight.

[0071] Specifically, collect multispectral data of typical rock samples in the target area, ensuring coverage of visible-near infrared (VNIR) and shortwave infrared (SWIR) bands. Identify key spectral features such as reflection peaks (maximum reflectivity at specific wavelengths such as 550nm, 650nm), absorption valleys (minimum reflectivity at specific wavelengths such as 2200nm, 2340nm), and characteristic indices (calculate spectral slope such as 650-900nm red edge slope). Define a spectral fingerprint template for each lithology, including characteristic wavelength position, reflectivity threshold, and mimic risk index (such as limestone with high mimic risk due to strong absorption characteristics, index set to 0.91).

[0072] Load the spectral bands of the image to be detected (which need to include the bands defined in the rock feature library, such as 550nm, 2200nm), and load the terrain data (slope, slope direction) simultaneously to assist spatial analysis. For each pixel in the image, extract its spectral features and calculate the similarity with the templates in the rock library. Set a matching threshold (this application defines similarity > 0.85), mark the pixels that meet the conditions as suspected mimic areas, and perform morphological processing (such as dilation, erosion, etc.) on the suspected areas to merge fragmented pixels, generating a continuous mimic interference mask. Combine the terrain data to evaluate the mimic risk level, and prioritize high-risk areas (such as areas with slope > 30° and mimic index > 0.8).

[0073] The initial weights of the terrain features (such as slope discontinuity, DEM elevation difference) and optical features (such as spectral reflectance, texture entropy) are defined, and the default terrain weight = 0.5, and the optical weight = 0.5 (this weight can be adjusted according to experiments). When a pixel is labeled as a suspected mimic area, the weight adjustment is triggered, and for each pixel, the weighted boundary confidence is calculated: , where the terrain confidence can be obtained by DEM gradient or slope discontinuity detection, and the optical confidence comes from the initial spectral boundary detection result.

[0074] Based on the weighted confidence map, the final boundary is extracted using non-maximum suppression (NMS). For mimic suppression areas, the terrain boundary (obtained by DEM contour fitting) is preferred to replace the optical boundary to ensure boundary continuity. By comparing the boundaries before and after weight adjustment with the ground measured data (RTK-GPS measurement), the boundary offset amount improvement ratio is calculated. By superimposing the original boundary, the adjusted boundary, and the mimic area mask, it is verified whether the boundary of the steep slope area is clear and free of lithology misjudgment. According to the verification result, the matching threshold (such as from 0.85 to 0.88) and the weight proportion (such as the terrain weight from 0.8 to 0.75) are dynamically optimized to balance false positives and false negatives. For different lithology distribution areas (such as limestone dense area vs. granite dense area), customized rock feature library and weight adjustment strategy are used to improve the universality of the scheme.

[0075] wherein the weight adjustment includes: terrain weight increase, i.e. increasing the terrain feature weight to 0.7-0.8 (such as terrain weight = 0.8), to strengthen the contribution of terrain boundaries (such as ridges, slope discontinuity) to the final boundary; and optical weight reduction, i.e. reducing the optical feature weight to 0.2-0.3 (such as optical weight = 0.2), to suppress the interference of spectral mimicry on boundary detection.

[0076] S323, using multi-angle image fusion and distortion coefficient modeling, to eliminate reflection distortion caused by steep slope terrain and generate a reliable boundary candidate set resistant to distortion interference.

[0077] Specifically, three-angle images (sub-pixel level registration of forward-looking, backward-looking and nadir images) of the same target area are obtained using multi-view sensors (nadir image sensor vertical to the ground observation incident angle ≈ 0°, forward-looking image sensor forward inclined observation incident angle ≈ 15° and backward-looking image sensor backward inclined observation incident angle ≈ 15°) to ensure the position alignment of the same ground object in the three images (error < 0.5 pixels).

[0078] Steepness and Aspect of each pixel are extracted from high-precision DEM to generate Steepness map and Aspect map. Steep slopes (Steepness > 25°) will cause two main types of reflection distortion: one is specular reflection distortion, when the solar incidence angle is close to the Aspect (e.g. the difference between the solar azimuth and the Aspect is < 30°), the ground surface produces specular reflection, resulting in saturation of reflectance (loss of boundary features); the other is shadow-light mixed distortion, local shadow and light appear alternately on the slope, resulting in reflectance mutation in the boundary area (pseudo-edge interference). According to the Steepness, solar elevation angle (obtained from S321), and incidence angle (determined by the image angle), a reflection distortion correction coefficient matrix is established: wherein, is the terrain Steepness, is the solar azimuth, is the Aspect, is an empirical coefficient (calibrated by experiment, defined in this application = 0.5).

[0079] For each image (nadir, forward, and backward), according to its incidence angle and the Steepness and Aspect of the corresponding pixel, the reflectance is corrected using the distortion coefficient C: wherein, is the original reflectance, is the corrected reflectance, according to the size of the correction coefficient C, mark the high-distortion risk area (e.g. C > 1.5), and reduce its weight in subsequent fusion.

[0080] For each corrected image, use the edge detection algorithm to extract the initial boundary and generate three boundary confidence maps. Retain the boundary line segments that appear in all three boundary maps (i.e. the intersection of the forward and backward boundaries) as high-reliability candidate boundaries. For the boundaries in the nadir image that are not covered by the intersection but have high confidence, if they are located in the low-distortion area (C < 1.5), they are included in the candidate set. Perform non-maximum suppression (NMS) and connected component analysis on the candidate boundaries to remove fragmented boundaries and generate a continuous anti-distortion boundary candidate set.

[0081] S324, combine terrain, optical, and lithological features to perform conflict arbitration and guarantee the boundary accuracy in different slope intervals.

[0082] Specifically, calculate the Steepness and Aspect of each pixel from the high-precision DEM to generate a terrain feature map. Define the terrain risk level: low-risk area with Steepness < 15° (flat terrain, weak reflection distortion); medium-risk area with 15° ≤ Steepness ≤ 30° (undulating terrain, possible local shadow); high-risk area with Steepness > 30° (steep slope, significant specular reflection and shadow-light mixed distortion).

[0083] Spectral reflectance, texture features and edge confidence are extracted from multispectral / hyperspectral imagery. Based on the rock spectral feature library (constructed in step S322), the lithology type of each pixel (such as granite, sandstone, limestone) is matched, and the lithology sensitive area is marked. High sensitive lithology refers to the lithology whose spectral features are easily confused with terrain (such as limestone which is easily misjudged as shadow due to strong absorption characteristics). Low sensitive lithology refers to the lithology whose spectral features are stable (such as granite whose reflectance is smooth and less affected by terrain). For each pixel, if its terrain risk level is inconsistent with the optical / lithology features (such as high-risk terrain area but high optical boundary confidence), it is marked as a conflict pixel and needs to be arbitrated later.

[0084] A conflict arbitration rule library is constructed. In terms of terrain-optical arbitration rules, low-risk areas (slope < 15°) prefer to trust optical boundaries (weight = 0.8), and terrain is only used as auxiliary verification (such as boundary parallel to contour line, confidence +0.1). In medium-risk areas (15°≤ slope ≤ 30°), the weight is dynamically adjusted, , where θ is the slope, to ensure that the terrain weight increases with the increase of the slope; in high-risk areas (slope > 30°), the terrain boundary is preferred (weight = 0.7), and the optical boundary needs to meet additional conditions (such as angle < 15° with terrain boundary) to be accepted. In terms of lithology-optical arbitration rules, the high sensitive lithology area reduces the optical boundary confidence threshold (from 0.7 to 0.6) to avoid misjudgment of the boundary due to spectral mimicry. If the terrain boundary is consistent with the lithology distribution (such as limestone boundary coinciding with steep slope foot), the confidence is +0.2. The low sensitive lithology area (such as granite) maintains the default optical boundary confidence threshold (0.7), and the terrain is only used to correct local anomalies (such as isolated noise points).

[0085] The comprehensive arbitration model is to calculate the comprehensive arbitration score of each conflict pixel where is the terrain boundary confidence (calculated by DEM gradient, range [0, 1]), is the optical boundary confidence (output by edge detection algorithm, range [0, 1]), is the lithology matching score (complete match = 1, no match = 0), , , is the dynamic weight (sum = 1), determined by the slope interval and the lithology sensitivity.

[0086] According to the boundary precision grading guarantee, the precision grading standard is: first precision (high reliability) requires arbitration score S>0.85, and meets one of the two conditions of being located in a low-risk terrain area and the optical boundary confidence being greater than 0.9, or being located in a high-risk terrain area but the terrain and the optical boundary completely overlap (the included angle is less than 5°); second precision (medium reliability) requires 0.6<=S<0.85, which needs to be verified manually or combined with auxiliary data to confirm; third precision (low reliability) requires S<0.6, which is marked as a to-be-corrected area and needs to reacquire data or adjust algorithm parameters. The grading processing strategy is: the first precision boundary is directly output as the final result without subsequent processing; the second precision boundary is superimposed with the terrain contour or lithology distribution map in GIS, and the boundary rationality is checked manually, and the arbitration score is recalculated by using local window analysis (such as 3*3 pixels) for suspicious areas; the third precision boundary expands the search range to reextract features, and if the conflict cannot be solved, it is marked as an "uncertain area" and the coordinates are recorded for subsequent on-site verification.

[0087] The technical solutions in the embodiments of the application have at least the following technical effects or advantages: By extracting multiple source features and marking conflicts, constructing terrain-optical, lithology-optical arbitration rules and a comprehensive arbitration model, the arbitration score of the conflict pixels can be accurately calculated. According to the precision grading standard and the processing strategy, the boundary precision level can be effectively distinguished, the high-reliability result is directly output, the medium and low-reliability results are processed in a targeted manner, the boundary extraction accuracy is improved, the manual intervention is reduced, and reliable data support is provided for geological analysis and the like.

[0088] Further, the embodiment of the application also provides a land use-based remote sensing image segmentation system.

[0089] Figure 2 is a structural schematic diagram of a land use-based remote sensing image segmentation system according to the embodiment of the application.

[0090] As shown in Figure 2 , a land use-based remote sensing image segmentation system includes a collection and labeling module, a zoning processing module, an identification and delineation module, a calculation and fusion module, a clustering and segmentation module, a model correction module, and a precision verification module.

[0091] The collection and labeling module is used to collect continuous period remote sensing images of the same area, label and record the real land type and period seasonal factor vector of each area; The zoning processing module is used to determine whether to zone according to the latitude span of the real-time remote sensing image; The identification and delineation module is used to identify the micro-plot boundary after zoning processing, calculate the boundary identification degree, and delineate the protection area; The computing fusion module is configured to compute a non-protection zone single-season factor vector and a protection zone adjusted seasonal factor vector, and fuse them into a global seasonal factor. The clustering segmentation module is configured to preliminarily cluster and segment the image into multiple layers of pixel points by using a fuzzy C-means clustering model in combination with the global seasonal factor. The clustering model correction module is configured to correct the fuzzy C-means clustering model in combination with the micro-plot seasonal characteristics and the boundary recognition degree. The model correction module is configured to correct the fuzzy C-means clustering model in combination with the micro-plot seasonal characteristics and the boundary recognition degree. The precision verification module is configured to perform image re-segmentation on the target region based on the corrected fuzzy C-means clustering model, and verify the segmentation precision.

[0092] It should be noted that other specific implementation contents of the remote sensing image segmentation system based on land use in the embodiments of the present application can refer to the remote sensing image segmentation method based on land use.

[0093] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0094] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0095] The present application is described with reference to flowcharts and / or block diagrams according to the method, device (system), and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a machine that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device for performing the functions specified in one flow or multiple flows and / or blocks Figure 1 The device for performing the functions specified in one flow or multiple flows and / or blocks

[0096] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0097] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that are executed on the computer or other programmable apparatus provide steps for implementing the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0098] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such additional variations and modifications as fall within the scope of the present application. What is claimed is:

[0099] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A land use based remote sensing image segmentation method, characterized in that, The method comprises: S1, collecting remote sensing images of the same region in consecutive periods, labeling and recording the real land type of each region and the seasonal factor vector of the period; S2, determining whether to divide into zones according to the latitude span of the real-time remote sensing image; S3, identifying the boundary of the micro-land block after zoning processing, calculating the boundary recognition degree and delineating the protection area; S4, calculating the single-season factor vector of the non-protection area and the adjusted seasonal factor vector of the protection area, and fusing them into a global seasonal factor; S5, using a fuzzy C-means clustering model combined with the global seasonal factor to preliminarily cluster and segment the image to divide multiple layers of pixel points; S6, combining the seasonal characteristics of the micro-land block and the boundary recognition degree to correct the fuzzy C-means clustering model; S7, based on the corrected fuzzy C-means clustering model, re-segmenting the target area image to verify the segmentation accuracy. 2.The land use based remote sensing image segmentation method according to claim 1, characterized in that, S2, according to the latitude span of the real-time remote sensing image to determine whether to divide into zones, comprising: Obtain the real-time remote sensing image to be segmented, record the shooting date and geographical location, calculate the latitude span of the image coverage area, if the latitude span is greater than the pre-set threshold, divide the image by latitude, and calculate the single-season factor for each zone. 3.The land use based remote sensing image segmentation method according to claim 2, characterized in that, In S4, the single-season factor vector of the non-protection area and the adjusted seasonal factor vector of the protection area are calculated, comprising: According to the micro-protected area boundary, the image is divided into protected area and non-protected area two types of regions, for all pixels in the non-protected area, extract color features, vegetation shape and texture features, as a matrix form as a unified single-season factor vector of non-protected area ; The micro-plot boundaries identified in each protected area divide the plots into high-identification plots and low-identification plots. For each high-identification plot, its monthly RGB, NDVI, and GLCM contrast features are extracted separately, the local seasonal factor vector of the plot is calculated, and the local seasonal factor vector is weighted according to the boundary recognition degree Recognition_Score of the plot: wherein is an adjustment coefficient, represents the local seasonal factor value of a single micro-plot at a time point t. For low-identification degree plots, find spatially adjacent and seasonally similar plots in the high-identification degree plots around it, interpolate the current plot's vector with the local seasonal factor vector of the adjacent plots, and if there is no suitable adjacent plot, use the single seasonal factor vector of the non-protected area as the default value; for the adjusted seasonal factor vector of all micro-plots in the protected area, generate the adjusted seasonal factor vector of the protected area by area-weighted average: wherein is the area of the i-th plot, and M is the total number of plots.

4. The land use based remote sensing image segmentation method according to claim 1, characterized in that, S3, comprising: Based on the micro-land block sub-image after zoning processing, enhance the spectral difference of the ridge through multi-spectral image, and correct the fuzzy boundary combined with laser radar terrain data; For micro-terrain interference, implement filtering processing according to slope classification, and dynamically adjust the boundary shape by using curvature; Set a buffer zone at the latitude zoning boundary, calculate the fusion weight according to the distance from the pixel to the zoning center, and constrain the boundary angle difference and elevation mutation value difference of adjacent zoning; Real-time response to agricultural machinery trajectory to dynamically redraw the boundary, and output multi-level precision segmentation results according to application scenarios.

5. The land use based remote sensing image segmentation method according to claim 4, characterized in that, The slope classification filtering processing comprises: The slope value of each pixel in the image is calculated, a filtering method is selected according to the slope range, and the boundary sharpening intensity is dynamically adjusted according to the slope value: The filtered image is subjected to sharpening processing; Based on the boundary confidence map to extract the initial boundary line, calculate the curvature radius of each boundary line to distinguish concave and convex boundaries; Correct the boundary position through a dynamic adjustment formula Corrected_Boundary=initial boundary x curvature compensation coefficient; wherein the curvature radius of the concave boundary is negative, and a contraction compensation coefficient is used; the curvature radius of the convex boundary is positive, and an expansion compensation coefficient is used.

6. The land use based remote sensing image segmentation method according to claim 4, characterized in that, The boundary angle difference and elevation mutation value difference of adjacent zoning are constrained, comprising: A buffer zone is established at the boundary of latitude zones, and the fusion weight is calculated according to the distance from the pixel to the center of the zone: The weight of the pixel close to the center of the zone is high, and the weight of the edge pixel is low. The boundary characteristic value in the buffer zone is weighted and averaged to eliminate the parameter jump at the boundary of the zone. The difference between the boundary angle of adjacent zones and the difference of the height mutation value are forced to be constrained. If the constraint condition is not met, the laser radar terrain data is used as the dominant factor to recalculate the boundary parameters of the conflict area.

7. The land use based remote sensing image segmentation method according to claim 6, characterized in that, The slope classification filtering processing comprises: Predict the shadow coverage area through the sun azimuth and elevation model, and dynamically correct the slope boundary confidence under shadow interference; Identify the mimic interference area based on the rock spectral feature library, suppress the lithology misjudgment by adjusting the terrain and optical feature weight; Use multi-angle image fusion and distortion coefficient modeling to eliminate the reflection distortion caused by steep slope terrain, and generate a reliable boundary candidate set resistant to distortion interference; Combine the terrain, optical and lithological features to arbitrate conflicts and guarantee the boundary accuracy in different slope intervals. 8.The land use based remote sensing image segmentation method according to claim 7, characterized in that, The dynamic correction of the slope boundary confidence under shadow interference comprises: The sun vector is obtained by calculating the sun azimuth and elevation based on astronomical algorithm; a high-precision digital elevation model (DEM) matching the image space range is loaded; the slope and slope direction of each pixel are calculated through the DEM to generate a slope map and a slope direction map; According to the sun azimuth and elevation, the shadow projection direction and length of each terrain point are calculated in combination with the DEM data; based on the sun elevation and terrain slope, the shadow coverage distance is calculated through L=h / tan(Elevation), wherein h is the terrain elevation difference and Elevation is the elevation angle; for each pixel, the proportion of pixels covered by the shadow in the neighborhood is counted to generate a shadow coverage rate map; the pixels with a shadow coverage rate exceeding a threshold value are taken as the key objects for confidence correction; The boundary confidence map is extracted from the boundary detection result, and the value range thereof is [0, 1]. The boundary confidence of the shadow coverage area is dynamically attenuated, and the formula is as follows: wherein k is an attenuation coefficient.

9. The land use based remote sensing image segmentation method of claim 8, wherein, The method further comprises: For each corrected image, an initial boundary is extracted using an edge detection algorithm to generate three boundary confidence maps; the boundary line segments appearing in the three boundary maps at the same time, i.e., the intersection of the front-view and rear-view boundaries, are retained as high-reliability candidate boundaries; for the boundary in the sub-solar point image which is not covered by the intersection but has a high confidence, if the boundary is located in a low-distortion region, the boundary is included in the candidate set; the candidate boundaries are subjected to non-maximum suppression and connected component analysis to remove fragmented boundaries, thereby generating a continuous anti-distortion boundary candidate set.

10. A land use based remote sensing image segmentation system applied to the land use based remote sensing image segmentation method according to any one of claims 1 to 9, characterized in that, The system comprises: A collection and labeling module is configured to collect remote sensing images of the same region in consecutive periods, and label and record the real land type and seasonal factor vector of each region and period; A zonification processing module is configured to determine whether to perform zonification according to the latitude span of the real-time remote sensing image; An identification and demarcation module is configured to identify the micro-plot boundary after zonification, calculate the boundary recognition degree, and demarcate the protection area; A calculation and fusion module is configured to calculate the single-season factor vector of the non-protection area and the adjusted seasonal factor vector of the protection area, and fuse them into a global seasonal factor; A clustering and segmentation module is configured to preliminarily cluster and segment the image into multiple layers of pixel points by using a fuzzy C-means clustering model in combination with the global seasonal factor; A clustering model correction module is configured to correct the fuzzy C-means clustering model in combination with the seasonal characteristics of the micro-plot and the boundary recognition degree; A model correction module is configured to correct the fuzzy C-means clustering model in combination with the seasonal characteristics of the micro-plot and the boundary recognition degree; An accuracy verification module is configured to perform image re-segmentation on the target region based on the corrected fuzzy C-means clustering model, and verify the segmentation accuracy.

Images