A method for extracting fuzzy boundaries of hilly agricultural plots in a hot zone based on remote sensing images

Abstract

The present application relates to the technical field of land boundary recognition and extraction, and specifically discloses a method for extracting fuzzy boundaries of hilly agricultural land in a hot region based on remote sensing images, which comprises the following steps: step S1, obtaining remote sensing image data, using image processing, deep learning and statistical models to construct a land boundary fuzziness index, classifying and scaling the image according to the index, and obtaining an image with obvious boundary features; step S2, using the spectrum, texture and spatial information of the remote sensing image to construct a multi-branch cascaded deep neural network model for land boundary recognition; and step S3, using the algorithms of maximum suppression binarization, connectivity detection and smoothness detection to post-process the recognition result, and obtaining the clear boundaries of hilly farmland in the hot region. The method for extracting fuzzy boundaries of hilly agricultural land in a hot region based on remote sensing images can realize intelligent extraction of fuzzy boundaries of hilly agricultural land.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of plot boundary recognition and extraction, and particularly relates to a method for extracting fuzzy boundaries of hilly agricultural plots in a hot region based on remote sensing images. BACKGROUND

[0002] China is a large agricultural producer, and agriculture plays a vital role in the national economy of China. With the vigorous development of agricultural informatization, promoting the construction of smart agriculture and precision agriculture is of great significance to improve the competitiveness of China's agriculture. The vector boundary of the plot is an important feature necessary to determine the location, shape and spatial range of the agricultural unit, and is also the basis for realizing the precision and scale production of agriculture. For the tropical crop planting areas in southern China, the crops in this region are mostly planted in the mountainous and hilly areas with fragmented farmland and complex topography (such as about 40% in Hainan and about 70% in Guangxi), which leads to fuzzy boundaries of plots in remote sensing images in the hot crop region, and also makes it difficult to identify various crop types in the image, affecting the relevant departments to timely grasp the local agricultural information, and making it difficult to carry out smart management of plots, agricultural yield prediction, agricultural risk assessment, etc. Therefore, realizing the accurate and rapid automatic extraction of fuzzy plot boundaries in complex environments such as hilly and mountainous areas is a key problem to be solved in the development of agriculture in the hot crop region.

[0003] However, using the traditional manual digitization method to extract plot boundaries not only takes time and effort, but also requires interpreters to have rich professional knowledge. With the rapid development of image processing and computer vision, based on remote sensing image segmentation technology and deep learning technology, plot information can be quickly and accurately extracted. The plot boundary extraction method based on image segmentation mainly divides the image according to the features such as gray level, color, surface texture and geometric structure of the image. Using deep learning technology to extract plot boundaries is a hot research topic at home and abroad, and the extraction accuracy of plot boundaries is further improved through deep learning. However, most of the existing research on plot boundary extraction is aimed at single crop plots in plain areas, and there is less research on fuzzy plot boundaries in complex environments such as hilly and mountainous areas, especially lacking research on tropical crop planting areas in southern China. SUMMARY

[0004] The present application aims to at least solve one of the above-mentioned technical problems, and provides a method for extracting fuzzy boundaries of hilly agricultural plots in a hot region based on remote sensing images, which can realize intelligent extraction of fuzzy boundaries of hilly agricultural plots.

[0005] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: a method for extracting fuzzy boundaries of hilly agricultural plots in a hot region based on remote sensing images, comprising the following steps:

[0006] Step S1, acquiring remote sensing image data, using image processing, deep learning, statistical model, constructing plot boundary fuzzy index, grading and scaling the image according to the index, obtaining the image with obvious boundary features;

[0007] Step S2, using the spectrum, texture and spatial information of the remote sensing image, constructing a multi-branch cascade deep neural network model for plot boundary recognition;

[0008] Step S3, using the maximum value suppression binarization, connectivity detection and smoothness detection algorithm to post-process the recognition result, obtaining the clear boundary of the hilly farmland in the tropical region.

[0009] Preferably, in step S1, the specific steps of constructing the plot boundary fuzzy index and grading and scaling the image according to the fuzzy index include:

[0010] S11. Cut the remote sensing image into image blocks of uniform size, calculate the matrix variance of the Laplace operator of all image blocks, and divide the image into two parts according to the boundary clear and boundary fuzzy;

[0011] S12. Use the image of the boundary clear part to train the RCF boundary recognition model, and then infer the boundary fuzzy image based on the obtained model to obtain the boundary gray image, and analyze the boundary graph gray value distribution of all images;

[0012] S13. Combine the Laplace variance and the gray distribution, use the statistical model to construct the fuzzy degree index, and the calculation formula of the fuzzy degree index is as follows:

[0013] B=α×B G +(1-α)×B L (1)

[0014] In the formula, B is the fuzzy index of the plot boundary; α is used to weight the gray distribution index and the Laplace variance, and in actual operation, α is 0.7; B L is the Laplace fuzzy index; B G is the gray fuzzy index, wherein the calculation formula of B L is as follows:

[0015]

[0016] In the formula, B L is the Laplace fuzzy index; σ 2 is the Laplace operator matrix variance of the image block; v max and v min are the maximum and minimum variance values of the image set respectively,

[0017] σ 2 is calculated as follows:

[0018]

[0019] In the formula, σ 2 is the variance; M and N are the length and width of the image I respectively; L is the matrix of the image I And

[0020] By the logarithmic function, the larger variance value can be compressed, and the change becomes smooth, and 1 is added to avoid negative infinity or zero in logarithmic calculation. L The value of B G ranges between [0, 1], and the closer the value is to 1, the clearer the picture is; on the contrary, the closer the value is to 0, the more blurred the picture is.

[0021] B G The calculation formula is as follows:

[0022]

[0023] In the formula, B G is the gray blur index; P left , P mid and P right represent the wave peaks of the left, middle and right three gray values (0 to 255) in the image gray histogram respectively; C is a constant used to control the gray blur index to fall within [0, 1], which is set to 9 according to experience in practice;

[0024] The closer the value of B G is to 1, the clearer the picture is; on the contrary, the closer the value of B G is to 0, the more blurred the picture is.

[0025] S14. According to the blur index and the model recognition situation feedback, all images are classified, and the images with blurred field boundaries are scaled adaptively according to the classification.

[0026] Preferably, the step 2 of identifying the field boundary by the multi-branch cascaded deep neural network model comprises the following steps:

[0027] S21. Based on the semantic segmentation network, Mamba is introduced to divide the image into subsets of land types such as cultivated land, forest land, grassland, garden land, water area and construction land, and the division of other non-agricultural land blocks helps to improve the extraction accuracy of agricultural field boundaries. For different subsets, multiple weights are trained using the edge extraction network, and the corresponding weights are selected adaptively for reasoning during reasoning;

[0028] S22. The edge extraction network adopts a double-task branch structure, and uses the boundary information and attribute information of the field at the same time, wherein the boundary information is used to supervise the edge extraction task, and the attribute information is used to supervise the semantic segmentation task;

[0029] S23. Determine the plot boundaries of each land class using the trained fuzzy boundary extraction model library, and extract the boundaries of the hilly agricultural plots therefrom.

[0030] Preferably, the post-processing in step S3 specifically includes the following steps:

[0031] S31. Perform binaryzation using the NMS maximum value suppression algorithm, simplify the gray scale distribution, and thin the lines to facilitate subsequent processing;

[0032] S32. Perform connectivity detection based on the Region Growing algorithm, scan from the line breakpoints, consider the directionality, fit the nearby pixels to realize breakpoint connection;

[0033] S33. Apply the Douglas-Peucker algorithm to simplify the number of pixel points of the lines while maintaining the shape and characteristics of the original curve, realize the smoothing processing of the boundaries, and finally obtain the boundaries of the hilly farmland in the tropical region.

[0034] The beneficial effect is that, compared with the prior art, the method for extracting fuzzy boundaries of hilly agricultural plots in hot regions based on remote sensing images adopts a deep learning method, is based on high-resolution remote sensing images, and fully extracts and applies the spectral, texture, and spatial information of the images through steps such as developing a self-adaptive fuzzy boundary image scale algorithm, constructing a multi-branch cascaded deep neural network, and performing fine processing on the extraction results, to realize intelligent extraction of the fuzzy boundaries of hilly farmland. BRIEF DESCRIPTION OF DRAWINGS

[0035] The specific embodiments of the present application will be further described in detail below in combination with the accompanying drawings, in which:

[0036] Figure 1 The flowchart of the method for extracting fuzzy boundaries of hilly agricultural plots in hot regions based on remote sensing images of the present application;

[0037] Figure 2 The Laplacian variance distribution of the three different size sampling pictures of the present application;

[0038] Figure 3 The image boundary clear and fuzzy partition result graph of the present application;

[0039] Figure 4 The image boundary clear and fuzzy gray scale histogram example of the present application;

[0040] Figure 5 The hilly farmland boundary extraction result graph of the present application. DETAILED DESCRIPTION

[0041] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0042] It should be noted that when a component is referred to as being "fixed" to another component, it can be directly on the other component or there can be intervening components. When a component is referred to as being "connected" to another component, it can be directly connected to the other component or there can be intervening components. When a component is referred to as being "disposed" on another component, it can be directly on the other component or there can be intervening components. When a component is referred to as being "disposed in the middle", it is not only disposed in the middle position, but also within the range defined by the two ends. The terms "vertical", "horizontal", "left", "right", and similar terms used herein are for illustrative purposes only.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0044] As shown in Figure 1 The application discloses a method for extracting fuzzy boundaries of hilly agricultural plots in a hot zone based on remote sensing images, which comprises the following steps:

[0045] Step S1: Obtain remote sensing image data, use image processing, deep learning, and statistical models to construct a plot boundary fuzziness index, and perform grading and scaling on the image according to the index to obtain an image with obvious boundary features;

[0046] Step S2: Use the spectrum, texture, and spatial information of the remote sensing image to construct a multi-branch cascaded deep neural network model to identify plot boundaries;

[0047] Step S3: Use the algorithms of maximum value suppression binarization, connectivity detection, and smoothness detection to post-process the identification result to obtain clear boundaries of hilly farmland in the hot zone.

[0048] Specifically, in step S1, the specific steps of constructing a plot boundary fuzziness index and grading and scaling the image according to the fuzziness index include:

[0049] S11. Cut the remote sensing image into image blocks of uniform size (such as 1024*1024), calculate the matrix variance of the Laplace operator of all image blocks, and divide the image into two parts with clear boundaries and fuzzy boundaries;

[0050] S12. Train the RCF boundary recognition model using the image of the clear boundary part, and then infer the fuzzy boundary image based on the obtained model to obtain the boundary gray image, and analyze the boundary gray value distribution of all images;

[0051] S13. Combine the Laplace variance and the gray distribution, use a statistical model to construct a fuzzy degree index, and the calculation formula of the fuzzy degree index is as follows:

[0052] B = a x B G + (1-a) x B L (1)

[0053] In the formula, B is the fuzzy index of the land boundary; a is used to weight the gray distribution index and the Laplace variance, and a is 0.7 in actual operation; B L is the Laplace fuzzy index; B G is the gray fuzzy index, wherein the calculation formula of B L is as follows:

[0054]

[0055] In the formula, B L is the Laplace fuzzy index; σ 2 is the Laplace operator matrix variance of the image block; v max and v min are the maximum and minimum variance values of the image set respectively,

[0056] The calculation formula of σ 2 is as follows:

[0057]

[0058] In the formula, σ 2 is the variance; M and N are the length and width of the image I respectively; L is the matrix of the image I

[0059] Through the logarithmic function, the larger variance value can be compressed to make the change smooth, and 1 is added to avoid negative infinity or zero in logarithmic calculation. The value of B L ranges between 0 and 1, and the closer the value is to 1, the clearer the picture; on the contrary, the closer the value is to 0, the more fuzzy the picture.

[0060] The calculation formula of B G is as follows:

[0061]

[0062] In the formula, B G is a gray scale blur index; P left , P mid and P right respectively represent the wave peaks of left, middle and right gray scale values (0 to 255) in the image gray scale histogram; C is a constant used to control the gray scale blur index to fall within [0, 1], which is set to 9 according to experience in practice, B G The value is closer to 1, the clearer the picture is; on the contrary, the value is closer to 0, the more blurred the picture is. G

[0063] S14. According to the blur index and the model recognition situation feedback, all images are classified, and the images with blurred land boundaries are scaled adaptively according to the classification.

[0064] As shown in the Laplace variance distribution of the three different size sampling pictures in Figure 2 . In order to divide the image samples into clear boundary and blurred boundary two parts for subsequent boundary extraction model training, the present application first cuts the remote sensing images into image blocks of uniform size (such as 1024*1024), then calculates the matrix variance of the Laplace operator of all image blocks, and finally constructs the Laplace blur index B 2 according to the calculated variance σ max and the maximum variance value v min and the minimum variance value v L of the image set. Among them, the maximum variance value v max and the minimum variance value v min are obtained by sampling the variances of three different size images, as shown in the figure.

[0065] As shown in the image boundary clear and fuzzy partition result graph in Figure 3 . The value of the Laplace blur index B L ranges between [0, 1], the closer to 1, the clearer the picture is; on the contrary, the closer to 0, the more blurred the picture is. According to the calculation result of the index B L , the images can be divided into clear boundary and blurred boundary two parts, as shown in the figure. The image with blurred boundary cannot directly extract the boundary of the land block, and it is difficult to distinguish the specific content and distribution range of various land types; while the image with clear boundary can directly extract the boundary of the land block, and the distribution range of various land types is relatively clear. Dividing the image samples into blurred and clear can be used as the input of the subsequent boundary recognition model, and provides the basis for constructing the blur degree index.

[0066] As shown in the image boundary clear and fuzzy partition result graph in Figure 4 ​The illustrated image boundary clear and fuzzy gray histogram examples. Using the image of the clear part of the boundary to train the RCF boundary recognition model, and then based on the resulting model to infer the boundary fuzzy image, get the boundary gray image. As shown in the histogram, the horizontal axis represents the gray value (0 to 255), the vertical axis represents the number of pixels with the gray value. Analyzing the boundary map gray value distribution of the image, the clearer the picture, the more accurate the boundary extraction result, the performance on the gray image is that the gray pixels are few, and the black and white pixels are many. Conversely, the more fuzzy the picture, the more gray pixels on the gray image. By extracting the left, middle and right three gray values (0 to 255) of the image gray histogram P left mid right The gray fuzzy index B G is constructed, which provides the premise for the calculation of the fuzzy index B of the land boundary.

[0067] Specifically, in step 2, the multi-branch cascaded deep neural network model for land boundary recognition includes the following steps:

[0068] S21. Based on the semantic segmentation network, introduce Mamba to divide the remote sensing image into subsets of cultivated land, forest land, grassland, garden land, water area, construction land and other land types. The division of other non-agricultural land blocks helps to improve the extraction accuracy of agricultural land boundary. For different subsets, use the edge extraction network to train multiple weights, and adaptively select the corresponding weight for inference during inference;

[0069] S22. The edge extraction network adopts a double-task branch structure, which simultaneously uses the boundary information and attribute information of the land block. Among them, the boundary information is used to supervise the edge extraction task. Since there is no edge position belongs to the same land object, the edge extraction task can alleviate the problem of broken segmentation results of the semantic segmentation task. The attribute information is used to supervise the semantic segmentation task. Since there is no edge inside the land object, the position closer to the land boundary is more likely to be an edge, so the semantic segmentation task can alleviate the pseudo-edge problem of the edge extraction task;

[0070] S23. Use the trained fuzzy boundary extraction model library to determine the land boundary of each land type, and extract the boundary of the hilly agricultural land block from it.

[0071] Specifically, the post-processing in step S3 specifically includes the following steps:

[0072] S31. The boundary result identified by the deep neural network is a gray image, which needs to be post-processed before it can be used as a land boundary. Use NMS maximum value suppression algorithm for binaryzation, simplify gray distribution, and thin lines to facilitate subsequent processing;

[0073] ​​S32. Connectivity detection is performed based on the Region Growing algorithm. The algorithm scans from the line breakpoint and considers directional fitting of nearby pixels to achieve breakpoint continuation.

[0074] S33. The Douglas-Peucker algorithm is used to simplify the number of pixels of the lines while preserving the shape and features of the original curves, thereby achieving smooth boundary processing and finally obtaining the boundary of hilly farmland in tropical regions.

[0075] like Figure 5 The diagram shows the results of extracting the boundaries of hilly farmland in tropical regions. As can be seen from the original image, agricultural land in this area, including woodland, grassland, and farmland, is densely vegetated, resulting in blurred and unclear boundaries. This makes it difficult to distinguish the specific distribution range of different land use types and interferes with further crop identification. After initial identification using a trained neural network, the boundaries of the plots in the image were initially obtained. Although the boundaries at this stage had some outline, they were still grayscale images and relatively blurry. Therefore, based on the previous step, algorithms such as maximum suppression binarization, connectivity detection, and smoothness detection were further used to post-process the identification results, ultimately obtaining clear boundaries of hilly farmland in tropical regions. This provides a foundation for accurate identification of various crop types and helps relevant departments in the region to promptly grasp local agricultural information and promote intelligent land management.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of the technical solutions of the present invention.

Claims

1. A method for extracting fuzzy boundaries of hilly agricultural field based on remote sensing images, characterized in that, The method comprises the following steps: Step S1, acquiring remote sensing image data, using image processing, deep learning, and statistical models to construct a land parcel boundary blur index, classifying and scaling the image according to the index to obtain an image with obvious boundary features, the specific steps comprising: S11. Cutting the remote sensing image into image blocks of uniform size, calculating the matrix variance of the Laplace operator of all image blocks, and dividing the image into two parts according to the boundary clarity and boundary blur; S12. Using the image of the boundary clear part to train the RCF boundary recognition model, and then based on the obtained model to infer the boundary fuzzy image to obtain a boundary gray image, and analyze the boundary graph gray value distribution of all images; S13. Combining the Laplace variance and the gray distribution, using a statistical model to construct a blur degree index, the calculation formula of the blur degree index is as follows: （1） where B is a blur index of the parcel boundary; For weighting the gray scale distribution index and Laplacian variance, in practice is 0.7; is a Laplacian blur index; is a gray scale blur index, where The calculation formula of B is as follows: （2） wherein is the Laplace blur index; is the Laplace operator matrix variance of the image block; and are the maximum and minimum variance values of the image set, respectively, The calculation formula is as follows: （3） wherein is the Laplacian matrix variance of the image block; M and N are the length and width of the image , respectively; L is the matrix of the image ; and ; i and j represent the row and column position of the pixel in the image, respectively. The calculation formula is as follows: （4） wherein, is the gray scale blur index; , and represent the left, middle and right peaks of the gray scale histogram of the image, respectively; is a constant used to control the gray scale blur index to fall between and S14. According to the blur index and the model recognition situation feedback, all images are classified, and the land parcel boundary fuzzy image is scaled according to the level; Step S2, using the spectrum, texture and spatial information of the remote sensing image, a multi-branch cascaded deep neural network model is constructed for land parcel boundary recognition; Step S3, using the maximum value suppression binarization, connectivity detection, and smoothness detection algorithm to post-process the recognition result to obtain the clear boundary of the hilly farmland in the tropical region. 2.The method of claim 1, wherein, The multi-branch cascaded deep neural network model in step 2 for land parcel boundary recognition comprises the following steps: S21. Based on the semantic segmentation network, Mamba is introduced to divide the remote sensing image into different subsets of land classes, and for different subsets, multiple weights are trained using the edge extraction network, and the corresponding weight is selected for inference during inference; S22. The edge extraction network adopts a double-task branch structure, which simultaneously uses the boundary information and attribute information of the land parcel, wherein the boundary information is used to supervise the edge extraction task, and the attribute information is used to supervise the semantic segmentation task, and since there is no edge inside the ground object; S23. Using the trained fuzzy boundary extraction model library, the land parcel boundaries of each land class are determined, and the boundaries of the hilly agricultural land parcel are extracted. 3.The method of claim 1, wherein the method further comprises: determining a slope of the contour line; and determining whether the slope is greater than a predetermined threshold value. The post-processing in step S3 specifically comprises the following steps: S31. Using the NMS maximum value suppression algorithm for binarization to simplify the gray distribution and thin the lines for subsequent processing; S32. Based on the Region Growing algorithm for connectivity detection, scanning from the line breakpoint, considering the directionality to fit the nearby pixels to realize the breakpoint connection; S33. Using the Douglas-Peucker algorithm to simplify the number of line points while maintaining the shape and characteristics of the original curve to realize the smoothing of the boundary, and finally obtaining the boundary of the hilly farmland in the tropical region.

Images