Farmland extraction method of high-resolution remote sensing images based on multi-level semantic boundary segmentation

Through the multi-level semantic boundary segmentation method, the edge detection network and Zhang-Suen algorithm are used to solve the problem of incomplete farmland boundaries, and the high-precision and high-integrity farmland extraction effect is achieved.

CN114219815BInactive Publication Date: 2024-11-08ZHEJIANG UNIV OF TECH +1
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202111304541.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-05
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the prior art, in farmland extraction, the degree of fit between the farmland boundaries and the actual boundaries is low. Improper threshold setting leads to incomplete boundaries or the retention of non-agricultural boundaries, and it is impossible to accurately extract complete farmland plots.

Method used

Using a multi-level semantic boundary segmentation method, an edge detection network is used to generate a farmland edge intensity map, segmentation and refinement through gradually decreasing thresholds, and combining with the Zhang-Suen algorithm to achieve boundary completion and refinement, and finally construct a complete farmland vector. Object.

Benefits of technology

The positioning accuracy and integrity of farmland boundaries are improved, ensuring that the extraction results are in line with the real boundaries, and a complete farmland plot is formed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114219815B_ABST
    Figure CN114219815B_ABST
Patent Text Reader

Abstract

A method for extracting farmland from high-resolution remote sensing images with multi-level semantic boundary segmentation includes: Step 1: Extracting farmland semantic boundaries: Step 2: Setting farmland boundary thresholds: Step 3: Threshold segmentation of farmland boundaries: Step 4: Refinement of farmland boundaries: Step 5: Iterative segmentation and refinement of farmland boundaries: Step 6: Construction of farmland vector objects. The present invention uses the boundary strength map obtained by the edge detection network as input, and adopts a multi-level threshold segmentation method to extract fine boundaries. Among them, the high threshold segmentation result can ensure that the segmentation result fits the real boundary as much as possible, and the low threshold makes the constructed ground object as complete as possible. Thereby, the refinement result can maintain a high boundary positioning accuracy and have good integrity.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image processing, and in particular to a method for extracting farmland from high-resolution remote sensing images based on semantic boundaries. Background Art

[0002] Cultivated land distribution information is one of the important agricultural information, which is of great value for applications such as farmland growth detection, crop yield prediction, and agricultural risk assessment (Liu Wei, Wu Zhifeng, Luo Jiancheng, et al. A method for extracting high-resolution remote sensing information of cultivated land in hilly and mountainous areas based on deep learning [J]. Acta Geodaetica et Cartographica Sinica, 2021, 50(01): 105-116.). Remote sensing images have the advantages of wide coverage, easy data acquisition, and low price. Remote sensing images can quickly and accurately extract cultivated land distribution information, which is an important means to determine cultivated land distribution information.

[0003] Traditional farmland extraction methods based on remote sensing images use spectral, spatial and texture features, combined with various classification algorithms (such as support vector machine method and random forest method) to identify farmland plots (Jing S, Liu J, Lin X, et al. Cropland Extraction from Very High Spatial Resolution Satellite Imagery by Object-Based Classification Using Improved Mean Shift and One-Class Support Vector Machines [J]. Sensor Letters, 2011, 9 (3): 997-1005.). Due to the diversity of crop types in farmland, their growth cycles and farmland management methods are different, resulting in the spectral characteristics of farmland in remote sensing images being varied and the texture characteristics being complex, which is easily confused with other land types, and the extraction results cannot meet the actual extraction accuracy requirements.

[0004] In recent years, deep neural networks have been gradually used for farmland extraction because they can change the connection structure and weight of the neural network according to sample information, so that features can be adjusted adaptively. With the increase in the number of neural network layers, the semantic information from shallow features to deep features is gradually enriched, which has better effects in farmland identification and classification. Farmland extraction generally uses a semantic segmentation network with an encoding-decoding structure. Typical networks include U-Net (Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation [C] / / International Conference on Medical Image Computing and Computer-assisted Intervention, 2015: 234-241.) This network classifies each pixel in the image (such as cultivated land, grassland or river, etc.), and then aggregates objects of the same category to form plot objects. However, due to the multi-level convolution, the image cannot accurately correspond to the original image, and the farmland extraction results have problems such as low fit between the farmland boundary and the actual boundary.

[0005] In this regard, the edge detection network can accurately locate the boundary of the object. The RCF edge detection network (LiuYun, Cheng Ming-Ming, et al. Richer Convolutional Features for Edge Detection. [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(8).) can also be applied to farmland detection methods. This method convolves the original image with a specifically constructed convolution kernel to obtain a feature map reflecting the edge of the farmland in the image. Then, the boundary of a specific intensity is extracted according to the set threshold, and then its skeleton is extracted to finally obtain the farmland boundary. This method solves the problem of low fit between the farmland boundary and the actual boundary in the traditional semantic segmentation network.

[0006] However, current contour-oriented farmland extraction methods all use threshold segmentation combined with thinning methods to obtain farmland objects. Therefore, threshold segmentation is a key issue that needs to be solved in farmland extraction. If the threshold is set too large, the edge information below the threshold will be lost, the boundary will be broken and incomplete, and a complete farmland plot cannot be constructed; if the threshold is too small, a large number of non-agricultural boundaries will be retained in the segmentation result, and the thinning algorithm generally retains the skeleton of the object, resulting in it being unable to accurately fit the real plot boundary. Summary of the invention

[0007] The present invention aims to overcome the above-mentioned limitations of the prior art and proposes a method for extracting farmland from high-resolution remote sensing images by multi-level semantic boundary segmentation.

[0008] Firstly, the edge detection network is used to extract the edge of farmland and generate the edge intensity map of farmland. The edge intensity of farmland is segmented by successively decreasing thresholds, and pixels greater than the threshold are retained as edges, and pixels less than the threshold are retained as non-edges. Then, the edge of farmland is refined to complete the incomplete boundary and obtain a more complete boundary. Finally, the farmland vector object is constructed according to the refinement result to realize the extraction of farmland.

[0009] The present invention uses the boundary strength map obtained by the edge detection network as input and adopts a multi-level threshold segmentation method to extract fine boundaries. Among them, the high threshold segmentation result can ensure that the segmentation result fits the real boundary as much as possible, and the low threshold makes the constructed ground object as complete as possible. Therefore, the refinement result can maintain a high boundary positioning accuracy and have good integrity.

[0010] According to the above principles, the high-resolution remote sensing image farmland extraction method of the present invention with multi-level semantic boundary segmentation specifically comprises the following steps:

[0011] Step 1: Extraction of farmland semantic boundaries:

[0012] In this step, edge detection networks (such as Bi-Directional Cascade Network, BDCN) are used to extract farmland boundaries from remote sensing images and obtain farmland boundary intensity maps K, where the boundary intensity range of K is [0,255]. In the boundary intensity map, the boundary probability increases with the value k. i ∈K increases. When the boundary strength value of a pixel is 0, it indicates that the possibility that the pixel is a boundary is the lowest; and when the boundary strength value is 255, it indicates that the possibility that the pixel is a boundary is the highest.

[0013] Step 2: Setting the farmland boundary threshold:

[0014] In this step, multiple thresholds are selected from large to small to obtain a threshold set.

[0015] First, a larger value is selected as the first threshold g1, and then each time the value d (d < 255, d ∈ N * ) as the step size, multiple thresholds are selected in the interval [0,255], and the i-th threshold is expressed as:

[0016] g i =g1-(i-1)×d,0≤g i <g1≤255 (1)

[0017] Get a multi-threshold set G that is an arithmetic decreasing sequence:

[0018] G={g1,g2,g3,…,g m} (2)

[0019] Among them, the threshold satisfies g i >g i+1 ,i=1,2,..,m-1, m represents the number of thresholds.

[0020] Step 3: Field boundary threshold segmentation:

[0021] According to the selected g i ∈G performs threshold segmentation on the farmland boundary intensity map K obtained in step 1 to obtain the threshold segmentation map p i ,

[0022]

[0023] p i (r,c) represents the threshold segmentation map P i Whether the rth row and cth column in the image is an edge, the boundary strength value is greater than or equal to g i The pixel intensity value is set to 255, indicating the boundary; the boundary intensity value is less than g i The pixel value is set to 0, indicating a non-boundary. This step results in a binary image consisting of 0 and 255.

[0024] Step 4: Field boundary refinement:

[0025] This step uses the thinning algorithm to refine the binary image p obtained in step 3. i Thinning is performed, and the thinning algorithm uses the Zhang-Suen algorithm. This step obtains a single-pixel width boundary, which is represented by 255; the non-boundary pixel intensity value is 0, and the thinned boundary of the entire image is represented by T i .

[0026] During refinement, the new refinement result retains the previous refinement result, that is, T i Contains the refinement result T of the i-1th round i-1 , which can be expressed as:

[0027]

[0028] The thinning operation uses a pixel width as the boundary line width for topological connection, using the binary map p i Relative to p i-1 Better connectivity, connected T i-1 The discontinuous boundaries in the image are improved to improve the integrity of the boundaries.

[0029] Step 5: Iterative segmentation and refinement of farmland boundaries:

[0030] This step is based on the multi-threshold set g i ∈G, i=1,2,...,m, segment and refine according to steps 3 to 4. i ∈G, i=1,2,...,m, the resulting farmland boundary refinement graph set T is:

[0031] T={T1,T2,...,T m} (5)

[0032] Among them, T i ∈T represents the threshold g i ∈G is the boundary obtained by segmentation and refinement. Since the new refined boundary can maintain the previous refinement result, it satisfies:

[0033]

[0034] Better boundary fit and integrity are achieved through layer-by-layer refinement.

[0035] Step 6: Farmland vector object construction:

[0036] This step is the last threshold g obtained in step 5. m Get the farmland boundary refinement result T m Perform vectorization operation, that is, use the boundary lines in the complete farmland boundary map as the outer frame outline, and use the closed area formed by the outer frame outline as the farmland plot to complete the construction of the farmland vector object in the study area.

[0037] The advantages of the present invention are: the farmland boundary intensity map is segmented in sequence by thresholds from high to low, wherein the high threshold refinement result closely fits the farmland boundary, the low threshold refinement result improves its integrity, and the high to low threshold segmentation results realize information fusion, so that a complete and accurate farmland boundary can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the method of the present invention.

[0039] Figure 2 The image is the input image of the present invention. The image adopts a 0.5-meter Google Earth image, and the selected area is Tongxiang City, Jiaxing City, Zhejiang Province.

[0040] Figure 3 It is the implementation process of the present invention. Figure 3 (a)~ Figure 3 (m) are the same parts of Tongxiang area. Figure 3 (a) is the Tongxiang image of the area; Figure 3(b) is the edge intensity map of some farmlands obtained after the image is input into the edge detection network, with an edge intensity range color band, the lowest intensity value is 0 and the highest is 255; Figure 3 (c) is the threshold segmentation result of the edge intensity map when the threshold is set to 200; Figure 3 (d) is the result image obtained after the segmentation image with the threshold of 200 is refined; Figure 3 (e) The vector graph obtained after vectorization with a threshold of 200 is displayed on the map; Figure 3 (f) shows the threshold segmentation result of the edge intensity map when the threshold is set to 150; Figure 3 (g) is the result image obtained after the threshold segmentation image of 150 is refined; Figure 3 (h) The vector graph obtained after vectorization with a threshold of 150 is displayed on the map; Figure 3 (i) is the threshold segmentation result of the edge intensity map when the threshold is set to 100; Figure 3 (j) The result image obtained after the segmentation image with the threshold of 100 is refined; Figure 3 (k) The vector graph obtained after vectorization with a threshold of 100 is displayed on the map; Figure 3 (l) is the result obtained by fusion of all threshold refinement result images; Figure 3 (m) shows the effect of a certain number of objects on the map after the fusion result map is vectorized.

[0041] Figure 4 It is the farmland object constructed by the present invention. Figure 4 (a) is a vector result map of the entire Tongxiang area obtained by the method of the present invention; Figure 4 (b)~ Figure 4 (d) shows the vector results of three local areas in Tongxiang. Figure 4 The locations of these three areas in Tongxiang are marked in (a). DETAILED DESCRIPTION

[0042] In order to facilitate those skilled in the art to understand and implement the present invention, the following embodiments and appended Figure 1 The invention process of the present invention is further described in detail. It should be understood that the specific embodiments described herein are only used to explain the present invention, not to limit the present invention, and the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Therefore, based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0043] The study area of ​​this embodiment is Tongxiang City, Jiaxing City, Zhejiang Province. The study area uses a 0.5-meter resolution Google Earth remote sensing image (see attached Figure 2 shown).

[0044] The present invention provides a method for extracting farmland from high-resolution remote sensing images based on multi-level semantic boundary segmentation, and its specific implementation is as follows:

[0045] Step 1: Extraction of farmland semantic boundaries:

[0046] In this embodiment, the remote sensing image of Tongxiang (as shown in the attached Figure 3 (a) Input the trained BDCN edge detection network for detecting farmland edges to obtain the farmland boundary intensity map of the study area, with the boundary intensity range of [0,255]. Figure 3 (b) The farmland edge intensity result map obtained after a portion of the Tongxiang image was input into BDCN. The intensity value of the pixel is proportional to the grayscale value, that is, the larger the intensity value of the pixel, the darker the pixel display color.

[0047] Step 2: Setting the farmland boundary threshold:

[0048] Using the value range [0,255] of the farmland boundary intensity map obtained in step 1, select multiple thresholds. The first threshold selected is 250, and then the threshold is selected from the boundary intensity range [0,255] with a difference of 10 each time until the threshold is 0, and an arithmetic decreasing multi-threshold set is obtained, and the number of thresholds in the set is 25. The subsequent steps are explained using the threshold 200 as an example.

[0049] Step 3: Field boundary threshold segmentation:

[0050] The farmland boundary intensity map obtained in step 1 (see attached) is compared with the set threshold of 200. Figure 3 (b) Threshold segmentation is performed. Each pixel in the intensity map is judged, and the intensity value of the pixel with an intensity not less than 200 is set to 255 (i.e., boundary), and the intensity value of the pixel with an intensity less than 200 is set to 0 (i.e., non-boundary), and a binary map composed of 0 and 255 is obtained, as shown in the attached figure. Figure 3 (c) Threshold segmentation result.

[0051] Step 4: Field boundary refinement:

[0052] The threshold segmentation result diagram (attached) when the threshold value is 200 obtained in step 3 is obtained by using the "Zhang-Suen" thinning algorithm Figure 3 (c)) is refined while retaining the refinement result when the threshold value was 210. Figure 3(d) shows the farmland boundary refinement result map, which is also a binary map composed of 0 and 255. The refinement result map retains the refinement result when the threshold is 210, and uses the better connectivity and more detailed features of the current threshold segmentation result map to connect the discontinuous boundaries in the refinement result of the 210 threshold, and obtain a more complete refinement result connected by a pixel width.

[0053] Step 5: Iterative segmentation and refinement of farmland boundaries:

[0054] Select the next threshold 190 in the threshold set as the new threshold, repeat steps 3 to 4, perform a new round of threshold segmentation and thinning operations, and obtain a farmland boundary thinning result map when the threshold is 190. After that, continue to select the next new threshold from the threshold set until the last threshold 0 in the set is also selected to perform threshold segmentation and thinning operations, and obtain the farmland boundary thinning result map corresponding to each threshold in the multi-threshold sequence (as shown in the attached figure). Figure 3 (c) Attachment Figure 3 (k)).

[0055] Step 6: Farmland vector object construction:

[0056] The final farmland boundary refinement result with complete boundaries obtained in step 5 (as shown in the attached Figure 3 (l)) is vectorized, that is, the boundary lines in the obtained complete farmland boundary map are used as the outer frame outline, and the closed area formed by the outer frame outline is used as the farmland plot, as shown in the attached figure. Figure 3 (j) Farmland vector objects of the study area shown.

[0057] The above process mainly illustrates the specific implementation process of the method used in the present invention.

[0058] This embodiment uses the farmland vector object obtained by the method of the present invention, and a partial display is shown in the attached figure. Figure 4 (b) Attachment Figure 4 (d) From the attached Figure 4 (b) Attachment Figure 4 It can be seen from (d) that the boundary integrity of the farmland boundary extracted by the method of the present invention is improved, which is helpful to obtain a complete farmland boundary, and the extracted farmland boundary basically overlaps with the actual farmland boundary, which illustrates the effectiveness of the method of the present invention.

[0059] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also provides equivalent technical means that can be thought of by those skilled in the art based on the inventive concept.

Claims

1. A method for extracting farmland from high-resolution remote sensing images using multi-level semantic boundary segmentation. The following steps are involved: Step 1: Extraction of farmland semantic boundaries: The edge detection network Bi-Directional Cascade Network is used to extract the farmland boundaries of remote sensing images, and the farmland boundary intensity map K is obtained. The boundary intensity range of K is [0,255]. In the boundary intensity map, the boundary possibility increases with the value k. i ∈K increases. When the boundary strength value of a pixel is 0, it indicates that the possibility that the pixel is a boundary is the lowest; and when the boundary strength value is 255, it indicates that the possibility that the pixel is a boundary is the highest. Step 2: Setting the farmland boundary threshold: Select multiple thresholds from large to small to obtain a threshold set; First, select the first threshold g 1 , and then, each time the value d is used as the step size, where d < 255, d ∈ N * , select multiple thresholds in the interval [0,255], and the i-th threshold is expressed as: g i =g 1 -(i-1)×d,0≤g i <g 1 ≤255 (1) Get a multi-threshold set G that is an arithmetic decreasing sequence: G={g 1 ,g 2 ,g 3 ,...,g m } (2) Among them, the threshold satisfies g i >g i+1 ,i=1,2,..,m-1, m represents the number of thresholds; Step 3: Field boundary threshold segmentation: According to the selected g i ∈G performs threshold segmentation on the farmland boundary intensity map K obtained in step 1 to obtain the threshold segmentation map p i , p i (r,c) represents the threshold segmentation map P i Whether the rth row and cth column in the image is an edge, the boundary strength value is greater than or equal to g i The pixel intensity value is set to 255, indicating the boundary; the boundary intensity value is less than g i The pixel value is set to 0, indicating a non-boundary; the resulting image is a binary image consisting of 0 and 255; Step 4: Field boundary refinement: Using the thinning algorithm, the binary image p obtained in step 3 is i Thinning is performed using the Zhang-Suen algorithm. This step obtains a single-pixel-width boundary, represented by 255. The intensity value of non-boundary pixels is 0, and the thinned boundary of the entire image is represented by T i ; During refinement, the new refinement result retains the previous refinement result, that is, T i Contains the refinement result T of the i-1th round i-1 , which can be expressed as: The thinning operation uses a pixel width as the boundary line width for topological connection, using the binary map p i Relative to p i-1 Better connectivity, connect T i-1 Discontinuous boundaries in the structure improve the integrity of the boundaries; Step 5: Iterative segmentation and refinement of farmland boundaries: According to the multi-threshold set g i ∈G, i=1,2,...,m, segment and refine according to steps 3 to 4; get the threshold value according to the threshold g i ∈G, i=1,2,...,m, the resulting farmland boundary refinement graph set T is: T={T 1 ,T 2 ,…,T m } (5) Among them, T i ∈T represents the threshold g i ∈G is the boundary obtained by segmentation and refinement. Since the new refined boundary can maintain the previous refinement result, it satisfies: Get better boundary fit and integrity through layer-by-layer refinement; Step 6: Farmland vector object construction: For the last threshold g obtained in step 5 m Get the farmland boundary refinement result T m Perform vectorization operation, that is, use the boundary lines in the complete farmland boundary map as the outer frame outline, and use the closed area formed by the outer frame outline as the farmland plot to complete the construction of the farmland vector object in the study area.

Citation Information

Patent Citations

  • Semantic edge dominated high-resolution remote sensing image segmentation method

    CN110555857A

  • Image automatic segmentation method and device based on deep learning edge detection

    CN110570440A

  • AUPR891701A0