A method and system for automatically extracting cultivated land plots from remote sensing images
By improving the semantic segmentation network and watershed segmentation algorithm, combined with the Douglas-Puk algorithm, the problem of low extraction accuracy of cultivated land plots in complex scenarios is solved, and high-precision land boundary recognition and extraction are achieved.
Patent Information
- Application Number
- CN202210071016.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-21
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-01-21
AI Technical Summary
The existing technology has low accuracy in extracting arable land plots in complex and changing scenarios. Traditional methods cannot achieve accurate boundary recognition. The deep convolutional network has weak ability to extract linear objects features, resulting in discontinuity of boundaries and cannot be applied to large scene plots. extract.
By improving the semantic segmentation network, adding shallow modules to identify linear information, using the boundary feature extraction model to obtain the boundary probability of each pixel point, combining the watershed segmentation algorithm to process oversegment images, determine the boundary of the plot, and finally using the Douglas-Puk algorithm Perform linearization to improve the continuity of boundary recognition.
It improves the continuity and extraction accuracy of plot boundary recognition, and is suitable for arable land plot extraction in complex and variable scenarios.
Smart Images

Figure CN114419443B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of remote sensing image farmland plot extraction, and in particular to a remote sensing image farmland plot automatic extraction method and system. Background Art
[0002] The plot extraction based on traditional detection is carried out through simple edge detection or region segmentation methods, that is, using traditional edge detection operators to depict the boundaries of plots, or using watershed segmentation, multi-resolution segmentation and other methods to perform object-oriented segmentation in software such as eCognition. However, both methods have their own limitations. The edge operator can quickly divide the image into boundaries and non-boundaries by detecting discontinuities between pixels to identify edge strength, but the identified boundaries are prone to breakage, resulting in discontinuities in the identified boundaries, and it is impossible to achieve the transformation from line to surface from the plot boundary to the plot area; in addition, most edge detectors only detect gradient information and are sensitive to noise, which can also easily lead to boundary recognition errors; object-oriented segmentation methods merge pixels into regions through a certain homogeneity standard, but there may be large differences within the farmland plots, which will cause over-segmentation, and this type of method often stops growing before reaching the actual boundary, resulting in a single narrow polygon on the edge of the boundary. In addition, due to the limitations of the method itself, the traditional method can only identify the shallow and intuitive features of cultivated land, so the accuracy is poor and it is only suitable for simple and small areas. It is not suitable for the extraction of cultivated land plots in complex and changeable large scenes.
[0003] To extract cultivated land in complex and changing scenarios, it is necessary to be able to express the deeper features of cultivated land. Deep convolutional neural networks make this possible. In recent years, many methods for extracting cultivated land have emerged. However, it is worth noting that in the process of extracting features of cultivated land areas, convolution operations need to capture long-term spatial dependencies. These dependencies help identify classes, but they lose location information and the boundary information of objects cannot be accurately determined. As a result, this type of extraction method can only obtain the overall range of cultivated land areas, but cannot achieve the extraction of cultivated land plots that require precise boundaries.
[0004] In addition, there are also many studies that use deep convolutional networks to obtain land boundary information. This type of research regards land boundary as a type of land feature and performs semantic recognition by learning the image features of land boundary. However, deep convolutional networks have weak feature extraction capabilities for linear objects, and the recognized boundaries are discontinuous, resulting in low recognition accuracy. Summary of the invention
[0005] The purpose of the present invention is to provide a method and system for automatically extracting cultivated land plots from remote sensing images, which can improve the extraction accuracy of automatically extracting cultivated land plots from remote sensing images.
[0006] To achieve the above object, the present invention provides the following solutions:
[0007] A method for automatically extracting cultivated land plots from remote sensing images, comprising:
[0008] Obtain remote sensing images of the study area;
[0009] Each pixel value of the remote sensing image of the study area is input into the boundary feature extraction model to obtain the boundary probability of each pixel point in the remote sensing image of the study area; the boundary feature extraction model is obtained by training the improved semantic segmentation network by using the labeled historical remote sensing image of the study area; the improved semantic segmentation network is obtained by adding a shallow module to the semantic segmentation network; the shallow module is used to identify linear information;
[0010] According to the boundary probability of each pixel in the remote sensing image of the study area, the over-segmented image of the cultivated land plot is determined;
[0011] The over-segmented images of cultivated land are processed using the watershed segmentation algorithm to obtain the over-segmented area set and the common boundary set.
[0012] Determining a land parcel boundary recognition amount according to the over-segmented region set and the common boundary set;
[0013] The cultivated land plots in the remote sensing image of the study area are divided according to the plot boundary identification amount.
[0014] Optionally, before acquiring the remote sensing image of the study area, the method further includes:
[0015] Obtain historical remote sensing images of the study area;
[0016] The boundary probability of the pixel points at the boundary in the historical remote sensing image of the study area is marked as 1, and the boundary probability of the pixel points at the non-boundary in the historical remote sensing image of the study area is marked as 0, so as to obtain the marked historical remote sensing image of the study area;
[0017] The pixel values of the historical remote sensing images marking the study area are used as input, and the boundary probability of the pixels is used as output, and the improved semantic segmentation network is trained to obtain the boundary feature extraction model.
[0018] Optionally, determining the over-segmented image of the cultivated land plot according to the boundary probability of each pixel point in the remote sensing image of the study area specifically includes:
[0019] The boundary probability of each pixel in the remote sensing image of the study area is taken as the gray value of the corresponding pixel to construct a boundary probability image;
[0020] All gray values in the boundary probability image are binarized to obtain the over-segmented image of the cultivated land.
[0021] Optionally, the watershed segmentation algorithm is used to process the over-segmented image of the cultivated land to obtain the land boundary recognition amount, specifically including:
[0022] Let the number of iterations n be 1;
[0023] The nth level structure hierarchy is constructed with the study area as the root node and the over-segmented regions in the over-segmented region set as leaf nodes;
[0024] Determine a weight value of each common boundary in the common boundary set; the weight value is an average value of boundary probabilities at all pixel points in the common boundary;
[0025] Merge the two sub-regions adjacent to the common boundary with the smallest weight value to obtain an updated over-segmented region set and an updated common boundary set, increase the value of the iteration number n by 1, and return to the step of "taking the study area as the root node and the over-segmented regions in the over-segmented region set as leaf nodes to construct the n-th layer of the structural hierarchy" until the iteration number n reaches the hierarchy threshold, and determine that the plot boundary corresponding to the n-th layer of the structural hierarchy is the plot boundary identification amount; the over-segmented region includes multiple sub-regions; the number of sub-regions adjacent to the common boundary is greater than one; the number of over-segmented regions adjacent to the non-common boundary is 1 and the number of sub-regions adjacent to the non-common boundary is 1.
[0026] Optionally, dividing the cultivated land plots in the remote sensing image of the study area according to the identification amount of the plot boundaries specifically includes:
[0027] Converting the land parcel boundary identification amount into land parcel boundary vector data;
[0028] Performing denoising on the land parcel boundary vector data to obtain denoised land parcel boundary vector data;
[0029] The Douglas-Peucker algorithm is used to linearize the denoised plot boundary vector data to obtain the linearized plot boundary vector data;
[0030] The cultivated land plots in the remote sensing image of the study area are divided according to the linearized plot boundary vector data.
[0031] A remote sensing image cultivated land automatic extraction system, comprising:
[0032] The remote sensing image acquisition module of the study area is used to obtain the remote sensing image of the study area;
[0033] A boundary probability determination module is used to input each pixel value of the remote sensing image of the study area into a boundary feature extraction model to obtain the boundary probability of each pixel point in the remote sensing image of the study area; the boundary feature extraction model is obtained by training an improved semantic segmentation network using the labeled historical remote sensing image of the study area; the improved semantic segmentation network is obtained by adding a shallow module to the semantic segmentation network; the shallow module is used to identify linear information;
[0034] The over-segmented image determination module of cultivated land is used to determine the over-segmented image of cultivated land according to the boundary probability of each pixel point in the remote sensing image of the study area;
[0035] The module for determining the land parcel boundary recognition quantity is used to process the over-segmented image of the cultivated land parcel using the watershed segmentation algorithm to obtain the land parcel boundary recognition quantity;
[0036] A set extraction module is used to process the over-segmented images of cultivated land plots using a watershed segmentation algorithm to obtain an over-segmented region set and a common boundary set;
[0037] A land parcel boundary recognition amount determination module, used to determine a land parcel boundary recognition amount according to the over-segmented region set and the common boundary set;
[0038] The cultivated land parcel extraction module is used to divide the cultivated land parcels in the remote sensing image of the study area according to the parcel boundary recognition quantity.
[0039] Optionally, the system further includes:
[0040] Historical remote sensing image acquisition module, used to obtain historical remote sensing images of the study area;
[0041] A marking module is used to mark the boundary probability of pixel points at the boundary of the historical remote sensing image of the study area as 1, and mark the boundary probability of pixel points at the non-boundary of the historical remote sensing image of the study area as 0, so as to obtain a marked historical remote sensing image of the study area;
[0042] The boundary feature extraction model determination module is used to take the pixel value of the historical remote sensing image marking the study area as input and the boundary probability of the pixel as output, train the improved semantic segmentation network, and obtain the boundary feature extraction model.
[0043] Optionally, the farmland plot over-segmentation image determination module specifically includes:
[0044] The boundary probability image construction unit is used to construct the boundary probability image by taking the boundary probability of each pixel point in the remote sensing image of the study area as the gray value of the corresponding pixel point;
[0045] The over-segmented image determination unit of cultivated land is used for binarizing all gray values in the boundary probability image to obtain the over-segmented image of the cultivated land.
[0046] Optionally, the land parcel boundary identification quantity determination module specifically includes:
[0047] An iteration number assignment unit, used to set the iteration number n to 1;
[0048] A structural hierarchy construction unit is used to construct the nth structural hierarchy with the study area as the root node and the over-segmented regions in the over-segmented region set as leaf nodes;
[0049] A weight value determination unit, used to determine a weight value of each common boundary in the common boundary set; the weight value is an average value of boundary probabilities at all pixel points in the common boundary;
[0050] The over-segmented region merging unit is used to merge two sub-regions adjacent to the common boundary with the smallest weight value to obtain an updated over-segmented region set and an updated common boundary set, increase the value of the iteration number n by 1, and call the structure hierarchy construction unit until the iteration number n reaches the hierarchy threshold, and determine that the plot boundary corresponding to the nth structural hierarchy is the plot boundary identification amount; the over-segmented region includes multiple sub-regions; the number of sub-regions adjacent to the common boundary is greater than one; the number of over-segmented regions adjacent to the non-common boundary is 1 and the number of sub-regions adjacent to the non-common boundary is 1.
[0051] Optionally, the farmland plot extraction module specifically includes:
[0052] A land parcel boundary identification quantity conversion unit, used for converting the land parcel boundary identification quantity into land parcel boundary vector data;
[0053] A denoising unit, used for denoising the land parcel boundary vector data to obtain denoised land parcel boundary vector data;
[0054] A linearization processing unit is used to perform linearization processing on the denoised land parcel boundary vector data using the Douglas-Peucker algorithm to obtain linearized land parcel boundary vector data;
[0055] The cultivated land parcel extraction unit is used to divide the cultivated land parcels in the remote sensing image of the study area according to the linearized parcel boundary vector data.
[0056] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0057] The purpose of the present invention is to provide a method and system for automatically extracting cultivated land plots from remote sensing images. The method improves the semantic segmentation network to improve the sensitivity of the semantic segmentation network to linear information, and uses the improved semantic segmentation network to determine the boundary probability of each pixel point in the remote sensing image of the study area, thereby dividing the cultivated land plots in the remote sensing image of the study area, improving the continuity of plot boundary recognition, and thus improving the extraction accuracy of automatic extraction of cultivated land plots in remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0059] Figure 1 This is a flow chart of a method for automatically extracting cultivated land plots from remote sensing images in an embodiment of the present invention;
[0060] Figure 2 This is a schematic diagram of the boundaries corresponding to the first 90% hierarchical structure in an embodiment of the present invention;
[0061] Figure 3 This is a schematic diagram of the boundaries corresponding to the first 50% hierarchical structure in an embodiment of the present invention;
[0062] Figure 4 This is a schematic diagram of the boundaries corresponding to the first 10% hierarchical structure in an embodiment of the present invention;
[0063] Figure 5 Schematic diagram of the results of extracting cultivated land plots from remote sensing images in an embodiment of the present invention. DETAILED DESCRIPTION
[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0065] The purpose of the present invention is to provide a method and system for automatically extracting cultivated land plots from remote sensing images, which can improve the extraction accuracy of automatically extracting cultivated land plots from remote sensing images.
[0066] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0067] like Figure 1The present invention provides a method for automatically extracting cultivated land plots from remote sensing images, comprising:
[0068] Step 101: Acquire remote sensing images of the study area;
[0069] Step 102: Input each pixel value of the remote sensing image of the study area into the boundary feature extraction model to obtain the boundary probability of each pixel in the remote sensing image of the study area; the boundary feature extraction model is obtained by training the improved semantic segmentation network using the labeled historical remote sensing image of the study area; the improved semantic segmentation network is obtained by adding a shallow module to the semantic segmentation network; the shallow module is used to identify linear information; the shallow module is a single convolutional layer, which is used to strengthen the shallow boundary information, and the shallow module is superimposed with the output of the semantic segmentation network before the improvement in the fully connected layer to identify linear information. The semantic segmentation network is PSPNet (Pyramid Scene Parsing Network).
[0070] Step 103: determining an over-segmented image of the cultivated land plot according to the boundary probability of each pixel point in the remote sensing image of the study area; the number of boundaries of the cultivated land plot in the over-segmented image of the cultivated land plot is greater than the number of actual boundaries of the cultivated land plot;
[0071] Step 104: using a watershed segmentation algorithm to process the over-segmented image of the cultivated land to obtain an over-segmented region set and a common boundary set;
[0072] Step 105: determining the land parcel boundary recognition amount according to the over-segmented region set and the common boundary set;
[0073] Step 106: Divide the cultivated land plots in the remote sensing image of the study area according to the identified plot boundary values.
[0074] The method for automatically extracting cultivated land plots from remote sensing images provided by the present invention further includes, before step 101:
[0075] Obtain historical remote sensing images of the study area;
[0076] The boundary probability of the pixel points at the boundary in the historical remote sensing image of the study area is marked as 1, and the boundary probability of the pixel points at the non-boundary in the historical remote sensing image of the study area is marked as 0, so as to obtain the marked historical remote sensing image of the study area;
[0077] The pixel values of the historical remote sensing images marking the study area are used as input, and the boundary probability of the pixels is used as output. The improved semantic segmentation network is trained to obtain the boundary feature extraction model.
[0078] Specifically, step 103 specifically includes:
[0079] The boundary probability of each pixel in the remote sensing image of the study area is taken as the gray value of the corresponding pixel to construct a boundary probability image;
[0080] All gray values in the boundary probability image are binarized to obtain the over-segmented image of the cultivated land.
[0081] Preferably, step 105 specifically includes:
[0082] Let the number of iterations n be 1;
[0083] The nth level structure hierarchy is constructed with the study area as the root node and the over-segmented regions in the over-segmented region set as leaf nodes;
[0084] Determine the weight value of each common boundary in the common boundary set; the weight value is the average value of the boundary probabilities at all pixel points in the common boundary;
[0085] The two sub-regions adjacent to the common boundary with the smallest weight value are merged to obtain the updated over-segmented region set and the updated common boundary set, the value of the iteration number n is increased by 1, and the step "taking the study area as the root node and the over-segmented regions in the over-segmented region set as the leaf nodes to construct the nth layer of the structural hierarchy" is returned until the iteration number n reaches the layer threshold, and the land boundary corresponding to the nth layer of the structural hierarchy is determined as the land boundary identification amount. When the iteration number n reaches the layer threshold, the recognition result is most consistent with the actual land parcel scale; the over-segmented region includes multiple sub-regions; the number of sub-regions adjacent to the common boundary is greater than one; the number of over-segmented regions adjacent to the non-common boundary is 1 and the number of sub-regions adjacent to the non-common boundary is 1.
[0086] Furthermore, step 106 specifically includes:
[0087] Convert the land parcel boundary identification quantity into land parcel boundary vector data;
[0088] De-noising the plot boundary vector data to obtain de-noised plot boundary vector data;
[0089] The Douglas-Peucker algorithm is used to linearize the denoised plot boundary vector data to obtain the linearized plot boundary vector data;
[0090] The cultivated land plots in the remote sensing images of the study area are divided according to the linearized plot boundary vector data.
[0091] Specifically, the method for automatically extracting cultivated land plots from remote sensing images provided by the present invention is as follows:
[0092] Step 1: Semantic information detection:
[0093] The probability that each pixel on the image belongs to the boundary of the plot is detected by a deep convolutional neural network, which is a pixel-scale classification problem. Generally, a semantic segmentation model (i.e., PSPNet pyramid scene parsing network) is used, and the semantic segmentation model can output the category probability of each pixel. PSPNet combines a deep residual network with a pyramid pooling module, and has high recognition accuracy, but the process of multi-layer convolution of the network will lose position information, making the linear extraction ability weaker, which is inconsistent with the purpose of detecting the boundary information of the plot. Therefore, the present invention adopts the MPSPNet model, which is an improved model of PSPNet. On the basis of the PSPNet deep residual network and the pyramid pooling module, a shallow module for recognizing linear information is added, which strengthens the model's ability to linear extraction.
[0094] Training: The sample set of the study area is divided into three data sets according to the ratio of 12:3:85: training sample data set, validation sample data set and test sample data set. The sample annotated sample categories are "plot boundary class" and "non-plot boundary class". The adam algorithm is used as the gradient descent optimization method during training, and the validation sample set is used to select the optimal model.
[0095] Semantic information detection: Input the test sample set into the optimal model to detect the probability that each pixel in the image belongs to the "land boundary class". The land boundary probabilities of all pixels constitute the entire boundary probability image.
[0096] Step 2: Extracting the land parcel boundary:
[0097] Acquisition of over-segmented regions: The boundary probability image is a grayscale image. Watershed segmentation is performed on the boundary probability image. Watershed segmentation is a regional segmentation method that can convert grayscale images into binary segmented regions. Specifically, the image is regarded as a terrain surface, and the grayscale value of each pixel represents the height. Assuming that each isolated valley (local minimum) is filled with water, when the water rises to a certain height, it will form a watershed at the moment when it is about to overflow the valley. This is the principle of this method to segment the region. All local minima on the boundary probability image are used as seed points for watershed segmentation to obtain a completely over-segmented region.
[0098] Hierarchical structure construction: R represents the set of all over-segmented regions, K represents the set of common edges of all adjacent regions, and W(K i ) represents the weight value of all common edges, and the weight value is calculated by the average value of the boundary strength of all pixels on the line segment. Construct the graph G(R, K, W(K i )) Based on the graph theory greedy algorithm, the adjacent regions with the smallest common edge weight value are merged each time, and a hierarchical structure of regions is constructed in the process of region merging. The leaf nodes are all the over-segmented regions, and the root node is the whole graph region.
[0099] Optimal hierarchical threshold selection: In the hierarchical structure constructed based on the region, different hierarchical levels represent different segmentation scales. The present invention selects an appropriate hierarchical threshold to make the segmentation result most consistent with the scale of the plot on the image. Since the hierarchical threshold is only affected by the structure of the cultivated land, only a single optimal threshold needs to be selected on a large-scale image with similar planting structure to obtain the result of the plot boundary. On a general cultivated land image, selecting the first 20% of the hierarchical structure can roughly represent the boundary of the cultivated land. Specifically, the cultivated land boundary obtained by selecting the first 90% of the hierarchy is as follows: Figure 2 As shown in the figure, the cultivated land boundaries obtained by selecting the first 50% of the level are as follows Figure 3 As shown in the figure, the cultivated land boundaries obtained by selecting the first 10% of the level are as follows Figure 4 shown.
[0100] Step 3: Plot extraction:
[0101] Conversion of plot boundaries into vectors: Considering that the boundary result obtained in step 2 is raster data, and the goal of the present invention is to obtain the vector surface of the plot, this step requires the use of the ArcGIS raster-to-vector tool to convert the plot boundary result of step 2 into vector data.
[0102] Noise area removal: Since some pixels that are not the boundaries of the plots will be detected when detecting the probability information of the plot boundaries, they need to be filtered out: set the area threshold to remove vector surfaces with an area less than 100 square meters.
[0103] Straightening the boundary lines: Considering that the normal farmland boundaries are often straight, the boundary detection process of the present invention will blur the straight line features, resulting in a more tortuous boundary. Therefore, the Douglas-Peucker algorithm is finally used to straighten the boundary lines. Then, the cultivated land plots in the remote sensing image are extracted according to the boundaries after the line straightening process. The extraction results are as follows: Figure 5 .
[0104] In addition, the present invention also provides a remote sensing image cultivated land automatic extraction system, comprising:
[0105] The remote sensing image acquisition module of the study area is used to obtain the remote sensing image of the study area;
[0106] The boundary probability determination module is used to input each pixel value of the remote sensing image of the study area into the boundary feature extraction model to obtain the boundary probability of each pixel point in the remote sensing image of the study area; the boundary feature extraction model is obtained by training the improved semantic segmentation network using the labeled historical remote sensing images of the study area; the improved semantic segmentation network is obtained by adding a shallow module to the semantic segmentation network; the shallow module is used to identify linear information;
[0107] The over-segmented image determination module of cultivated land is used to determine the over-segmented image of cultivated land according to the boundary probability of each pixel point in the remote sensing image of the study area;
[0108] The module for determining the land parcel boundary recognition quantity is used to process the over-segmented image of the cultivated land parcel using the watershed segmentation algorithm to obtain the land parcel boundary recognition quantity;
[0109] A set extraction module is used to process the over-segmented images of cultivated land plots using a watershed segmentation algorithm to obtain an over-segmented region set and a common boundary set;
[0110] A land parcel boundary recognition amount determination module, used to determine the land parcel boundary recognition amount according to the over-segmented region set and the common boundary set;
[0111] The cultivated land parcel extraction module is used to divide the cultivated land parcels in the remote sensing image of the study area according to the parcel boundary identification quantity.
[0112] In addition, the remote sensing image cultivated land plot automatic extraction system provided by the present invention also includes:
[0113] Historical remote sensing image acquisition module, used to obtain historical remote sensing images of the study area;
[0114] A marking module is used to mark the boundary probability of pixel points at the boundary of the historical remote sensing image of the study area as 1, and mark the boundary probability of pixel points at the non-boundary of the historical remote sensing image of the study area as 0, so as to obtain a marked historical remote sensing image of the study area;
[0115] The boundary feature extraction model determination module is used to take the pixel value of the historical remote sensing image marking the study area as input and the boundary probability of the pixel as output, train the improved semantic segmentation network, and obtain the boundary feature extraction model.
[0116] Specifically, the over-segmented image determination module for cultivated land parcels includes:
[0117] The boundary probability image construction unit is used to construct the boundary probability image by taking the boundary probability of each pixel point in the remote sensing image of the study area as the gray value of the corresponding pixel point;
[0118] The over-segmented image determination unit of cultivated land is used for binarizing all gray values in the boundary probability image to obtain the over-segmented image of the cultivated land.
[0119] Specifically, the land parcel boundary identification quantity determination module specifically includes:
[0120] An iteration number assignment unit, used to set the iteration number n to 1;
[0121] A structural hierarchy construction unit is used to construct the nth structural hierarchy with the study area as the root node and the over-segmented regions in the over-segmented region set as leaf nodes;
[0122] A weight value determination unit, used to determine the weight value of each common boundary in the common boundary set; the weight value is the average value of boundary probabilities at all pixel points in the common boundary;
[0123] The over-segmented region merging unit is used to merge two sub-regions adjacent to the common boundary with the smallest weight value to obtain an updated over-segmented region set and an updated common boundary set, increase the value of the iteration number n by 1, and call the structure hierarchy construction unit until the iteration number n reaches the hierarchy threshold, and determine the plot boundary corresponding to the nth structural hierarchy as the plot boundary identification amount; the over-segmented region includes multiple sub-regions; the number of sub-regions adjacent to the common boundary is greater than one; the number of over-segmented regions adjacent to the non-common boundary is 1 and the number of sub-regions adjacent to the non-common boundary is 1.
[0124] Specifically, the farmland plot extraction module includes:
[0125] A land parcel boundary identification quantity conversion unit, used for converting the land parcel boundary identification quantity into land parcel boundary vector data;
[0126] A denoising unit, used for denoising the land parcel boundary vector data to obtain denoised land parcel boundary vector data;
[0127] A linearization processing unit is used to perform linearization processing on the denoised land parcel boundary vector data using the Douglas-Peucker algorithm to obtain linearized land parcel boundary vector data;
[0128] The cultivated land parcel extraction unit is used to divide the cultivated land parcels in the remote sensing image of the study area according to the linearized parcel boundary vector data.
[0129] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0130] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A method for automatically extracting cultivated land plots from remote sensing images, characterized in that: The method comprises: Obtain remote sensing images of the study area; Each pixel value of the remote sensing image of the study area is input into the boundary feature extraction model to obtain the boundary probability of each pixel point in the remote sensing image of the study area; the boundary feature extraction model is obtained by training the improved semantic segmentation network by using the labeled historical remote sensing image of the study area; the improved semantic segmentation network is obtained by adding a shallow module to the semantic segmentation network; the shallow module is used to identify linear information; According to the boundary probability of each pixel in the remote sensing image of the study area, the over-segmented image of the cultivated land plot is determined; The over-segmented images of cultivated land are processed using the watershed segmentation algorithm to obtain the over-segmented area set and the common boundary set. Determining a land parcel boundary recognition amount according to the over-segmented region set and the common boundary set; Divide the cultivated land plots in the remote sensing image of the study area according to the plot boundary identification amount; The determining of the land parcel boundary recognition amount according to the over-segmented region set and the common boundary set specifically includes: Let the number of iterations n be 1; The nth level structure hierarchy is constructed with the study area as the root node and the over-segmented regions in the over-segmented region set as leaf nodes; Determine a weight value of each common boundary in the common boundary set; the weight value is an average value of boundary probabilities at all pixel points in the common boundary; Merge multiple sub-regions adjacent to the common boundary with the smallest weight value to obtain an updated over-segmented region set and an updated common boundary set, increase the value of the iteration number n by 1, and return to the step of "taking the study area as the root node and the over-segmented regions in the over-segmented region set as leaf nodes to construct the nth layer of the structural hierarchy" until the iteration number n reaches the hierarchy threshold, and determine that the plot boundary corresponding to the nth layer of the structural hierarchy is the plot boundary identification amount; the over-segmented region includes multiple sub-regions; the number of sub-regions adjacent to the common boundary is greater than one; the number of over-segmented regions adjacent to the non-common boundary is 1 and the number of sub-regions adjacent to the non-common boundary is 1.
2. The method for automatically extracting cultivated land plots from remote sensing images according to claim 1, characterized in that: Before obtaining the remote sensing image of the study area, the method also includes: Obtain historical remote sensing images of the study area; The boundary probability of the pixel points at the boundary of the historical remote sensing image of the study area is marked as 1. Will The boundary probability of non-boundary pixels in the historical remote sensing images of the study area is marked as 0, and the historical remote sensing images of the study area are marked; The pixel values of the historical remote sensing images marking the study area are used as input, and the boundary probability of the pixels is used as output, and the improved semantic segmentation network is trained to obtain the boundary feature extraction model.
3. The method for automatically extracting cultivated land plots from remote sensing images according to claim 1, characterized in that: Determining the over-segmented image of the cultivated land plot according to the boundary probability of each pixel point in the remote sensing image of the study area specifically includes: The boundary probability of each pixel in the remote sensing image of the study area is taken as the gray value of the corresponding pixel to construct a boundary probability image; All gray values in the boundary probability image are binarized to obtain the over-segmented image of the cultivated land.
4. The method for automatically extracting cultivated land plots from remote sensing images according to claim 1, characterized in that: The dividing of the cultivated land plots in the remote sensing image of the study area according to the land plot boundary recognition amount specifically includes: Converting the land parcel boundary identification amount into land parcel boundary vector data; Performing denoising on the land parcel boundary vector data to obtain denoised land parcel boundary vector data; The Douglas-Peucker algorithm is used to linearize the denoised plot boundary vector data to obtain the linearized plot boundary vector data; The cultivated land plots in the remote sensing image of the study area are divided according to the linearized plot boundary vector data.
5. A remote sensing image cultivated land automatic extraction system, characterized in that: The system comprises: The remote sensing image acquisition module of the study area is used to obtain the remote sensing image of the study area; A boundary probability determination module is used to input each pixel value of the remote sensing image of the study area into a boundary feature extraction model to obtain the boundary probability of each pixel point in the remote sensing image of the study area; the boundary feature extraction model is obtained by training an improved semantic segmentation network using the labeled historical remote sensing image of the study area; the improved semantic segmentation network is obtained by adding a shallow module to the semantic segmentation network; the shallow module is used to identify linear information; The over-segmented image determination module of cultivated land plots is used to determine the over-segmented image of cultivated land plots according to the boundary probability of each pixel point in the remote sensing image of the study area; The module for determining the land parcel boundary recognition quantity is used to process the over-segmented image of the cultivated land parcel using the watershed segmentation algorithm to obtain the land parcel boundary recognition quantity; A set extraction module is used to process the over-segmented images of cultivated land plots using a watershed segmentation algorithm to obtain an over-segmented region set and a common boundary set; A land parcel boundary recognition amount determination module, used to determine a land parcel boundary recognition amount according to the over-segmented region set and the common boundary set; A cultivated land parcel extraction module, used for dividing the cultivated land parcels in the remote sensing image of the study area according to the parcel boundary recognition amount; The land parcel boundary identification quantity determination module specifically includes: An iteration number assignment unit, used to set the iteration number n to 1; A structural hierarchy construction unit is used to construct the nth structural hierarchy with the study area as the root node and the over-segmented regions in the over-segmented region set as leaf nodes; A weight value determination unit, used to determine a weight value of each common boundary in the common boundary set; the weight value is an average value of boundary probabilities at all pixel points in the common boundary; The over-segmented region merging unit is used to merge two sub-regions adjacent to the common boundary with the smallest weight value to obtain an updated over-segmented region set and an updated common boundary set, increase the value of the iteration number n by 1, and call the structure hierarchy construction unit until the iteration number n reaches the hierarchy threshold, and determine that the plot boundary corresponding to the nth structural hierarchy is the plot boundary identification amount; the over-segmented region includes multiple sub-regions; the number of sub-regions adjacent to the common boundary is greater than one; the number of over-segmented regions adjacent to the non-common boundary is 1 and the number of sub-regions adjacent to the non-common boundary is 1.
6. The automatic extraction system of cultivated land plots from remote sensing images according to claim 5 is characterized in that: The system further comprises: Historical remote sensing image acquisition module, used to obtain historical remote sensing images of the study area; Marking module, used to The boundary probability of the pixel points at the boundary in the historical remote sensing image of the study area is marked as 1, and the boundary probability of the pixel points at the non-boundary in the historical remote sensing image of the study area is marked as 0, so as to obtain the marked historical remote sensing image of the study area; The boundary feature extraction model determination module is used to take the pixel value of the historical remote sensing image marking the study area as input and the boundary probability of the pixel as output, train the improved semantic segmentation network, and obtain the boundary feature extraction model.
7. The automatic extraction system of cultivated land plots from remote sensing images according to claim 6 is characterized in that: The over-segmented image determination module of the cultivated land plot specifically comprises: The boundary probability image construction unit is used to construct the boundary probability image by taking the boundary probability of each pixel point in the remote sensing image of the study area as the gray value of the corresponding pixel point; The over-segmented image determination unit of cultivated land is used for binarizing all gray values in the boundary probability image to obtain the over-segmented image of the cultivated land.
8. The automatic extraction system of cultivated land plots from remote sensing images according to claim 5 is characterized in that: The farmland plot extraction module specifically includes: A land parcel boundary identification quantity conversion unit, used for converting the land parcel boundary identification quantity into land parcel boundary vector data; A denoising unit, used for denoising the land parcel boundary vector data to obtain denoised land parcel boundary vector data; A linearization processing unit is used to perform linearization processing on the denoised land parcel boundary vector data using the Douglas-Peucker algorithm to obtain linearized land parcel boundary vector data; The cultivated land parcel extraction unit is used to divide the cultivated land parcels in the remote sensing image of the study area according to the linearized parcel boundary vector data.
Citation Information
Cited By
A complex terrain farmland block multi-dimensional feature fine identification and classification system
CN122761206A
A method for remotely sensing and segmenting the boundary of a hilly farmland plot
CN122866204A
