Land parcel fine identification method and system based on boundary buffer deep learning

Abstract

The present application relates to the field of data recognition technology and provides a method and system for fine land parcel recognition based on boundary buffer deep learning. The method comprises: first, using a deep learning model, identifying the land parcel area of the image to be recognized to obtain a first land parcel recognition result; wherein the deep learning model is trained based on pre-acquired image data and label data, and the label data is obtained by performing a first buffering process on the original label data; then, the first binary image is sequentially subjected to a second buffering process and a connected domain process to obtain a first seed image; finally, based on the first seed image, the input image is segmented using a watershed algorithm to obtain a fine land parcel recognition result. In this way, the label data obtained by performing the first buffering process on the original label data is input into the deep learning model as training data, and then robust and accurate subsequent processing steps are adopted to improve the accuracy of land parcel recognition.

Description

Technical Field

[0001] The present application relates to the field of data recognition technology, and in particular to a method and system for fine land parcel recognition based on boundary buffer deep learning. Background Art

[0002] Comprehensive, accurate, and real-time farmland boundary information is essential data for supporting crop identification, detection, and statistics. Farmland boundaries vary across remote sensing imagery at different scales. Traditional edge detection algorithms often lack precision and accuracy in extracting farmland objects based on these boundaries. Consequently, the accuracy of the resulting objectified farmland parcels falls short of practical application requirements.

[0003] Therefore, it is necessary to provide an improved technical solution to the above-mentioned deficiencies in the prior art. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for fine land parcel identification based on boundary buffer deep learning to solve or alleviate the problems existing in the above-mentioned prior art.

[0005] In order to achieve the above objectives, this application provides the following technical solutions:

[0006] This application provides a method for fine-grained land parcel identification based on deep learning of boundary buffers, including:

[0007] Identify the land parcel area in the image to be identified using a deep learning model to obtain a first land parcel identification result; wherein the deep learning model is trained based on pre-acquired image data and label data, and the label data is obtained by performing a first buffering process on the original label data; wherein the first buffering process is to expand N pixels on both sides of the boundary of the land parcel in the original label data as a center line, where N is a positive integer greater than 1;

[0008] The first binary image is sequentially subjected to a second buffering process and a connected component process to obtain a first seed image; wherein the first binary image is obtained by binarizing the land parcel area in the first land parcel identification result; the second buffering process is performed by shrinking each pixel in the first binary image by M pixels toward the center of the land parcel area in the first binary image; M is determined based on the closest distance between each pixel in the land parcel area in the first binary image and a non-land parcel pixel in the first binary image; M is a positive integer greater than 1;

[0009] According to the first seed image, the input image is segmented by a watershed algorithm to obtain a fine land parcel recognition result; wherein, the input image is obtained by enhancing the boundary of the image to be recognized according to the predicted boundary area in the first land parcel recognition result.

[0010] In some embodiments, the method further comprises:

[0011] performing binarization processing and uniformly assigning grayscale values to the non-land parcel area in the first land parcel recognition result to obtain a non-land parcel area seed image; wherein the grayscale value of each pixel in the non-land parcel area seed image is different from the grayscale value of each pixel in the first seed image;

[0012] appending the non-plot area seed image to the first seed image;

[0013] According to the first seed image, the input image is segmented by the watershed algorithm to obtain the fine land recognition result.

[0014] In some embodiments, based on the predicted boundary area in the first land parcel recognition result, the boundary of the image to be recognized is enhanced to obtain the input image, specifically:

[0015] performing binarization processing and boundary skeleton extraction on the predicted boundary area in the first land parcel recognition result in sequence to obtain a boundary skeleton image;

[0016] Performing band-by-band grayscale enhancement on the image to be identified to obtain an enhanced image to be identified;

[0017] The first boundary area in the enhanced image to be identified is uniformly assigned band by band to obtain an input image, wherein the first boundary area is an area corresponding to the position of the boundary area in the boundary skeleton image on the enhanced image to be identified.

[0018] In some embodiments, the method further comprises:

[0019] Vector conversion is performed on the land parcel fine identification result to obtain land parcel vector data.

[0020] In some embodiments, the vector conversion is performed on the land parcel fine recognition result to obtain the land parcel vector data, specifically:

[0021] Converting the land parcel boundary in the land parcel fine recognition result into polyline vector data;

[0022] Converting the polyline vector data into polygon vector data;

[0023] Determining the category of each polygon in the polygon vector data according to the multiple categories in the land parcel fine recognition result, and obtaining polygon vector data containing the land parcel category;

[0024] The polygonal vector data containing the land parcel category is optimized to obtain land parcel vector data.

[0025] In some embodiments, the deep learning model includes at least one convolutional attention module;

[0026] The convolutional attention module is used to perform attention calculation on the feature map output by the residual network of the deep learning model, and input the obtained calculation results into the pyramid segmentation network of the deep learning model.

[0027] In some embodiments, the convolutional attention module includes a channel attention submodule and a spatial attention submodule;

[0028] The channel attention submodule is configured to compress the feature map output by the residual network in the spatial dimension to obtain a channel attention feature map;

[0029] The spatial attention submodule is configured to globally weight the channel attention feature map to obtain an input feature map of the pyramid segmentation network.

[0030] In some embodiments, during the training process, the deep learning model calculates a training loss value based on the average of the focus loss function and the Dice loss function.

[0031] In some embodiments, the method further comprises:

[0032] Training a plurality of convolutional neural networks according to the pre-acquired image data and the label data to obtain a plurality of candidate convolutional neural networks;

[0033] The deep learning model is determined based on the parameter quantity of each candidate convolutional neural network and the accuracy of the candidate convolutional neural network.

[0034] The present application also provides a land parcel fine recognition system based on boundary buffer deep learning, including:

[0035] A land parcel recognition unit is configured to identify a land parcel area in the image to be recognized using a deep learning model to obtain a first land parcel recognition result; wherein the deep learning model is trained based on pre-acquired image data and label data, and the label data is obtained by performing a first buffering process on the original label data; wherein the first buffering process is to expand N pixels on both sides of the boundary of the land parcel in the original label data as a center line, where N is a positive integer greater than 1;

[0036] a buffer processing unit configured to sequentially perform a second buffer process and a connected domain process on the first binarized image to obtain a first seed image; wherein the first binarized image is obtained by binarizing the land parcel area in the first land parcel identification result; the second buffer process is to shrink each pixel in the first binarized image by M pixels toward the center of the land parcel area in the first binarized image; M is determined based on the closest distance between each pixel in the land parcel area in the first binarized image and a non-land parcel pixel in the first binarized image; and M is a positive integer greater than 1;

[0037] The fine segmentation unit is configured to: segment the input image according to the first seed image using the watershed algorithm to obtain a fine land parcel recognition result; wherein the input image is obtained by enhancing the boundaries of the image to be recognized based on the predicted boundary area in the first land parcel recognition result.

[0038] Beneficial effects:

[0039] In the present application, a deep learning model is first used to identify the plot area of the image to be identified to obtain a first plot identification result; wherein the deep learning model is trained based on pre-acquired image data and label data, and the label data is obtained by performing a first buffering process on the original label data; wherein the first buffering process is to extend N pixels on both sides of the boundary of the plot in the original label data as the center line, where N is a positive integer greater than 1; then, the first binarized image is sequentially subjected to a second buffering process and a connected domain process to obtain a first seed image; wherein the first binarized image is obtained by binarizing the plot area in the first plot identification result; the second buffering process is to shrink each pixel in the first binarized image by M pixels toward the center of the plot area in the first binarized image; M is determined based on the nearest distance between each pixel in the plot area in the first binarized image and the non-plot pixel in the first binarized image; M is a positive integer greater than 1; finally, based on the first seed image, the input image is segmented by a watershed algorithm to obtain a fine plot identification result. In this way, the image to be identified is preliminarily identified through the deep learning model to obtain the preliminary plot identification result of the plot boundary. Then, the exact area range of the plot and non-plot is determined as the seed image through the second buffer processing and connected domain processing. The image to be identified is enhanced based on the boundary skeleton image as the input image, and the watershed algorithm is performed to improve the accuracy of plot boundary extraction and plot identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings and descriptions that constitute part of this application are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. Among them:

[0041] Figure 1 A schematic flow chart of a method for fine land parcel identification based on boundary buffer deep learning according to some embodiments of the present application;

[0042] Figure 2 A technical logic diagram of a land parcel fine identification method based on boundary buffer deep learning according to some embodiments of the present application;

[0043] Figure 3 A schematic diagram of pre-acquired image data and label data provided according to some embodiments of the present application;

[0044] Figure 4 A schematic diagram of the structure of a deep learning model provided according to some embodiments of the present application;

[0045] Figure 5 A schematic diagram of a process for determining a plot area and a non-plot area according to some embodiments of the present application;

[0046] Figure 6 A schematic diagram of a boundary skeleton extraction process provided according to some embodiments of the present application;

[0047] Figure 7 A schematic diagram of land parcel vector data obtained by a land parcel fine recognition method based on boundary buffer deep learning provided in some embodiments of the present application;

[0048] Figure 8 A schematic structural diagram of a land parcel fine recognition system based on boundary buffer deep learning provided according to some embodiments of the present application. DETAILED DESCRIPTION

[0049] The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments. Each example is provided by way of explanation of the present application and does not limit the present application. In fact, it will be clear to those skilled in the art that modifications and variations can be made in the present application without departing from the scope or spirit of the present application. For example, a feature shown or described as part of one embodiment can be used in another embodiment to produce yet another embodiment. Therefore, it is expected that the present application includes such modifications and variations within the scope of the appended claims and their equivalents.

[0050] Exemplary Methods

[0051] In the related art, the identification and segmentation of cultivated land parcel boundaries are usually performed based on traditional edge detection algorithms, which use edge detection operators to depict the boundaries of the parcels, and then use image segmentation methods to segment the parcels. However, the principle of the edge detection algorithm is to use edge detection operators to detect discontinuities between pixels in image data to identify edge strength. Its recognition results are prone to breakage, resulting in discontinuous identification boundaries, making it impossible to accurately identify the parcel boundaries, resulting in the inability to automatically extract cultivated land parcel vector objects, and thus the inability to unitize and objectify cultivated land parcels. Due to the above-mentioned problems in the automated extraction of cultivated land parcels, in most cases, high-precision cultivated land parcel boundary identification is still performed by manual visual interpretation, which has high labor costs, low efficiency, and difficulty in achieving frequent updates and corrections.

[0052] To address the problems of insufficient accuracy and low efficiency in the prior art of identifying and extracting cultivated land plots and their boundaries, an embodiment of the present application provides a method for fine-grained land parcel identification based on boundary buffering deep learning. This method performs a first boundary buffering process on training samples, and on this basis, trains a deep learning model for boundary identification. Then, based on the deep learning model, preliminary identification of cultivated land parcel boundaries is performed to obtain preliminary land parcel identification results of the parcel boundaries. Subsequently, the exact area range of the parcel and non-parcel is obtained through a second buffering process and a connected domain process. This is used as a seed image, and the boundary of the image to be identified is enhanced based on the boundary skeleton image. The result is used as the input image, and the seed image and the input image are segmented using the watershed algorithm to obtain the automatic extraction result of cultivated land parcels. Experiments show that the automatic extraction result of cultivated land parcels reaches the same level of precision and accuracy as the cultivated land parcel results obtained by manual visual interpretation and manual vectorization, thereby realizing the automatic extraction of cultivated land parcels and their boundaries and improving the efficiency of the extraction of cultivated land parcels and their boundaries.

[0053] Figure 1 A schematic flow chart of a method for fine land parcel identification based on boundary buffer deep learning according to some embodiments of the present application; Figure 2 A technical logic diagram of a land parcel fine identification method based on boundary buffer deep learning according to some embodiments of the present application; the method includes:

[0054] Step S101: Identify the land area of the image to be identified through a deep learning model to obtain a first land identification result; wherein, the deep learning model is trained based on pre-acquired image data and label data, and the label data is obtained by performing a first buffering process on the original label data; wherein, the first buffering process is to expand N pixels on both sides of the boundary of the land in the original label data as the center line, where N is a positive integer greater than 1.

[0055] It should be noted that, in the embodiment of the present application, the plot is specifically a cultivated land plot, which is manifested as a continuous planting area of the same or similar crops. The boundary area of the plot can be a land entity such as a ridge, a road, a water system, etc. that can divide a large area into multiple continuous planting areas. In remote sensing image data, a plot or a plot boundary is a continuous pixel with roughly the same grayscale value; in vector plot data, a plot is a polygon enclosed by a plot boundary, and a plot boundary is a series of coordinate points. It is precisely because of the difference in data organization between plots and plot boundaries in image data and vector plot data that the plot and plot boundary information in remote sensing image data that can be easily identified by human visual inspection still has the problem of low accuracy and poor practicality when automatically identified and extracted based on computer programs, which leads to low statistical efficiency of cultivated land plots.

[0056] In the embodiment of the present application, the image to be identified is a high-resolution remote sensing image, and the image to be identified contains information of cultivated land plots to be identified, and the high resolution is a resolution of 2m or above.

[0057] In the embodiment of the present application, the land area of the image to be identified is identified by a deep learning model to obtain a first land identification result. Figure 2 The deep learning processing unit includes a deep learning model training process and a model recognition process after the training is completed. The deep learning model training process trains the model by acquiring a small range of image data and its corresponding label data (small sample), and uses a large range of image data and its corresponding label data (large sample) to tune the model performance, thereby obtaining a trained deep learning model. The model recognition process of the deep learning model can also be called a model production process, in which the trained deep recognition model is used to perform boundary recognition on the image data to be recognized, and obtain the first plot recognition result. In this way, the deep learning model's ability to abstract the image features layer by layer and in high dimensions is combined with a deep convolutional neural network to encode and extract the semantic information of the image to be recognized, and obtain a preliminary plot recognition result (first plot recognition result).

[0058] The deep learning model is trained based on pre-acquired image data and labeled data. The image data used in the model training process is high-resolution imagery with a resolution of 2 meters or higher, including red, green, and blue bands. The original labeled data used in the model training is labeled samples obtained from 0.8-meter resolution image data, and is spatially aligned with the image data.

[0059] In the embodiment of the present application, the original label data is subjected to a first buffering process to obtain label data for training the model, wherein the first buffering process is to extend N pixels on both sides of the boundary of the land parcel in the original label data as the center line, where N is a positive integer greater than 1, such as Figure 3 As shown, Figure 3 The i1 part in is the pre-acquired image data. Figure 3 The l1 portion of the data represents the label data. As can be seen from the figure, after the first buffering process, the resulting label data includes three types of labels: plot, non-plot, and boundary. Furthermore, the boundary region is widened. By extending the plot boundary to both sides, the boundary region features of the plots in the training samples are increased, ensuring that the deep learning model can learn the boundary features of the plots and improving its ability to recognize plot boundaries.

[0060] In specific applications, N can be determined based on the size of the geographical area covered by the training image data or based on prior knowledge. For example, the value of N can be 7 pixels.

[0061] To facilitate deep learning model training, the image and label data were first organized by county administrative region, with each county administrative region corresponding to both image and label data. The image and label data were then cropped with a 25% overlap, resulting in multiple 256-pixel x 256-pixel image blocks. Finally, the cropped image blocks were fed into the deep learning model for training.

[0062] In some embodiments, the boundary buffer deep learning-based land parcel identification method further includes: training multiple convolutional neural networks based on pre-acquired image data and label data to obtain multiple candidate convolutional neural networks; and determining a deep learning model based on the parameters and accuracy of each candidate convolutional neural network. In this way, multiple convolutional neural networks are screened using the same training data to improve the output accuracy of the deep learning model in land parcel boundary identification.

[0063] Exemplarily, multiple convolutional neural networks are trained synchronously using the same training data to obtain candidate convolutional neural networks, wherein the performance comparison of the candidate convolutional neural networks is shown in Table 1, which is as follows:

[0064] Table 1 Performance comparison of alternative convolutional neural networks

[0065]

[0066] As can be seen from Table 1, taking into account the parameter quantities and accuracy of each alternative convolutional neural network, the embodiment of the present application selects ResNet34-PSP as the deep learning model for land parcel boundary recognition.

[0067] In other optional embodiments, the deep learning model includes at least one convolutional block attention module (CBAM); the convolutional attention module is used to perform attention calculation on the feature map output by the residual network (ResNet) of the deep learning model, and input the calculation results into the pyramid segmentation network (PSP) of the deep learning model.

[0068] In the embodiment of the present application, the network structure of the deep learning model is as follows Figure 4 shown. Figure 4 In the figure, the deep learning model (ResNet34-PSP) is represented by the ResNet backbone network on the left and the PSP network on the right. ResNet consists of four layers, with the bottom layer, Layer 4, connected to the CBAM module. The CBAM module calculates the ResNet output feature map, and the resulting calculations are fed into the PSP network, enhancing its ability to capture target features at land boundary boundaries.

[0069] In some further embodiments, the convolutional attention module includes a channel attention sub-module and a spatial attention sub-module; the channel attention sub-module is configured to compress the feature map output by the residual network in the spatial dimension to obtain a channel attention feature map; the spatial attention sub-module is configured to globally weight the channel attention feature map to obtain an input feature map of the pyramid segmentation network.

[0070] Among them, the channel attention submodule performs global maximum pooling and global average pooling on the feature map output by the residual network in the channel dimension to obtain two one-dimensional feature vectors, and assigns feature weights to all one-dimensional feature vectors through the fully connected layer to obtain a channel attention feature map with enhanced ability to capture the target features of the land boundary in the channel domain.

[0071] The spatial attention submodule compresses the channel attention feature map output by the channel attention submodule in the spatial dimension to generate a two-dimensional feature map, and uses the convolution operation to perform global weighted processing on the two-dimensional feature map to complete the secondary capture capability enhancement of the plot boundary target features in the spatial domain, thereby obtaining the output result of the double enhancement of the plot boundary target feature capture capability.

[0072] By adding a convolutional attention module between the residual network and the pyramid segmentation network, the feature maps of different scales output by the residual network are combined before being input into the pyramid segmentation network. This allows for the simultaneous recognition of both planar and existing features in the image to be identified. This strengthens the deep learning model's ability to capture the target features of land parcel boundaries, effectively improving recognition accuracy and conserving deep learning model parameters. Experimental data shows that the addition of the convolutional attention module improves the overall accuracy of the deep learning model's land parcel boundary recognition results by approximately 2% compared to a deep learning model without the convolutional attention module.

[0073] In some optional embodiments, during the training process of the deep learning model, a training loss value is calculated based on the average of the focus loss function and the Dice loss function.

[0074] Since there are fewer plot boundaries, the number of training samples is relatively small compared to the number of plot and non-plot samples. At this time, if the traditional loss function is used for model training, it is easy for the model to fall into a local minimum and the training process is unstable. To avoid the above problems to a certain extent, in the embodiment of the present application, the training loss value is calculated based on the average of the focal loss function and the Dice loss function, which can alleviate the problem of imbalance in the number of training samples to a certain extent.

[0075] Among them, the focal loss function is as follows:

[0076] L f =-(1-p t ) γ log(p t )

[0077]

[0078] Where, L f represents the focal loss function; p represents the predicted value, that is, the result of deep learning model recognition; p t Indicates the size of the predicted probability; g represents the label value corresponding to the pre-acquired image data; γ is a hyperparameter set during the training process, γ>0.

[0079] The Dice loss function is as follows:

[0080]

[0081] Where, L D represents the Dice loss function, N represents the total number of pixels in the deep learning model recognition result; P i Represents the predicted value of the i-th pixel in the recognition result; T i Indicates the label value corresponding to the i-th pixel.

[0082] The joint loss of training is the average of the focal loss function and the Dice loss function:

[0083] Loss=(L f +L D ) / 2

[0084] Where Loss represents the loss function of the deep learning model. When the loss function is minimized, the model converges.

[0085] Step S102: performing a second buffering process and a connected domain process on the first binary image in sequence to obtain a first seed image; wherein the first binary image is obtained by performing a binarization process on the land area in the first land parcel recognition result; the second buffering process is to shrink each pixel in the first binary image by M pixels toward the center of the land area in the first binary image; M is determined based on the closest distance between each pixel in the land area in the first binary image and the non-land parcel pixel in the first binary image; M is a positive integer greater than 1.

[0086] Figure 5 A schematic diagram of a process for determining a plot area and a non-plot area according to some embodiments of the present application is provided. Figure 5 As shown, the first land parcel recognition result obtained by the deep learning model is a raster image, and each pixel in the image is one of the three types: land parcel, non-land parcel, and boundary.

[0087] In order to delineate the exact area of the plot in the first plot recognition result, the plot area of the first plot recognition result is first binarized to obtain a first binary image, thereby isolating the pixels predicted to be plots in the first plot recognition result into the first binary image, avoiding interference of non-plot and boundary types in the subsequent processing.

[0088] Then, a second buffering process is performed on the first binary image, wherein the second buffering process is to shrink each pixel in the first binary image by M pixels toward the center of the land area in the first binary image; M is determined according to the closest distance between each pixel in the land area in the first binary image and the non-land area pixel in the first binary image; M is a positive integer greater than 1.

[0089] In practical applications, a pixel value of 1 in the first binary image indicates that the pixel is a plot, and a value of 0 indicates that the pixel is not a plot. The second buffer processing process is as follows: establish a mapping image, the size of which is the same as that of the first binary image; traverse the pixels with a value of 1 in the first binary image, find the pixel with a value of 0 closest to the current pixel, calculate the number of pixels between the pixel with a value of 0 and the current pixel, and the number of pixels is M; record the value of M at the corresponding pixel position in the mapping image; according to the value of each pixel in the above mapping image (that is, the value of M), shrink each pixel with a value of 1 toward the center of the plot area in the first binary image. In this way, the pixels at the edge of the plot in the first binary image that are uncertain whether they are plots are eliminated to obtain a pure plot area.

[0090] Subsequently, connected domain processing is performed on the result of the second buffering process, where connected land pixels are marked with the same pixel value and disconnected land pixels are isolated with different pixel values, thereby obtaining a first seed image.

[0091] In the embodiment of the present application, the connected domain processing adopts an optimized block connected domain marking algorithm, where the "block" refers to a sliding window of 2 pixels × 2 pixels in size. The image of the result of the second buffer processing is scanned line by line based on the optimized block connected domain marking algorithm to determine whether the pixels in the sliding window meet the connectivity conditions, and the pixels that meet the connectivity conditions are marked with the same value based on the sliding window. Compared with other connected domain algorithms, the optimized block connected domain marking algorithm has higher processing efficiency and processing speed for high-resolution images, is easy to implement, and can significantly reduce the number of memory accesses during the connected domain processing of high-resolution image data, thereby improving computing efficiency.

[0092] In an embodiment of the present application, a second buffering process is performed on the first plot identification result through an adaptive buffer distance, and pixels at the edge of the plot in the first binary image that are uncertain whether they are plots are removed to obtain a pure plot area. Then, the pure plot area is marked with a connected domain based on an optimized block connected domain marking algorithm, and the result of the connected domain marking is used as the first seed image and input into the watershed algorithm for fine plot segmentation, thereby greatly improving the accuracy of segmentation.

[0093] In some optional embodiments, the method for fine land parcel identification based on boundary buffer deep learning also includes: binarizing and uniformly assigning grayscale values to the non-land parcel areas in the first land parcel identification result in sequence to obtain a seed image of the non-land parcel area; wherein the grayscale value of each pixel in the non-land parcel area seed image is different from the grayscale value of each pixel in the first seed image; appending the non-land parcel area seed image to the first seed image; and segmenting the input image according to the first seed image using a watershed algorithm to obtain a fine land parcel identification result.

[0094] See also Figure 5 As mentioned above, the first land parcel recognition result obtained by the deep learning model is a raster image, and each pixel in the image is one of the three types: land parcel, non-land parcel, and boundary. In the embodiment of the present application, the non-land parcel area in the first land parcel recognition result is binarized and uniformly assigned grayscale values in turn, so that the non-land parcel area in the recognition result of the deep learning model is extracted, assigned a grayscale value different from that in the first seed image, and appended to the first seed image to obtain the seed image required by the watershed algorithm. In this way, the identification of land parcels and non-land parcels is retained in the first seed image through different grayscale values, so that the types of land parcels and non-land parcels can be quickly obtained in the watershed algorithm segmentation result obtained based on the first seed image.

[0095] Step S103 : segmenting the input image using a watershed algorithm based on the first seed image to obtain a fine land parcel recognition result; wherein the input image is obtained by performing boundary enhancement on the image to be recognized based on the predicted boundary area in the first land parcel recognition result.

[0096] In some optional embodiments, the boundary of the image to be identified is enhanced according to the predicted boundary area in the first land parcel identification result to obtain an input image, specifically: the predicted boundary area in the first land parcel identification result is binarized and the boundary skeleton is extracted in sequence to obtain a boundary skeleton image; the image to be identified is grayscale enhanced band by band to obtain an enhanced image to be identified; the first boundary area in the enhanced image to be identified is uniformly assigned band by band to obtain an input image, wherein the first boundary area is the area corresponding to the position of the boundary area in the boundary skeleton image on the enhanced image to be identified.

[0097] It should be noted that the boundary skeleton refers to the central axis of the land parcel boundary. In the embodiment of the present application, the predicted boundary area in the first land parcel recognition result obtained by the deep learning model may have two situations: single-line or double-line: when the land parcel boundary is wider in actual situation (such as a road) and the image to be recognized is a high-resolution image, the predicted boundary of the first land parcel recognition result is a double-line boundary; when the land parcel boundary is narrower in actual situation (such as a ridge), the predicted boundary of the first land parcel recognition result is a single-line boundary. In order to avoid the influence of the polygonal vector object of the double-line boundary on the polygonal vector object of the land parcel during land parcel extraction, it is necessary to perform binarization processing and boundary skeleton extraction on the predicted boundary area in the first land parcel recognition result in turn, peel off the pixels representing the central axis of the land parcel layer by layer from the boundary area of the original land parcel, remove the pixels that are not the central axis in the original image, and maintain the shape of the original land parcel to obtain a boundary skeleton image, thereby shrinking the boundary of the land parcel from a surface to a line, and obtaining more accurate boundary information. The effect is as follows: Figure 6 shown.

[0098] In an embodiment of the present application, the grayscale enhancement of the image to be identified is performed band by band, and the enhanced image to be identified is obtained. The grayscale enhancement process is as follows: first, the image to be identified is divided into blocks and pixel-filled band by band, and then the mapping relationship between the original histogram of each block area image and the target histogram is calculated based on the restricted image contrast, wherein the target histogram is the histogram of the block image after histogram equalization, and finally, based on the mapping relationship between the original histogram and the target histogram, the contrast of the image to be identified is enhanced by the bilinear interpolation method to obtain the grayscale enhanced image to be identified. In this way, the details of the image to be identified are enhanced, the clarity of the land boundary in the image to be identified is improved, and the effect of suppressing background noise is achieved.

[0099] Then, the first boundary area in the enhanced image to be identified is uniformly assigned a value band by band, that is, a higher brightness value is assigned, such as a grayscale value of 255, to obtain the input image, where the first boundary area is the area corresponding to the position of the boundary area in the boundary skeleton image on the enhanced image to be identified, thereby obtaining an accurate plot boundary.

[0100] In this embodiment of the present application, a watershed algorithm is used to segment the input image based on the first seed image to obtain a refined land parcel identification result, which is raster land parcel data with refined land parcel boundaries. Because the first seed image contains accurate land parcel areas and non-land parcel areas, and the input image has precise land parcel boundaries, the watershed algorithm can accurately segment the image to be identified based on the first seed image and the input image, resulting in highly accurate and refined land parcel identification results.

[0101] In other optional embodiments, the method for fine land parcel identification based on boundary buffer deep learning further includes: performing vector conversion on the fine land parcel identification results to obtain land parcel vector data.

[0102] Among them, the vector conversion of the plot fine identification result is performed to obtain the plot vector data, specifically: the plot boundary in the plot fine identification result is converted into polyline vector data; the polyline vector data is converted into polygon vector data; according to multiple categories in the plot fine identification result, the category of each polygon in the polygon vector data is determined to obtain polygon vector data containing the plot category; the polygon vector data containing the plot category is optimized to obtain the plot vector data.

[0103] In an embodiment of the present application, the land parcel boundaries in the land parcel fine identification results are first converted into polyline vector data, and then the polyline vector data is converted into polygon vector data through the line-to-polygon algorithm in the geographic information system, thereby avoiding the topological errors caused by directly vectorizing the raster land parcel data with fine land parcel boundaries. Then, multiple categories in the land parcel fine identification results, i.e., identification information of land parcels and non-land parcels, are obtained according to the watershed algorithm, and attributes are assigned to each polygon data to obtain polygon vector data containing land parcel categories. Finally, the polygon vector data containing land parcel categories is optimized to obtain land parcel vector data. Among them, the polygon vector data is optimized, specifically: multiple small polygons of the same type with an area less than a preset threshold covered by a large polygon in the polygon vector data are merged with the large polygon, or multiple small polygons of the same type with an area less than a preset threshold connected are merged to obtain land parcel vector data. Among them, the specific value of the preset threshold can be determined by prior knowledge and passed into the method execution in the form of parameters. For example, the preset threshold can be set to a default of 200m 2 .

[0104] Figure 7 Schematic diagram of land parcel vector data obtained by the land parcel fine recognition method based on boundary buffer deep learning provided in some embodiments of the present application. Figure 7 It can be seen that the vector boundaries of the plot vector data obtained by the method provided in the embodiment of the present application are highly consistent with the plot boundaries in the high-resolution image to be identified, indicating that the plot identification is fine and accurate, and can meet the needs of automated plot extraction and rapid acquisition of unitized and objectified cultivated land plots, and has good practicality and application promotion value.

[0105] In summary, the method provided in the embodiment of the present application first identifies the plot area of the image to be identified through a deep learning model to obtain a first plot identification result; wherein, the deep learning model is trained based on pre-acquired image data and label data, and the label data is obtained by performing a first buffering process on the original label data; wherein, the first buffering process is to expand N pixels on both sides of the boundary of the plot in the original label data as the center line, where N is a positive integer greater than 1; then, the first binarized image is sequentially subjected to a second buffering process and a connected domain process to obtain a first seed image; wherein, the first binarized image is obtained by binarizing the plot area in the first plot identification result; the second buffering process is to shrink each pixel in the first binarized image by M pixels toward the center direction of the plot area in the first binarized image; M is determined based on the nearest distance between each pixel in the plot area in the first binarized image and the non-plot pixel in the first binarized image; M is a positive integer greater than 1; finally, based on the first seed image, the input image is segmented by a watershed algorithm to obtain a fine plot identification result. In this way, the image to be identified is preliminarily identified through the deep learning model to obtain the preliminary plot identification result of the plot boundary. Then, the exact area range of the plot and non-plot is determined as the seed image through the second buffer processing and connected domain processing. The image to be identified is enhanced based on the boundary skeleton image as the input image, and the watershed algorithm is performed to improve the accuracy of plot boundary extraction and plot identification.

[0106] Exemplary Systems

[0107] The present application also provides a land parcel fine recognition system based on boundary buffer deep learning. Figure 8 This is a structural diagram of the land parcel fine recognition system based on boundary buffer deep learning, as shown in Figure 8 As shown, the system includes: a land parcel identification unit 801, a buffer processing unit 802, and a fine segmentation unit 803.

[0108] The plot identification unit 801 is configured to identify the plot area of the image to be identified through a deep learning model to obtain a first plot identification result; wherein the deep learning model is trained based on pre-acquired image data and label data, and the label data is obtained by performing a first buffering process on the original label data; wherein the first buffering process is to expand N pixels on both sides of the boundary of the plot in the original label data as the center line, where N is a positive integer greater than 1.

[0109] The buffer processing unit 802 is configured to perform second buffer processing and connected domain processing on the first binary image in sequence to obtain a first seed image; wherein the first binary image is obtained by binarizing the land area in the first land identification result; the second buffer processing is to shrink each pixel in the first binary image by M pixels toward the center of the land area in the first binary image; M is determined based on the closest distance between each pixel in the land area in the first binary image and the non-land area pixel in the first binary image; M is a positive integer greater than 1.

[0110] The fine segmentation unit 803 is configured to: segment the input image according to the first seed image using the watershed algorithm to obtain a fine land parcel recognition result; wherein the input image is obtained by enhancing the boundaries of the image to be recognized based on the predicted boundary area in the first land parcel recognition result.

[0111] The land parcel fine identification system based on boundary buffer deep learning provided in the embodiment of the present application can implement the processes and steps of any of the above-mentioned land parcel fine identification methods based on boundary buffer deep learning, and achieve the same technical effects, which will not be described one by one here.

[0112] The foregoing description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A land parcel fine recognition method based on boundary buffer deep learning, characterized in that: include: Identify the land parcel area in the image to be identified using a deep learning model to obtain a first land parcel identification result; wherein the deep learning model is trained based on pre-acquired image data and label data, and the label data is obtained by performing a first buffering process on the original label data; wherein the first buffering process is to expand N pixels on both sides of the boundary of the land parcel in the original label data as a center line, where N is a positive integer greater than 1; The first binary image is sequentially subjected to a second buffering process and a connected component process to obtain a first seed image; wherein the first binary image is obtained by binarizing the land parcel area in the first land parcel identification result; the second buffering process is performed by shrinking each pixel in the first binary image by M pixels toward the center of the land parcel area in the first binary image; M is determined based on the closest distance between each pixel in the land parcel area in the first binary image and a non-land parcel pixel in the first binary image; M is a positive integer greater than 1; According to the first seed image, the input image is segmented by a watershed algorithm to obtain a fine land parcel recognition result; wherein, the input image is obtained by enhancing the boundary of the image to be recognized according to the predicted boundary area in the first land parcel recognition result.

2. The land parcel fine identification method based on boundary buffer deep learning according to claim 1 is characterized in that: The method further comprises: performing binarization processing and uniformly assigning grayscale values to the non-land parcel area in the first land parcel recognition result to obtain a non-land parcel area seed image; wherein the grayscale value of each pixel in the non-land parcel area seed image is different from the grayscale value of each pixel in the first seed image; appending the non-plot area seed image to the first seed image; According to the first seed image, the input image is segmented by the watershed algorithm to obtain the fine land recognition result.

3. The land parcel fine identification method based on boundary buffer deep learning according to claim 1 is characterized in that: According to the predicted boundary area in the first land parcel recognition result, the boundary of the image to be recognized is enhanced to obtain the input image, specifically: performing binarization processing and boundary skeleton extraction on the predicted boundary area in the first land parcel recognition result in sequence to obtain a boundary skeleton image; Performing band-by-band grayscale enhancement on the image to be identified to obtain an enhanced image to be identified; The first boundary area in the enhanced image to be identified is uniformly assigned band by band to obtain an input image, wherein the first boundary area is an area corresponding to the position of the boundary area in the boundary skeleton image on the enhanced image to be identified.

4. The land parcel fine identification method based on boundary buffer deep learning according to claim 1 is characterized in that: The method further comprises: Vector conversion is performed on the land parcel fine identification result to obtain land parcel vector data.

5. The land parcel fine identification method based on boundary buffer deep learning according to claim 4 is characterized in that: The vector conversion of the land parcel fine recognition result is performed to obtain land parcel vector data, specifically: Converting the land parcel boundary in the land parcel fine recognition result into polyline vector data; Converting the polyline vector data into polygon vector data; Determining the category of each polygon in the polygon vector data according to the multiple categories in the land parcel fine recognition result, and obtaining polygon vector data containing the land parcel category; The polygonal vector data containing the land parcel category is optimized to obtain land parcel vector data.

6. The land parcel fine identification method based on boundary buffer deep learning according to any one of claims 1 to 5, characterized in that: The deep learning model includes at least one convolutional attention module; The convolutional attention module is used to perform attention calculation on the feature map output by the residual network of the deep learning model, and input the obtained calculation results into the pyramid segmentation network of the deep learning model.

7. The land parcel fine identification method based on boundary buffer deep learning according to claim 6 is characterized in that: The convolutional attention module includes a channel attention submodule and a spatial attention submodule; The channel attention submodule is configured to compress the feature map output by the residual network in the spatial dimension to obtain a channel attention feature map; The spatial attention submodule is configured to globally weight the channel attention feature map to obtain an input feature map of the pyramid segmentation network.

8. The land parcel fine identification method based on boundary buffer deep learning according to claim 1 is characterized in that: During the training process, the deep learning model calculates the training loss value based on the average of the focus loss function and the Dice loss function.

9. The land parcel fine identification method based on boundary buffer deep learning according to claim 1 is characterized in that: The method further comprises: Training a plurality of convolutional neural networks according to the pre-acquired image data and the label data to obtain a plurality of candidate convolutional neural networks; The deep learning model is determined based on the parameter quantity of each candidate convolutional neural network and the accuracy of the candidate convolutional neural network.

10. A land parcel fine recognition system based on boundary buffer deep learning, characterized in that: include: A land parcel recognition unit is configured to identify a land parcel area in the image to be recognized using a deep learning model to obtain a first land parcel recognition result; wherein the deep learning model is trained based on pre-acquired image data and label data, and the label data is obtained by performing a first buffering process on the original label data; wherein the first buffering process is to expand N pixels on both sides of the boundary of the land parcel in the original label data as a center line, where N is a positive integer greater than 1; a buffer processing unit configured to sequentially perform a second buffer process and a connected domain process on the first binarized image to obtain a first seed image; wherein the first binarized image is obtained by binarizing the land parcel area in the first land parcel identification result; the second buffer process is to shrink each pixel in the first binarized image by M pixels toward the center of the land parcel area in the first binarized image; M is determined based on the closest distance between each pixel in the land parcel area in the first binarized image and a non-land parcel pixel in the first binarized image; and M is a positive integer greater than 1; The fine segmentation unit is configured to: segment the input image according to the first seed image using the watershed algorithm to obtain a fine land parcel recognition result; wherein the input image is obtained by enhancing the boundaries of the image to be recognized based on the predicted boundary area in the first land parcel recognition result.

Images