A method, device, equipment and medium for extracting plot boundaries

Abstract

The present application discloses a method, device, equipment and medium for extracting plot boundaries, relating to the field of remote sensing image processing. The method includes: performing plot boundary extraction on the satellite remote sensing image of the study area based on the cultivated land plot boundary extraction model to determine the plot boundary image of the study area; the plot boundary image includes a mask, a contour and a distance; the cultivated land plot boundary extraction model includes a residual neural network, a feature enhancement module and a convolutional output layer connected in sequence; the feature enhancement module includes a multi-scale feature integration block and a spatial group enhancement attention mechanism; the residual neural network performs feature extraction at different levels on the input image; the multi-scale feature integration block fuses the features at different levels; the spatial group enhancement attention mechanism performs spatial group enhancement on the fused features; the convolutional output layer performs mask prediction, contour extraction and distance estimation on the features after spatial group enhancement. The present application improves the extraction accuracy and efficiency of plot boundaries.

Description

Technical Field

[0001] The present application relates to the field of remote sensing image processing, and particularly to a method, device, equipment and medium for extracting plot boundaries by integrating multi-source high-resolution remote sensing satellite image information. Background Art

[0002] With the continuous growth of the global population and the acceleration of urbanization, new challenges have been posed to food security and land resource management. Accurately extracting the boundaries of cultivated land plots has become the key to scientific and technological assistance for land resource planning and agricultural development. Traditional ground observation methods have problems of low efficiency and high cost in extracting and monitoring the boundaries of cultivated land plots. Summary of the Invention

[0003] The purpose of the present application is to provide a method, device, equipment and medium for extracting plot boundaries, which can improve the extraction accuracy and efficiency of plot boundaries.

[0004] To achieve the above purpose, the present application provides the following solutions:

[0005] In the first aspect, the present application provides a method for extracting plot boundaries, including:

[0006] Obtaining satellite remote sensing images of the study area;

[0007] Performing plot boundary extraction on the satellite remote sensing images based on a cultivated land plot boundary extraction model to determine the plot boundary images of the study area; the plot boundary images include masks, contours and distances;

[0008] The cultivated land plot boundary extraction model is pre-trained using a training data set, the training data set includes multiple sample remote sensing images and the corresponding masks, contour maps and distance maps for each sample remote sensing image; the cultivated land plot boundary extraction model includes a residual neural network, a feature enhancement module and a convolutional output layer connected in sequence; the feature enhancement module includes a multi-scale feature integration block and a spatial group enhanced attention mechanism;

[0009] Among them, the residual neural network is used to perform feature extraction at different levels on the input image; the multi-scale feature integration block is used to fuse features at different levels; the spatial group enhanced attention mechanism is used to perform spatial group enhancement on the fused features; the convolutional output layer is used to perform mask prediction, contour extraction and distance estimation on the spatially group-enhanced features.

[0010] In the second aspect, the present application provides a device for extracting plot boundaries, including:

[0011] An image acquisition module, configured to acquire satellite remote sensing images of the study area;

[0012] A boundary extraction module, configured to perform plot boundary extraction on the satellite remote sensing image based on a cultivated land plot boundary extraction model to determine a plot boundary image of the study area; the plot boundary image includes a mask, a contour, and a distance;

[0013] The cultivated land plot boundary extraction model is pre-trained using a training data set, and the training data set includes multiple sample remote sensing images and a mask, a contour map, and a distance map corresponding to each sample remote sensing image; the cultivated land plot boundary extraction model includes a residual neural network, a feature enhancement module, and a convolutional output layer connected in sequence; the feature enhancement module includes a multi-scale feature integration block and a spatial group enhanced attention mechanism;

[0014] Among them, the residual neural network is used to perform feature extraction at different levels on the input image; the multi-scale feature integration block is used to fuse features at different levels; the spatial group enhanced attention mechanism is used to perform spatial group enhancement on the fused features; the convolutional output layer is used to perform mask prediction, contour extraction, and distance estimation on the spatially group-enhanced features.

[0015] In a third aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the above-mentioned plot boundary extraction method.

[0016] In a fourth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned plot boundary extraction method is implemented.

[0017] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0018] The present application provides a plot boundary extraction method, device, equipment, and medium. The constructed cultivated land plot boundary extraction model integrates multi-task learning, a residual neural network, and a feature enhancement module, and particularly introduces a multi-scale feature integration block and a spatial group enhanced attention mechanism, significantly improving the extraction accuracy and efficiency of plot boundaries, especially outperforming other deep learning methods in dealing with complex boundaries and features at different scales. Description of the Drawings

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 It is an application environment diagram of a method for extracting plot boundaries in an embodiment of the present application;

[0021] Figure 2 It is a schematic flowchart of a method for extracting plot boundaries provided in an embodiment of the present application;

[0022] Figure 3 It is a schematic diagram of the generation process of a contour map and a distance map provided in an embodiment of the present application;

[0023] Figure 4 It is a schematic diagram of a cultivated land plot boundary extraction model provided in an embodiment of the present application;

[0024] Figure 5 It is a diagram of the ablation experiment results of different models;

[0025] Figure 6 It is a schematic diagram of the results of boundary extraction by different models;

[0026] Figure 7 It is a schematic diagram of the functional modules of a plot boundary extraction device provided in an embodiment of the present application. Detailed implementation manners

[0027] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0028] The development of remote sensing technology has provided a new solution for the efficient, low-cost and high spatio-temporal resolution monitoring of cultivated land plots. The related remote sensing image plot boundary extraction methods mainly include traditional machine learning methods and deep learning methods.

[0029] Traditional machine learning methods usually distinguish cultivated land by formulating some rules and thresholds or analyzing image attributes, which is simple to operate but sensitive to noise. With the progress of technology, traditional machine learning methods are gradually being replaced by deep learning methods, which provide higher automation and adaptability.

[0030] Deep learning methods use deep learning models based on Convolutional Neural Network (CNN) to automatically extract plot boundaries, which can maintain high accuracy in complex or changing environments. This technology requires a large amount of computing resources and a large amount of training data.

[0031] With the continuous improvement of computing power and the strong support of high-resolution remote sensing images, in the field of extracting the boundaries of land parcels from remote sensing images, the extraction of cultivated land parcel boundaries based on deep learning methods is the focus of current research and development by scholars. At the same time, the fusion of multi-source heterogeneous data, the universality and efficiency of data processing, etc. have also been continuously enhanced.

[0032] For example, in the convolutional neural network of deep learning methods, U-Net is a popular deep learning architecture. Its uniqueness lies in its "U" shape structure, which includes a contracting path and a symmetric expanding path. This design helps the network capture certain detailed information while maintaining context information. However, when dealing with very small or slender structures, the model may not be able to fully capture enough details, which may be a limitation for the task of extracting cultivated land boundaries from high-resolution remote sensing images.

[0033] In deep learning methods, the semantic segmentation network (SegNet) uses an encoder-decoder architecture and uses the max-pooling index of the encoder stage to upsample the feature map in the decoder, thereby reducing the number of model parameters and improving the segmentation accuracy. However, SegNet is more sensitive to noise in the image, which will affect its performance in the actual application scenarios of remote sensing images.

[0034] The multi-scale attention network (MA-Net) is an improved deep learning architecture designed for medical image segmentation. MA-Net introduces an attention mechanism, which enhances the model's ability to focus on key features, thereby improving its segmentation accuracy. Although the introduction of the attention mechanism improves the feature extraction ability, it also increases the complexity of the model, limiting its application in resource-constrained environments. In addition, the optimization and training of MA-Net are more complex than standard convolutional networks, and more detailed training strategies and parameter adjustments are required to achieve optimal performance.

[0035] The fully convolutional neural network (FCN) is an image segmentation model that was first proposed to replace traditional fully connected layers with convolutional layers. It allows the network to process input images of any size and directly output the corresponding size segmentation map, making FCN efficient and flexible in dealing with image segmentation tasks. However, it has certain limitations in upsampling and feature fusion and does not make full use of multi-scale information, resulting in insufficient generalization ability in new scenarios with slight differences in actual use, and thus a decline in performance.

[0036] The Multi-task Network Based on Plot Extraction (BsiNet) adopts a multi-task learning architecture, emphasizing boundary accuracy in the image segmentation process. The main task of this model focuses on improving segmentation accuracy, while the additional task focuses on refining boundary processing. However, BsiNet is mainly applicable to specific application scenarios that require precise boundary analysis. In addition, the structure of this model leads to high computational resource requirements, including significant memory and processing time, and it is difficult to maintain performance in resource-constrained environments.

[0037] ResUNet-a combines the feature extraction ability of the residual network and the effective segmentation structure of U-Net. This structure enhances feature transmission through residual connections, making it perform well in processing images with complex backgrounds and details. However, due to the model containing multiple layers of residual blocks and a large number of parameters, its demand for computational resources is high, and the generality and usability of the model are poor in real-time or resource-constrained application scenarios.

[0038] In addition, when processing high-resolution multi-source image data, there are often problems such as insufficient data fusion, low boundary extraction accuracy, and low operation efficiency. For example, multi-source data often includes data with different bands and resolutions, such as high-resolution satellite images and Light Laser Detection and Ranging (LiDAR) data, radar images, etc. In these high-resolution images, the boundaries may be very complex and may exhibit different characteristics in different bands or sensors, which may prevent boundary detection methods from accurately capturing these complex boundaries; similarly, high-resolution multi-source image data is very large, and deep learning methods require huge storage space and computational resources for data access and operation.

[0039] Aiming at the problems of insufficient image data fusion, low boundary extraction accuracy, and low operation efficiency in multi-source high-resolution image data, this application mainly focuses on strategies such as designing effective feature extraction, fusion, and enhancement, and reducing resource waste.

[0040] To make the above objects, features, and advantages of this application more obvious and understandable, the following further detailed description of this application will be given in conjunction with the accompanying drawings and specific embodiments.

[0041] The plot boundary extraction method provided by the embodiments of this application can be applied to, for example Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set separately, integrated on the server 104, placed on the cloud or other servers. The terminal 102 can send the satellite remote sensing images of the research area to the server 104. After receiving the satellite remote sensing images of the research area, the server 104 performs plot boundary extraction on the satellite remote sensing images based on the cultivated land plot boundary extraction model to determine the plot boundary images of the research area. The server 104 can feedback the plot boundary images to the terminal 102. In addition, in some embodiments, the plot boundary extraction method can also be implemented separately by the server 104 or the terminal 102.

[0042] Among them, the terminal 102 can be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0043] In an exemplary embodiment, as Figure 2 shown, a method for extracting plot boundaries is provided. This method is executed by a computer device, and can be specifically executed alone by a computer device such as a terminal or a server, or jointly executed by a terminal and a server. In the embodiments of the present application, taking this method applied to Figure 1 the server 104 in as an example for illustration, it includes the following steps 201 and 202.

[0044] Step 201, obtain the satellite remote sensing images of the research area.

[0045] Step 202, perform plot boundary extraction on the satellite remote sensing images based on the cultivated land plot boundary extraction model to determine the plot boundary images of the research area. The plot boundary images include masks, contours, and distances.

[0046] The cultivated land plot boundary extraction model is obtained by pre-training with a training data set. The training data set includes multiple sample remote sensing images and the corresponding masks, contour maps, and distance maps for each sample remote sensing image.

[0047] In an exemplary embodiment, the training process of the cultivated land plot boundary extraction model includes the following steps 301 to 308.

[0048] Step 301, obtain the remote sensing images of a preset area within multiple preset time periods to obtain multiple preliminary remote sensing images.

[0049] In an exemplary embodiment, first, remote sensing detection information of a resource satellite in a preset area within multiple preset time periods is obtained. The remote sensing detection information of the resource satellite includes geographical information, airspace meteorological environment information, and remote sensing images. Then, denoising, cloud removal, geometric correction, and orthorectification processing are sequentially performed on each remote sensing image to obtain multiple preprocessed images. Next, multiple preliminary remote sensing images are determined based on the geographical information, airspace meteorological environment information, and preprocessed images corresponding to each remote sensing image.

[0050] Among them, first, the collected remote sensing images are denoised to remove random noise in the images. Then, cloud removal technology is applied to process the remote sensing images to remove clouds and shadows that affect image clarity. Next, the geographical distortion of the images is corrected to ensure the geographical location accuracy of the images. After that, the images are converted to a ground coordinate system to eliminate the influence of the inclination of buildings and terrain. Finally, images of different bands or different time points are fused to form a complete image of the study area.

[0051] Step 302: Perform image fusion on multiple preliminary remote sensing images to obtain a fused image.

[0052] Step 303: Crop the fused image to obtain multiple cropped images. Specifically, 20 1000×1000 images that best represent the characteristics of cultivated land and 10 1000×1000 images that clearly distinguish cultivated land from urban lines are cropped from the fused image to ensure that these images are evenly covered in the study area. Among them, the characteristics of cultivated land are identified according to the color of the products produced in the regional area or the observed undulation of the ridges of cultivated land by humans.

[0053] Step 304: Draw vector data of the cultivated land plot boundaries based on multiple cropped images, and perform rasterization processing on the vector data of the cultivated land plot boundaries to obtain rasterized images of each cropped image. The rasterized image is a mask corresponding to the cropped image.

[0054] Specifically, with the help of the 30 cropped 1000×1000 images, professional GIS software (such as ArcMap) is used to manually draw high-standard vector data of the cultivated land plot boundaries to form high-precision vector data. Then, through a self-built GDAL script library in Python language, rasterization processing is performed on the drawn vector data, and the raster data has the same resolution and alignment method as the original image.

[0055] Step 305: Sample multiple cropped images and the rasterized images corresponding to each cropped image using an overlapping sliding window to obtain multiple sample remote sensing images and masks of each sample remote sensing image.

[0056] Specifically, use an overlapping sliding window to uniformly crop 30 cropped images and their corresponding 30 rasterized images, and divide the entire image into multiple 256×256 remote sensing images to obtain the most primitive sample set.

[0057] Step 306: According to multiple sample remote sensing images, use the Euclidean distance transformation formula and the quasi-Euclidean distance transformation formula to determine the contour map and distance map of each sample remote sensing image. The contour map and distance map together with the most primitive sample set constitute the model data set.

[0058] Specifically, use the library function of scipy and implement the Euclidean distance transformation and the quasi-Euclidean distance transformation with Python code. Among them, the Euclidean distance transformation formula is:

[0059] The quasi-Euclidean distance transformation formula is:

[0060] Among them, d1 is the Euclidean distance, d2 is the quasi-Euclidean distance, x i and y i are the feature vectors of two individuals respectively, V is the variance of the corresponding feature vectors of the two individuals (it should be noted that the variance calculation method here is not to divide by n, but to divide by n - 1), and n is the number of individuals.

[0061] The generation process of the contour map and distance map is as Figure 3 shown.

[0062] Step 307: Perform geometric transformations on multiple sample remote sensing images, the contour map and distance map of each sample remote sensing image to obtain the training data set.

[0063] Specifically, perform operations such as horizontal flipping, 180-degree rotation, 90-degree clockwise rotation, and data augmentation on the sample remote sensing images, contour maps, and distance maps without changing the size of the images. Together with the most primitive sample set, they constitute the training data set, and perform unified standardized naming according to information such as the area (row and column) where the images are taken.

[0064] Step 308: Use the training data set to train the cultivated land plot boundary extraction model to obtain the trained cultivated land plot boundary extraction model.

[0065] Specifically, use the five-fold cross-validation method to verify and test the cultivated land plot boundary extraction model. Use the cultivated land plot boundary extraction model with the optimal round in the test results to complete the boundary extraction at the plot level in the study area, and perform raster vectorization and merging processing to form vector data.

[0066] Effect evaluation is carried out according to the evaluation criteria defined in this application, including Overall Accuracy (OA), Precision (P), Recall (R), Intersection over Union (IoU), and F1-score. According to the evaluation indicators, the optimal model parameter configuration is selected to obtain the final boundary prediction results of all sub-images.

[0067] This application continuously learns and improves the ability to extract plot boundaries during multiple iterative trainings of the training dataset. Compared with the related deep learning-based cultivated land boundary extraction models, this application retains the applicability of extracting plot boundaries from large-scale remote sensing images, reduces the parameters and computational amount of the model, improves the overall performance and computational efficiency of plot boundary extraction, and provides more possibilities and opportunities for the association between spatial geographic information and cultivated land plot information.

[0068] As Figure 4 shown, the cultivated land plot boundary extraction model includes a Residual Neural Network (ResNet), a feature enhancement module, and a convolutional output layer connected in sequence. The feature enhancement module includes a Multi-Scale Feature Integration Block (MSFI) and a Spatial Group-wise Enhance (SGE).

[0069] (1) The Residual Neural Network is used to extract features at different levels from the input image.

[0070] In an exemplary embodiment, the Residual Neural Network includes multiple layers of residual blocks, and each layer of residual block includes multiple convolutional layers and skip connections. Each convolutional layer gradually delves deeper into the image by extracting features at different levels. These feature maps capture the details to abstract features of the image from primary to high-level and output them to the subsequent feature enhancement module for processing. The Residual Neural Network uses the following formula to extract features at different levels from the input image:

[0071] x l =x l-1 +F(x l-1 ,W l );

[0072] where, x l is the output feature of the l-th layer of residual block, x l-1 is the output feature of the (l - 1)-th layer of residual block, W l is the weight of the l-th layer of residual block, F(xl-1 ,W l ) is the overlap of multiple residual layers, where 0 < l ≤ L and L is the number of layers in the residual block. When l = 0, the output feature of the 0th residual block is the image input to the residual neural network.

[0073] (2) The multi-scale feature integration block is used to fuse features at different levels, which helps the model capture the detailed information of the cultivated land plots.

[0074] The fused feature map M of the (l - 1)th layer l-1 is: M l-1 = MSFI(x l-1 ,M l ). First, the deepest feature map extracted by the residual neural network is upsampled (the upsampling factor is 2). The fused feature map of the previous layer is obtained by processing the combination of the layer above the deepest layer of the corresponding residual neural network and M l .

[0075] In an exemplary embodiment, the process of the multi-scale feature integration block fusing features at different levels includes:

[0076] 1) For the lth layer, the deep feature map of the (l + 1)th layer is upsampled to obtain the fused feature map of the (l + 1)th layer. When l = L, the fused feature map of the (L + 1)th layer is the output feature of the Lth residual block.

[0077] 2) The fused feature map of the (l + 1)th layer is combined with the output feature of the lth residual block to obtain the combined feature map of the lth layer.

[0078] Specifically, first, a deep feature map In ∈ R H×W×C and a low-level feature map Res-Layer are input. In is upsampled to obtain the fused feature map In” ∈ R 2H×2W×C . Then, using the concat function, In” ∈ R 2H×2W×C and the low-level feature map Res-Layer are combined into ln' ∈ R 2H×2W×2C .

[0079] 3) The combined feature map of the lth layer is subjected to multi-scale fusion to obtain the fused feature map of multiple scales of the lth layer.

[0080] Specifically, In is subjected to multi-scale fusion to respectively generate and The formulas are as follows:

[0081]

[0082] Among them, It is a fused feature map of three scales. In' is the feature map after merging in the l-th layer, and In'1 is the result after processing by Conv 1×1 (In'). Conv 1×1 represents a 1×1 convolution operation. Conv 3×3 represents a 3×3 convolution operation. BN represents batch normalization operation. (H×W×C') represents the size of the feature map and the number of output channels.

[0083] 4) Sum the fused feature maps of multiple scales in the l-th layer element-wise and perform activation processing to obtain the deep feature map of the l-th layer. Specifically, the RELU activation function is used for activation processing. Repeat the multi-scale feature integration, and finally output a feature map of 256×256×128.

[0084] (3) The spatial group enhanced attention mechanism is used to perform spatial group enhancement on the fused features. Using the spatial group enhanced attention mechanism pays more attention to the key regions in the image, such as the fine boundary regions, while suppressing the irrelevant background information. The formula of the spatial group enhanced attention mechanism is: F fused = SGE(M1, x1); where F fused is the feature after spatial group enhancement.

[0085] The process of performing spatial group enhancement by the spatial group enhanced attention mechanism includes:

[0086] ① Divide the convolutional feature matrix (i.e., the fused feature) Y with length H, width W, and channel dimension C into k groups, where Y = {y1, y2, y3, …, y M}, M = H×W, and represents the feature vector at a specific point in the H×W space. The importance of each point is re-determined by calculating the similarity between the feature vector of each point and the global feature vector g, and g is obtained by performing spatial averaging on all points:

[0087]

[0088] where f represents the spatial averaging function.

[0089] ② Determine the importance coefficient c m of y m . To avoid the influence of the coefficient size deviation between different samples, the importance coefficient c m is further normalized to obtain the normalized importance coefficient

[0090] c m = g·y m = ||g||·cos||y m ||·cos(θm )

[0091]

[0092] where θ m is the angle between y m and g in the feature space, u d is the mean value of the importance coefficient c m , is the variance of the importance coefficient c m , and ∈ is a constant to ensure the denominator is positive.

[0093] ③ Use a linear function to adjust to ensure that the normalization can represent the identity transformation and the normalization operation can be restored, obtaining the adjusted coefficient where α and β are function parameters. Here, α = β = k = 32.

[0094] ④ Scale the original feature vector y m by applying the Sigmoid function δ to the adjusted coefficient p m to obtain the enhanced feature vector

[0095] ⑤ The enhanced feature vectors obtained at this time form a feature map of 256×256×3. In the tasks of mask prediction and contour extraction, a softmax classifier is used to obtain the prediction result map, and in the distance map estimation task, a sigmoid classifier is used to obtain the prediction result map.

[0096] (4) The convolutional output layer is used for mask prediction, contour extraction, and distance estimation of the features enhanced by the spatial population.

[0097] Aiming at the problems of insufficient image data fusion, low boundary extraction accuracy, and low operation efficiency in multi-source high-resolution image data, this application adopts a deep learning method strategy and designs a cross-fusion multi-task learning network (CFMT-Net) including three core modules: multi-task learning, feature extraction, and feature enhancement. Through this network, a cultivated land plot boundary extraction model is constructed to improve the data fusion effect and boundary extraction effect of the cultivated land plot boundary extraction model. The cultivated land plot boundary extraction model demarcates cultivated land from high-resolution satellite images, adopts a multi-task learning method, and combines a residual neural network) and an improved feature enhancement module.

[0098] Among them, the multi-task learning strategy takes mask prediction as the core task, while contour extraction and distance map estimation as auxiliary tasks to improve the shape and boundary of the primary task, so as to further refine the prediction of the farmland mask and enable the model to obtain better generalization ability and stability.

[0099] Feature extraction is performed through a residual neural network, which not only increases the depth and complexity of feature extraction. The features include spatial location information, form information, etc. in the image, but also reduces the parameters and computational amount of the model, enabling the model to have a faster calculation speed.

[0100] After that, through the feature enhancement module, features at different levels are fused to capture more complex and abstract image features; in this process, the spatial resolution of the image is gradually restored through a series of upsampling layers and skip connections and different-level features are combined to restore the detailed information of the image.

[0101] MSFI and SGE are designed and introduced in the feature enhancement module. MSFI can capture feature information in different regions and at different scales for integration, weakening the influence of background noise. SGE constructs a single decoder from the decoders of three parallel tasks, enhancing the detection ability for complex boundaries, enabling the model to have fewer parameters and good accuracy, and thus improving the model's detection ability for complex boundaries.

[0102] Finally, the final output image is generated through the convolutional output layer, making CFMT-Net have better generalization and stability than traditional CNNs. During the network training process, the contour extraction and distance map estimation tasks are used to assist in optimizing the main mask prediction task. In addition, mask and boundary prediction are classification problems, and distance map estimation is a regression problem. During the model training process, the loss function (such as negative log-likelihood loss and regression error loss) is calculated based on the output of the model and the true labels, and the weights of the network are updated using the Adam optimizer to optimize the model performance. This application continuously learns from the training dataset during multiple iterative trainings and improves the ability to extract plot boundaries.

[0103] This application integrates the multi-task learning method during the training process, generates more sample data information through the distance formula for the training samples to be added to the model training, and combines the residual neural network and the feature enhancement module to improve the model's generalization ability and the effect of plot boundary extraction, improving the accuracy and efficiency of plot boundary extraction to a certain extent, and also providing more possibilities and opportunities for the association between spatial geographic information and cultivated land plot information.

[0104] In another exemplary embodiment, the plot boundary extraction method further includes the following steps 203 to 205.

[0105] Step 203: Vectorize the plot boundary images in the study area to obtain vector data of the study area.

[0106] Step 204: Conduct overlay analysis on the vector data of the study area and the non-cultivated land vector data, and remove non-cultivated land features from the vector data of the study area to obtain cultivated land vector data.

[0107] Among them, the non-cultivated land vector data includes water system surfaces, road network surfaces, and urbanization building points. From the intuitive perspective of human vision, the water systems and road network surfaces are the first to be observed in satellites. In addition, during the process of urbanization, there is no large-scale agricultural cultivation in cities. Based on this, use the arcmap image tool to cover the high-resolution images, outline the water systems (main and branch rivers), road networks (main roads, branch roads), and towns (urbanized villages or cities), and erase the overlapping parts. The specific overlay and erasure steps can be implemented in sequence in the arcmap tool.

[0108] Step 205: Repair boundary breaks in the cultivated land vector data to obtain the cultivated land boundaries in the study area. Specifically, repair the possible boundary breaks or discontinuities that may occur after the erasure process, and finally obtain all the repaired and accurate cultivated land boundaries in the study area.

[0109] In summary, the beneficial effects of this application at least include the following points:

[0110] (1) Improve the calculation efficiency:

[0111] The data preprocessing process of this application includes: image fusion of multiple resource satellite remote sensing images; cropping out images that conform to the study area; drawing vector data of the cultivated land plot boundaries of the images; standardizing the sample data, rasterizing the vector data, enhancing, etc. Among them, in the case of deep learning training, there is a lack of data, and geometric transformation is needed to enhance the sample set data, including operations such as horizontal flipping, 180-degree rotation, and 90-degree clockwise rotation, without changing the size of the image, and together with the most original sample set, they form an experimental data set. By using rasterized label data in this application, rasterized data helps with more spatial analysis and image processing operations, captures pixel-level spatial relationships, ensures data consistency, and improves the calculation efficiency.

[0112] (2) Reduce the impact of overfitting. Overfitting often occurs in the model training data in supervised learning methods, resulting in the inability to generalize to new data. Through the multi-task learning method in this application, more sample data information is generated from the training data through the distance formula and added to the model training to improve the generalization ability of the model and make it perform better on new data.

[0113] (3) Consider identifying deeper features and characteristics within the recognition samples. The residual neural network adopted in this application can increase the network depth, extract more abstract features while reducing the size, and utilize the feature fusion module to combine features at different levels to capture more complex and abstract image characteristics. These features are crucial for understanding the image content and performing accurate plot boundary extraction.

[0114] (5) Design and train a multi-task learning CNN network (mask prediction, contour extraction, and distance map estimation) for co-training, requiring the network to be complementary in feature extraction and information processing and maintain prediction consistency. Introduce a combination of multiple losses to ensure that the networks learn from each other's strengths in different tasks during training and maintain the consistency of the prediction results to achieve deep co-learning.

[0115] (6) Improve the running efficiency: The information of high-resolution multi-source image data is extremely large, and the method requires a large amount of storage space and computing resources for data access and operations. By using residual connections, this application reduces the computational amount, synthesizes different decoders of multiple tasks into a single decoder to complete multi-task learning, optimizes the network structure and parameter operations to reduce model parameters, and improves the overall performance and operation efficiency.

[0116] To verify the effectiveness of this application, ablation experiments were conducted on CFMT-Net. In the ablation experiments, the baseline model gradually introduced different components, and the processing effects of the model on fine regions such as segmenting the boundary between arable land and non-arable land and thin or elongated boundaries were observed and compared. As Figure 5 shown, Figure 5 in the first column from left to right in

[0117] are the ablation experiment results of the baseline model, the second column is the ablation experiment results of the baseline model + feature extraction module, the third column is the ablation experiment results of the baseline model + feature enhancement module, the fourth column is the ablation experiment results of the baseline model + multi-task learning, the fifth column is the ablation experiment results of the baseline model + multi-task learning + feature enhancement module, the sixth column is the ablation experiment results of this application, the seventh column is the schematic diagram of the mask, and the eighth column is the input remote sensing image. Figure 5 The baseline model performs well in the segmentation tasks of simple arable land and vast arable land, but performs poorly in the processing of fine structures and complex boundaries. As

[0118] shown in the first column of Figure 5 the misclassification problem between arable land and non-arable land is obvious.

[0119] The improved feature enhancement module replaces the decoder of the baseline model, enabling the model to reduce over-segmentation when dealing with the transition area between arable land and non-arable land, making the fine boundaries and edges finer and smoother. For example, Figure 5 the slender boundary lines in the second and third rows of the third column in

[0120] Based on the baseline model, the integration of multi-task learning strategy enhances the model's overall understanding of images, improves the consistency and accuracy of boundaries, and reduces misjudgment phenomena. For example, Figure 5 in the fourth column of

[0121] The model that combines multi-task learning and feature enhancement, although there are cases of breaks or continuous points at slender boundaries, has significantly outperformed the baseline model that combines a single strategy in accurately extracting arable land boundaries. For example, Figure 5 there are breakpoints at the slender boundary in the center of the image and the simple arable land boundary in the second row of the fifth column in

[0122] Finally, CFMT-Net combines all optimization strategies and performs excellently in extracting arable land boundaries in complex scenarios. Its output highly coincides with the real mask, showing the accuracy and consistency of the model in dealing with complex scenarios to a certain extent. For example, Figure 5 as shown in the sixth column of

[0123] The evaluation metrics under different model methods are shown in Table 1. When the input image size is 3×256×256, the value of the floating-point operation count (Giga Floating-point Operations Per Second, GFLOPs) of CFMT-Net is the smallest (19.41), which is more efficient in computing efficiency compared to other models and can complete the same task with fewer computing resources, especially when running on devices with limited resources. Although the number of model parameters of CFMT-Net proposed in this application is more than that of the MA-Net model, it is still significantly lower than models such as FCN and ResUNet-a, reducing the storage requirements of the model and improving the loading speed to a certain extent. The combination of this low GFLOPs and reasonable number of parameters makes CFMT-Net suitable for deployment on platforms with limited storage and memory, and can also meet the requirements of high-performance computing environments, being very suitable for real-time or near-real-time application scenarios that require fast response.

[0124] Table 1 Comparison of evaluation metrics under different model methods

[0125]

[0126] Among them, the floating-point operation count represents the number of giga floating-point operations per second, with the unit of G, and the number of parameters represents the complexity of the model, with the unit of M. 1M is equal to 10 6 .

[0127] Table 2 shows the performance comparison of CFMT-Net and seven other convolutional neural network models in extracting the cultivated land boundary on the Jilin-1 image dataset, including the results of overall accuracy, precision, recall, F1-score, and intersection over union (IoU).

[0128] In terms of overall accuracy, CFMT-Net is the highest, reaching 88.14%, indicating that the boundary extraction of this application is more excellent in terms of the accuracy performance of all test samples.

[0129] In terms of precision, the precision of CFMT-Net is 88.63%, higher than the precision of ResUNet-a, which is 87.63%. It is more excellent in the ability to avoid mislabeling non-cultivated areas.

[0130] In terms of recall, CFMT-Net leads other models, reaching 89.86%. It is very effective in identifying real cultivated land areas and can capture all relevant areas to the greatest extent.

[0131] In terms of F1-Score, CFMT-Net leads with an F1-Score of 88.95%, further confirming the superior performance of the model in accurately identifying the cultivated land boundary.

[0132] In terms of IoU, compared with the IoU results obtained by U-Net, SegNet, FCN, MA-Net, BsiNet, and ResUNet-a, the IoU of CFMT-Net increases by 27.04%, 8.49%, 5.91%, 3.85%, 4.94%, and 1.29% respectively. Overall, it is better than other models and has better performance.

[0133] Generally speaking, CFMT-Net performs better than the comparison models in these key performance indicators, especially showing excellent accuracy and consistency in identifying and accurately marking the cultivated land boundary.

[0134] Table 2 Comparison of experimental result indicators under different model methods

[0135] Model OA (%) P(％) R(％) F1-Score (%) IoU (%) U-Net 72.72 79.77 62.42 62.94 53.33 SegNet 82.33 82.39 84.27 82.90 71.88 FCN 83.88 82.74 87.31 84.72 74.46 MA-Net 85.90 87.71 85.91 85.95 76.52 BsiNet 84.45 83.27 88.71 85.41 75.43 ResUNet-a 87.33 87.46 88.98 87.78 79.08 CFMT-Net 88.14 88.63 89.86 88.95 80.37

[0136] As Figure 6 shown, the ground truth masks show the target segmentation regions that each model aims to replicate. These masks are binary, and the white regions represent the target regions that need to be segmented from the background. Figure 6The first row is the input remote sensing image, the second row is the mask, the third row is the result obtained by this application, the fourth row is the result obtained by ResUNet-a, the fifth row is the result obtained by MA-Net, the sixth row is the result obtained by FCN, the seventh row is the result obtained by SegNet, and the eighth row is the result obtained by U-Net. From the visual comparison, it can be observed that the prediction result of CFMT-Net shows a high degree of agreement with the real mask. Its segmentation boundary is clearly defined, and it rarely misjudges the background as the target, showing its high precision and accuracy.

[0137] In Figure 6 In the first, third, fourth, and fifth columns of, the two models of CFMT-Net and ResUNet-a show high precision when dealing with the boundaries of small or slender regions. However, ResUNet-a has over-segmentation (the fourth column) and under-segmentation (the third and fifth columns) at the connection points and corners of the slender boundaries, which makes the boundaries appear slightly rough. In contrast, MA-Net, FCN, and several other models perform mediocrely when dealing with these fine boundaries and fail to achieve the ideal segmentation effect.

[0138] In Figure 6 In the second, sixth, and seventh columns of, the segmentation results of cultivated land and towns, cultivated land and forests, and cultivated land and cultivated land can be observed. The prediction results of CFMT-Net are relatively close to the real mask and highly consistent, with fewer over-segmentation and under-segmentation problems. In contrast, ResUNet-a misjudges non-cultivated areas as cultivated land in some cases, showing a certain degree of over-segmentation, which may be due to the excessive generalization ability of the model in identifying the texture and color of forest areas. Although MA-Net conforms to the boundary curve of the real mask, its overall accuracy is relatively low. FCN and SegNet provide good segmentation results in most cultivated land areas. However, at the complex boundaries between different land types, such as the boundaries between cultivated land and towns or forests, they often lack sufficient perception ability to accurately divide different land types. In addition, U-Net often shows under-segmentation problems when dealing with continuous or complex terrain changes and may fail to fully capture small plots or continuous terrain changes. These results show the effects of each model in dealing with image boundaries in practical applications, visually showing that the prediction result of CFMT-Net has a high degree of agreement with the real mask, further proving that CFMT-Net has strong feature extraction and boundary recognition capabilities.

[0139] In summary, the present application has conducted extensive experimental verifications on the publicly available competition dataset (Huawei iFlytek Jilin-1 image dataset) and the self-constructed experimental dataset. The experimental results show that the CFMT-Net provided by the present application demonstrates excellent boundary extraction capabilities on the dataset. It not only achieves remarkable results in improving the accuracy and boundary clarity of the model, but also the efficient computing power of the model opens up new possibilities for the in-depth association and wide application of spatial geographic information and cultivated land plot information.

[0140] Based on the same inventive concept, the embodiments of the present application also provide a plot boundary extraction device for implementing the plot boundary extraction method involved above. The solution provided by this device for solving problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more of the following embodiments of the plot boundary extraction device can refer to the limitations on the plot boundary extraction method in the above text, and will not be elaborated here.

[0141] In an exemplary embodiment, as Figure 7 shown, a plot boundary extraction device is provided, including: an image acquisition module 701 and a boundary extraction module 702.

[0142] Among them, the image acquisition module 701 is used to acquire satellite remote sensing images of the study area.

[0143] The boundary extraction module 702 is used to perform plot boundary extraction on the satellite remote sensing image based on the cultivated land plot boundary extraction model to determine the plot boundary image of the study area; the plot boundary image includes a mask, a contour, and a distance.

[0144] The cultivated land plot boundary extraction model is pre-trained using a training dataset. The training dataset includes multiple sample remote sensing images and the corresponding mask, contour map, and distance map for each sample remote sensing image. The cultivated land plot boundary extraction model includes a residual neural network, a feature enhancement module, and a convolutional output layer connected in sequence. The feature enhancement module includes a multi-scale feature integration block and a spatial group enhancement attention mechanism.

[0145] Among them, the residual neural network is used to perform feature extraction at different levels on the input image. The multi-scale feature integration block is used to fuse features at different levels. The spatial group enhancement attention mechanism is used to perform spatial group enhancement on the fused features. The convolutional output layer is used to perform mask prediction, contour extraction, and distance estimation on the spatially group-enhanced features.

[0146] In another exemplary embodiment, the plot boundary extraction device further includes: a vector processing module 703, a removal module 704, and a fracture repair module 705.

[0147] The vector processing module 703 is used to perform vectorization processing on the plot boundary images in the study area to obtain the vector data of the study area.

[0148] The removal module 704 is used to perform overlay analysis on the vector data of the study area and the non-cultivated land vector data, and remove the non-cultivated land features from the vector data of the study area to obtain the cultivated land vector data.

[0149] The fracture repair module 705 is used to repair the boundary fractures in the cultivated land vector data to obtain the cultivated land boundary of the study area.

[0150] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0151] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0152] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0153] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0154] In this application, all actions of obtaining signals, information, or data are carried out on the premise of complying with the corresponding data protection regulations and policies of the country where the location is located and obtaining the authorization given by the owner of the corresponding device.

[0155] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tapes, floppy disks, flash memories, optical memories, high-density embedded non-volatile memories, resistive random access memories (ReRAM), magnetoresistive random access memories (MRAM), ferroelectric random access memories (FRAM), phase change memories (PCM), graphene memories, etc. Volatile memories can include random access memories (RAM) or external cache memories, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0156] The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., without limitation.

[0157] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

[0158] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A method for extracting land parcel boundaries, characterized in that: The land parcel boundary extraction method comprises: Obtain satellite remote sensing images of the study area; Extracting the plot boundary of the satellite remote sensing image based on the cultivated land plot boundary extraction model to determine the plot boundary image of the study area; the plot boundary image includes a mask, a contour and a distance; The farmland plot boundary extraction model is obtained by pre-training with a training data set, wherein the training data set includes a plurality of sample remote sensing images and masks, contour maps and distance maps corresponding to each sample remote sensing image; the farmland plot boundary extraction model includes a residual neural network, a feature enhancement module and a convolution output layer connected in sequence; the feature enhancement module includes a multi-scale feature integration block and a spatial group enhancement attention mechanism; The residual neural network is used to extract features at different levels from the input image; the multi-scale feature integration block is used to fuse features at different levels; the spatial group enhancement attention mechanism is used to perform spatial group enhancement on the fused features; the convolution output layer is used to perform mask prediction, contour extraction and distance estimation on the features after spatial group enhancement; The residual neural network uses the following formula to extract features at different levels from the input image: x l =x l-1 +F(x l-1 ,W l ); Among them, x l is the output feature of the l-th residual block, and x l-1 is the output feature of the (l - 1)-th residual block. W l is the weight of the l-th residual block. F(x l-1 , W l ) is the overlap of multiple residual layers, where 0 < l ≤ L and L is the number of layers of the residual block. When l = 0, the output feature of the 0-th residual block is the image input to the residual neural network; The process of fusing features at different levels by the multi-scale feature integration block includes: For the lth layer, the deep feature map of the l+1th layer is upsampled to obtain the fused feature map of the l+1th layer; when l=L, the fused feature map of the L+1th layer is the output feature of the residual block of the Lth layer; The fused feature map of the l+1th layer is merged with the output feature of the residual block of the lth layer to obtain the merged feature map of the lth layer; The merged feature maps of the lth layer are fused at multiple scales to obtain the fused feature maps of multiple scales of the lth layer; The fused feature maps of multiple scales in the lth layer are summed element by element and activated to obtain the deep feature map of the lth layer; The following formula is used to perform multi-scale fusion on the feature maps after merging the lth layer: in, is the fusion feature map of three scales, In' is the feature map after merging at the lth layer, Conv 1×1 Represents a 1×1 convolution operation, Conv 3×3 represents a 3×3 convolution operation, and BN represents a batch normalization operation.

2. The land parcel boundary extraction method according to claim 1, characterized in that: The training process of the farmland plot boundary extraction model includes: Acquire remote sensing images of a preset area within a plurality of preset time periods to obtain a plurality of preliminary remote sensing images; Perform image fusion on multiple preliminary remote sensing images to obtain a fused image; Crop the fused image to obtain multiple cropped images; Drawing farmland plot boundary vector data according to the plurality of cropped images, and performing rasterization processing on the farmland plot boundary vector data to obtain a rasterized image of each cropped image; the rasterized image is a mask corresponding to the cropped image; Using overlapping sliding windows to sample multiple cropped images and the rasterized images corresponding to each cropped image, multiple sample remote sensing images and the mask of each sample remote sensing image are obtained; According to multiple sample remote sensing images, the contour map and distance map of each sample remote sensing image are determined using the Euclidean distance transformation formula and the quasi-Euclidean distance transformation formula; Performing geometric transformation on multiple sample remote sensing images, contour maps and distance maps of each sample remote sensing image to obtain a training data set; The training data set is used to train the cultivated land parcel boundary extraction model to obtain a trained cultivated land parcel boundary extraction model.

3. The method for extracting land parcel boundaries according to claim 2, characterized in that: Acquire remote sensing images of a preset area within multiple preset time periods to obtain multiple preliminary remote sensing images, including: Acquire resource satellite remote sensing detection information of a preset area within a plurality of preset time periods; the resource satellite remote sensing detection information includes geographic information, airspace meteorological environment information and remote sensing images; Each remote sensing image is subjected to denoising, cloud removal, geometric correction and orthorectification processing in turn to obtain multiple pre-processed images; According to the geographic information, airspace meteorological environment information and pre-processed images corresponding to each remote sensing image, multiple preliminary remote sensing images are determined.

4. The method for extracting land parcel boundaries according to claim 1, characterized in that: The land parcel boundary extraction method further comprises: Vectorizing the boundary image of the plot of land in the study area to obtain vector data of the study area; Performing superposition analysis on the vector data of the study area and the vector data of non-cultivated land, removing non-cultivated land features from the vector data of the study area, and obtaining cultivated land vector data; The boundary breaks in the cultivated land vector data are repaired to obtain the cultivated land boundary of the study area.

5. A land parcel boundary extraction device, applied to the land parcel boundary extraction method according to any one of claims 1 to 4, characterized in that: The land block boundary extraction device comprises: Image acquisition module, used to obtain satellite remote sensing images of the study area; A boundary extraction module is used to extract the boundary of the plot from the satellite remote sensing image based on the cultivated land plot boundary extraction model to determine the plot boundary image of the study area; the plot boundary image includes a mask, a contour and a distance; The farmland plot boundary extraction model is obtained by pre-training with a training data set, wherein the training data set includes a plurality of sample remote sensing images and masks, contour maps and distance maps corresponding to each sample remote sensing image; the farmland plot boundary extraction model includes a residual neural network, a feature enhancement module and a convolution output layer connected in sequence; the feature enhancement module includes a multi-scale feature integration block and a spatial group enhancement attention mechanism; Among them, the residual neural network is used to extract features at different levels of the input image; the multi-scale feature integration block is used to fuse features at different levels; the spatial group enhancement attention mechanism is used to perform spatial group enhancement on the fused features; the convolution output layer is used to perform mask prediction, contour extraction and distance estimation on the features after spatial group enhancement.

6. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the land boundary extraction method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for extracting land parcel boundaries described in any one of claims 1 to 4 is implemented.

Images