A method and system for extracting farmland boundaries with high precision

Abstract

The present invention provides a method and system for extracting farmland boundaries with high precision, including: obtaining remote sensing images of multiple bands of the area to be extracted; inputting the remote sensing images of each band into a cyclic residual convolutional neural network to obtain multiple rough farmland boundary maps; after fusing all the rough farmland boundary maps, calculating the boundary strength in the fused map to obtain a primary farmland boundary map; using a watershed transformation algorithm to close the gaps in the boundary in the primary farmland boundary map to obtain the final farmland boundary. The method realizes the high-precision delineation of farmland, greatly saving manpower and material resources.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a method and system for extracting farmland boundaries with high precision. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In agricultural monitoring, accurate and efficient acquisition of the scope and location of farmland helps meet social needs such as refined land management and agricultural industrial structure adjustment, thereby promoting the development of agricultural modernization and supporting agricultural monitoring to become regionalized, refined and quantitative.

[0004] Nowadays, most of the research on land parcel classification and extraction is based on high-resolution images, focusing on key technologies such as edge detection algorithms and regional classification and recognition at the object, pixel, and sub-pixel scales. Compared with traditional field surveys, methods based on satellite images can greatly reduce costs and improve efficiency. How to accurately and efficiently extract farmland plots is a long-term challenge facing the field of agricultural remote sensing.

[0005] At present, the extraction of plots is mainly achieved through two methods: manual vectorization and high-resolution image segmentation. The accuracy of extracting boundaries through manual vectorization is high, but this method takes a lot of time and is inefficient when obtaining large-scale farmland boundary data. The use of high-resolution image segmentation, such as threshold-based segmentation methods, edge-based segmentation methods, and region-based segmentation methods, although time-saving and labor-saving, the optimal segmentation scale is affected by the type of land object, the contrast of the surrounding environment, and the internal heterogeneity. The effect is poor when extracting farmland boundaries in complex farming areas, the accuracy is low, and there is a problem of mismatch with geographic entities. Summary of the invention

[0006] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a method and system for extracting farmland boundaries with high precision, which realizes the high-precision delineation of farmland and greatly saves manpower and material resources.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A first aspect of the present invention provides a method for extracting farmland boundaries with high precision, comprising:

[0009] Acquire multiple bands of remote sensing images of the area to be extracted;

[0010] The remote sensing image of each band is input into the cyclic residual convolutional neural network to obtain multiple rough farmland boundary maps;

[0011] After fusing all rough farmland boundary maps, the boundary strength in the fused map is calculated to obtain a primary farmland boundary map;

[0012] The watershed transformation algorithm is used to close the gaps in the boundaries of the primary farmland boundary map to obtain the final farmland boundary.

[0013] Furthermore, the cyclic residual convolutional neural network includes an encoder; the encoder is used to downsample the remote sensing images of each band, extract the features of the remote sensing images, and generate feature maps of different scales.

[0014] Further, the encoder includes a plurality of sequentially connected encoding layers;

[0015] Each encoding layer consists of a convolutional layer, a batch normalization layer, and a linear rectification layer connected in sequence.

[0016] Furthermore, the cyclic residual convolutional neural network also includes a decoder;

[0017] The decoder performs feature fusion on the feature maps of different scales generated by the encoder to restore a rough farmland boundary map.

[0018] Further, the decoder comprises a plurality of decoding layers connected in sequence;

[0019] Each decoding layer consists of a deconvolution layer, a batch normalization layer, and a linear rectification layer connected in sequence.

[0020] Furthermore, the specific steps of the watershed transformation algorithm are:

[0021] Find the downstream pixel for each pixel of the input image and record it in an array;

[0022] For each pixel in the array, determine whether it is a local minimum. If so, assign a new label and assign the new label to the connected and locally minimum pixels until all local minimum pixels are labeled. If not, determine whether its downstream pixels have labels. If so, use the label of its downstream pixels as its own label. Otherwise, determine whether the downstream pixels of its downstream pixels have labels until a labeled pixel is found, and use the label of the labeled pixel as its own label.

[0023] Furthermore, the fusion of all rough farmland boundary maps specifically adopts a single-band fusion method.

[0024] A second aspect of the present invention provides a high-precision farmland boundary extraction system, comprising:

[0025] An image acquisition module is configured to: acquire remote sensing images of multiple bands of the area to be extracted;

[0026] The segmentation module is configured to: input the remote sensing image of each band into a cyclic residual convolutional neural network to obtain multiple rough farmland boundary maps;

[0027] A primary farmland boundary map acquisition module is configured to: after fusing all rough farmland boundary maps, calculate the boundary strength in the fused map to obtain a primary farmland boundary map;

[0028] The gap closing module is configured to: use a watershed transformation algorithm to close the gaps in the boundaries of the primary farmland boundary map to obtain a final farmland boundary.

[0029] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for high-precision extraction of farmland boundaries as described above.

[0030] The fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for high-precision extraction of farmland boundaries as described above are implemented.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention provides a method for extracting farmland boundaries with high precision. The method uses high-resolution satellite images as a data source, and obtains a farmland plot boundary delineation model by improving a U-Net model with a cyclic (recursive) convolutional layer and a residual unit. The method realizes high-precision delineation of farmland, greatly saves manpower and material resources, and can improve the situation in which land and resources surveys heavily rely on manually digitized plot boundaries.

[0033] The present invention provides a method for extracting farmland boundaries with high precision. The method uses a residual convolution module to replace the traditional forward convolution layer, continuously extracts image features based on different time steps, realizes feature accumulation, and helps to develop more effective deep models. It ensures better and stronger feature representation. The introduction of residual units also effectively avoids the network degradation problem caused by deepening the network depth.

[0034] The present invention provides a high-precision method for extracting farmland boundaries, which can determine the scope and location of farmland, provide real-time information on crop conditions for the development of an accurate agricultural monitoring system, and achieve accurate delineation of small plots in complex farming areas by building an algorithm model that adapts to the characteristics of farming areas, providing a reference for fine classification of large-scale plots, meeting social needs such as fine land management and dynamic monitoring of agricultural industrial structure adjustment, and promoting the modernization of agriculture. It has great universality in farmland boundary extraction, can serve smart agriculture, agricultural management, crop planting management, ecological services, etc., and has a wide range of applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0036] Figure 1 is a structural diagram of an improved R2U-Net model according to the first embodiment of the present invention;

[0037] Figure 2 This is a graph showing the plot boundary extraction results in four study areas using U-Net, ResU-Net, and R2U-Net according to the first embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0040] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0041] Embodiment 1

[0042] This embodiment provides a method for extracting farmland boundaries with high precision. Figure 1 As shown, the following steps are included:

[0043] Step 1, obtaining remote sensing images of multiple bands of the area to be extracted;

[0044] Step 2: Input the remote sensing image of each band into the trained cyclic residual convolutional neural network (improved R2U-Net model) to obtain multiple rough farmland boundary maps of the same area;

[0045] Step 3: After fusing all rough farmland boundary maps, calculate the boundary strength in the fused map, and obtain the primary farmland boundary map based on the boundary strength.

[0046] Step 4: Use a watershed transformation algorithm to close the gaps in the boundaries of the primary farmland boundary map to obtain the final farmland boundary.

[0047] The specific construction, training and evaluation steps of the cyclic residual convolutional neural network are as follows:

[0048] Step P1: Through visual interpretation, the training samples of the study area are obtained by vectorization to obtain the training data set of the target area.

[0049] Collect remote sensing images of the study area, obtain farmland boundary sample data of the study area through visual interpretation, and use manual vectorization to obtain the training data set of the target area, including the following specific steps:

[0050] Step P101: The study area in the present invention is Heilongjiang, and four study areas were selected in Heilongjiang, namely: Yangming County, Mudanjiang City (YM, study area 1), Huachuan County, Jiamusi City (HC, study area 2), Kedong County, Qiqihar City (KD, study area 3) and Wangkui County, Suihua City (WK, study area 4). Among them, YM is located in a mountainous area with fragmented cultivated land, and the other three counties are mainly distributed in plain areas. There are small farms in HC and WK areas, and a variety of crops such as corn, cotton and soybeans are planted in summer, while the farmlands in KD area are generally large and neatly shaped.

[0051] Using remote sensing images of 9 bands with resolutions of 10m and 20m provided by Sentinel-2, a total of 400 labeled subset images were generated in the four study areas. Each subset was 256×256 pixels, of which 350 were used for training and the remaining 50 were used for model testing.

[0052] Step P102: Perform preprocessing operations such as radiation correction, atmospheric correction, and geometric registration on the acquired 400 labeled subset images.

[0053] Step P103: Use ArcGIS to vectorize the farmland plots in the pre-processed remote sensing image to obtain manually digitized field boundary vector data.

[0054] Step P104: Mask extraction is performed on the farmland vector boundary data obtained after digitization to obtain the corresponding farmland plot sample data. The remote sensing image of the study area and the farmland plot boundary map are combined into a training sample set.

[0055] Step P2: Construct a farmland boundary delineation model based on the recurrent residual convolutional neural network (improved R2U-Net model).

[0056] Starting from the plot scale of crop remote sensing identification, the cyclic residual U-Net structure (cyclic residual convolutional neural network) originated from R2U-Net is adopted. R2U-Net was originally developed by taking advantage of the advantages of three recently developed CNNs, namely U-Net model, deep residual model and cyclic CNN (RCN) to segment medical images. In the present invention, by studying the U-Net model with a basic forward convolution layer, the U-Net model with a forward convolution layer and a residual unit, namely the residual U-Net (ResU-Net) and the U-Net model with a cyclic (recursive) convolution layer and a residual unit, namely R2U-Net, a farmland boundary delineation algorithm model adapted to the characteristics of agricultural planting areas is constructed, and the precise extraction of fragmented plots in complex agricultural planting areas is achieved, providing a reference for the refined classification of large-scale plots.

[0057] The cyclic residual convolutional neural network includes an encoder, a decoder and a predictor, wherein the encoder is used to downsample the input remote sensing image of each band, extract the features of the remote sensing image, and generate feature maps of different scales. The encoder includes multiple encoding layers connected in sequence, each of which includes a residual convolution module (convolution layer), batch normalization (BN layer) and ReLU nonlinear activation function (linear rectification layer) connected in sequence. The remote sensing image is input into the model, and the target image is feature extracted through the cyclic convolution layer (RCLs) with residual units, thereby cyclically exploring deep features. In order to accelerate the technology of deep network training, the output result is standardized, and this process can also alleviate the problem of internal covariate shift. In the convolutional neural network, the convolution operation is a linear operation. In order to improve the nonlinear expression ability of the network, the activation function is introduced to help the network learn complex abstract features in the data. Therefore, the feature map after batch normalization is input into the linear rectification layer, in which the ReLU function widely used in the convolutional neural network is used as the activation function, and the feature images of different scales after fusion are obtained.

[0058] The decoder performs feature fusion on a large number of feature maps of different scales with lower dimensions generated by the encoder to restore the segmented image (rough farmland boundary map). The decoder includes multiple decoding layers connected in sequence, each of which includes a residual cyclic deconvolution block (deconvolution layer), batch normalization (BN layer) and ReLU nonlinear activation function (linear rectification layer) connected in sequence. The feature map output by the encoding layer is input into the deconvolution module of the decoding layer to enlarge the size of the feature map and restore the size of the feature map. In order to better realize the fusion of feature maps, auxiliary supervision data is introduced before the fusion of the restored size feature map. The remote sensing image segmentation model will pay more attention to the edge part of the object, making the remote sensing image segmentation model more accurate in predicting the edge. The results of the previous step are batch normalized and the activation function is used to realize the fusion of features of different scales. The supervised edge feature map is fused with the main feature map to restore the segmented image.

[0059] The predictor is used to supervise the segmentation results. The predictor contains a convolution layer with a convolution kernel size of 3*3. The predictor maps the final feature map to the same dimension as the number of remote sensing image categories and uses the softmax loss function (cross entropy loss function) to supervise the segmentation results.

[0060] Among them, the encoder encodes the input into an intermediate state, where the input is a vector, and then the decoder network decodes the intermediate state into an output form. Standardization refers to Layer Normalization, because the change of the output of one layer will produce a highly correlated change in the input of the next layer, especially when using ReLU, its output changes greatly. Therefore, standardization is used, which is a technology to accelerate the training of deep networks. By normalizing the mean and variance, the problem of internal covariate shift is alleviated, and data is normalized for all neurons in a layer of the network. The first input of the circular convolution is the image map, and the output is the feature map. When entering the circular convolution, the input is the output feature map to continue feature extraction, and then the deep-level features are discovered in this cycle. Linear rectifier layer (Rectified Linear Units layer, ReLU layer), the activation function (Activation function) of this layer of neurons uses linear rectifier (Rectified Linear Units, ReLU) f(x) = max(0, x) f(x) = max(0, x).

[0061] The construction process of the cyclic residual convolutional neural network includes the following specific steps:

[0062] Step P201: The R2U-Net model used in the present invention is improved on the basis of the U-Net model with a cyclic (recursive) convolutional layer and a residual unit, retaining the convolutional encoding and decoding units of the original U-Net model. Compared with the basic forward convolutional layer in U-Net, a cyclic convolutional layer is used in R2U-Net, which helps to develop a more efficient and deeper model.

[0063] Step P202: Only concatenation operations are used in the R2U-Net model, and the cropping and copying units used in the convolutional network (U-Net) model are removed.

[0064] Step P203: Compared with U-Net, the present invention performs standardization in the R2U-Net model. Standardization is a technique that can accelerate deep network training and alleviate the problem of internal covariate shift by normalizing the mean and variance.

[0065] The specific operation process of standardization is as follows:

[0066] Find the mean of each training batch of data, the formula is:

[0067]

[0068] Find the variance of each training batch of data, the formula is:

[0069]

[0070] Use the obtained mean and variance to normalize the training data of this batch to obtain a 0-1 distribution. The formula is:

[0071]

[0072] Perform scale transformation and offset, the formula is:

[0073]

[0074] Among them, B is a batch of input data, μ B and are the mean and variance of B, ε is a small positive number used to avoid division by 0, γ and β are scale factors and translation factors, respectively, which are trainable parameters.

[0075] Standardization can greatly improve the speed of model training. The network without standardization needs to slowly adjust the learning rate. When standardization is added to the network, a large initial learning rate can be used, and then the learning rate decay rate is also very large. Therefore, the algorithm converges quickly and can significantly reduce the number of iterations that tend to converge, improving the final performance. After data standardization, it is equivalent to using only the linear part of the S-type activation function, which can alleviate the problem of gradient disappearance in the back propagation of the S-type activation function.

[0076] Step P204: Compared with the U-Net model with forward convolutional layers and residual units, a recurrent convolutional layer is added to the model, and an efficient feature accumulation method with different time steps is added to the added recurrent convolutional layer to continuously extract image features. The time step t is 1, 2, and 3 respectively. The input image first passes through two recurrent convolutional layers with a convolution kernel of 3×3. Figure 1 It represents the expansion process of the circular convolution operation at the time step of t=3. At t=0, the image is input to the circular convolution layer. At t=1, the first 3×3 forward convolution operation is performed on the image. At t=2, the input image and the output of the first forward convolution operation are used as the input of the second forward convolution operation. And so on. The output at t=3 is the output of the circular convolution layer.

[0077] The low-level feature information obtained after the circular convolution module operation is fused with the input image to achieve residual connectivity. The feature accumulation based on the circular convolution layer at different time steps ensures better and stronger feature representation, which helps to extract very low-level features. These features provide more detailed features for the remote sensing image segmentation task with complex features and improve the accuracy of segmenting farmland.

[0078] Step P205: Since the convolutional layer with increased stride outperforms the max pooling in multiple image recognition benchmarks, the max pooling layer is replaced with a 3×3 convolutional layer with a stride of 2 based on the R2U-Net model.

[0079] Step P3: Input training samples, learn representative and discriminative features based on the model, and train the farmland boundary delineation model.

[0080] Based on the training data set obtained after manual vectorization, the training samples are input into the constructed cyclic residual convolutional neural network model to learn representative and discriminative features, optimize various parameters, and train the farmland boundary delineation model, including the following specific steps:

[0081] Step P301: oversample the samples containing farmland in the training sample set, and select the same number of samples containing only background as the oversampled samples for undersampling.

[0082] Step P302: Input the oversampled and undersampled training sample sets into the cyclic residual convolutional neural network model, and use a step-by-step training method to continuously adjust and optimize the model parameters. Use the stochastic gradient descent algorithm to obtain a trained cyclic residual convolutional neural network model.

[0083] Step P303: Apply the trained cyclic residual convolutional neural network model to process the Sentinel-2 data of the study area to obtain a preliminary non-closed boundary map.

[0084] Step P4: Perform boundary connection and field generation.

[0085] The contour detection result based on R2U-Net is a fragmented contour, rather than dividing the image into closed parts. To solve this problem, it is necessary to use single-band fusion and edge tracking algorithms to integrate the fragmented boundaries, including the following specific steps:

[0086] Step P401: Each band of the Sentinel-2 multispectral image is ingested into the trained R2U-Net model once to obtain 9 rough farmland boundary layers of the same area.

[0087] Step P402: In the present invention, since the three red edges are very similar to the two SWIR bands, the rough farmland boundary layers obtained by different bands (9 bands) are superimposed to obtain a fusion image (the boundary image after band superposition is called a fusion image), and the frequency of each boundary of the rough farmland boundary layer appearing in the fusion image (i.e., boundary intensity) is calculated and normalized to 0-1.

[0088] Step P403: Perform hierarchical fusion according to a pre-set threshold. The optimal threshold is automatically investigated through the working receiver characteristic (ROC) curve. ROC is a graph showing the degree of contrast between true predictions and false predictions, which describes the relative trade-off between the true positive rate (TPR, plotted on the Y axis) and the false positive rate (FPR, plotted on the X axis).

[0089] In the present invention, the TPR and FPR of the fusion map were calculated in four study areas, and the ROC space was generated by setting the threshold range of 0.1-1.0. In the ROC space, each threshold creates a different point, and the best cutoff point is where the TPR is high and the FPR is low. The ROC curve shows the best cutoff points and different thresholds for different study areas. The results show that the best thresholds for the four study areas are 0.9, 0.8, 0.7 and 0.8, respectively.

[0090] Step P404: Select the boundaries in the normalized fusion image whose boundary strength is greater than the set threshold to obtain a primary farmland boundary map.

[0091] Step P405: Using a watershed transformation algorithm, the boundaries in the primary farmland boundary map are closed to obtain a final farmland boundary.

[0092] The watershed transform (OWT) algorithm is used to process the unclosed boundaries in the primary farmland boundary map. The watershed algorithm has a good response to weak edges and can obtain closed and continuous edges. However, noise in the image and slight grayscale changes on the surface of the object will cause over-segmentation. In order to reduce the over-segmentation caused by the watershed algorithm, the gradient function is usually modified. A simple method is to perform threshold processing on the gradient image to eliminate over-segmentation caused by slight changes in grayscale. The threshold is obtained by first applying a Gaussian filter (using a moving window of 9×9 pixels) to the aggregated edge image (obtained based on the Canny operator), and then calculating the weighted average of each pixel as the intermediate pixel value of the moving window to obtain a generalized edge image. Finally, the threshold is obtained from the standard deviation of the generalized image.

[0093] The final farmland boundary is obtained by iteratively merging the closed boundaries obtained based on the watershed transform (OWT) and combining the primary farmland boundary map in the segmentation map.

[0094] The watershed segmentation algorithm is a mathematical morphology segmentation method based on topological theory. Its basic idea is to regard the image as a topological landform in geodesy. The gray value of each pixel in the image represents the altitude of the point. Each local minimum and its affected area are called a watershed, and the boundary of the watershed forms a watershed. The specific operation steps are as follows (precipitation watershed transform):

[0095] (1) For each pixel, find the downstream pixel (i.e., the neighboring pixel with the smallest grayscale compared to it) and record it in an array. The array records the downstream pixels of all pixels.

[0096] (2) Identify the local minimum. For each pixel in the array, determine whether it is a local minimum (i.e., the pixel with the smallest grayscale value in the local image). If so, assign a new label. The same label is assigned to the connected areas that are all local minima until all the local minimum pixels are labeled.

[0097] (3) Identify the non-local minimum pixel. For each non-local minimum pixel p, there is always a downstream pixel. If the downstream pixel has been identified, assign this label to p. Otherwise, find the downstream pixel of the downstream pixel until the identified downstream pixel is found and assign the label to p. The labels of all pixels constitute the primary farmland boundary map. That is, if the pixel is not the local minimum, determine whether its downstream pixel has a label. If so, use the label of its downstream pixel as its own label. Otherwise, determine whether the downstream pixel of its downstream pixel has a label until a pixel with a label is found, and use the label of the found pixel with a label as its own label.

[0098] Step P406: Since the watershed segmentation algorithm is prone to over-segmentation and often produces a large number of small objects, some form of region merging needs to be performed after the watershed segmentation operation.

[0099] Step P5: Assess the accuracy of the generated farmland boundary. The accuracy of the final generated farmland boundary is evaluated using region-based and edge-based methods to analyze the applicability of the farmland boundary delineation model. The accuracy of the farmland boundary delineation model based on the cyclic residual convolutional neural network is analyzed by comparison with the conventional edge detection method, including the following specific steps:

[0100] Step P501: The suitability of the final farmland boundary is evaluated using a region-based approach, where the region-based metric is calculated using a confusion matrix consisting of 10,000 pixels, of which 5,000 "boundary" pixels are randomly selected within 10 m from the reference boundary, and 5,000 "non-boundary" pixels are randomly selected within a range of more than 10 m from the reference boundary.

[0101] The accuracy indicators include overall accuracy (OA), Kappa coefficient (K), commission error (CE), omission error (OE), and precision recall (PR). CE represents the false boundaries of the field, which is usually caused by over-segmentation of the field. OE provides an accuracy indicator for extracting boundaries along the reference field boundary, where a high error indicates that the segmentation does not describe the field boundary well. PR is a framework for calculating precision (P), recall (R), and F1 score (or F measure). For the farmland boundary map, the PR measure is calculated based on four terms: true positive, true negative, false positive, and false negative. Among them, P is the number of true positive predicted pixels divided by the number of all positive pixels returned by the network, which is used to measure the closeness of the detected boundary to the reference boundary; R is the number of true positive predicted pixels divided by all pixels that should be identified as positive, indicating the proportion of correctly detected reference boundaries; F1 score (F) is the harmonic mean of P and R, ranging from 0 to 1, and tends to the best state when it is close to 1. Its calculation formula is:

[0102]

[0103] Step P502: Use an edge-based method to evaluate the accuracy of the final farmland boundary. The edge-based accuracy is evaluated by four indicators (over-segmentation rate, under-segmentation rate, position offset and eccentricity), which respectively reflect the shape, size and offset of the extracted area relative to the target or reference area.

[0104] First, the overlapping area between the reference field and the delineated field is extracted, and then four indicators are calculated: (1) the relative area of ​​the overlapping area and the reference field (over-segmentation rate, RAor); (2) the relative area of ​​the overlapping area and the extracted field (under-segmentation rate, RAos); (3) the position difference between the extracted field (s) and the reference field (position offset, Dsr), calculated as the average distance between the center point of the extracted field and the center point of the reference field; (4) the absolute difference in shape (eccentricity coefficient, ε);

[0105] The calculation formula is as follows:

[0106]

[0107]

[0108]

[0109] ε=||Eccentricity or -Eccentricity os ||;

[0110] Where n represents the number of fields generated by the R2U-Net model, A o (i) is the area of ​​the i-th overlapping region relative to the reference field, A r is the area of ​​the reference field, As(i) is the area of ​​the ith extracted field, D sr is the average distance between the center point of the extracted field and the center point of the reference field, representing the position accuracy of the delineated field; X s (i) and Y s (i) is the coordinate of the center point of the i-th extraction field, X r and Y r are the coordinates of the center point of the reference field; RA or and Ra os It is used to evaluate the topological (partial) similarity between the extracted region and the reference region, providing a ratio value from 0 to 100: the closer the value is to 100, the less likely it is to be over-segmented or under-segmented. or , R.A. os and D srThe average value of represents the overall segmentation quality; the eccentricity indicates the degree to which the shape of a region deviates from a circle (eccentricity = 0). Table 1 shows the area-based and edge-based metric results of U-Net, ResU-Net and R2U-Net in the four study areas. Figure 2 The plot boundary extraction results of U-Net, ResU-Net and R2U-Net in the four study areas are shown.

[0111] Table 1 Area-based and edge-based measurement results

[0112]

[0113]

[0114] Step P503: Compare with conventional edge detection methods to detect the accuracy of the farmland boundary delineation model. In the present invention, a representative object-based image analysis (OBIA) method is used as a reference to evaluate the classification performance of the proposed method.

[0115] One of the widely used image segmentation methods in OBIA is the multi-resolution fractal network evolution method (FNEA) embedded in commercial definition professional software. FNEA is a bottom-up region growing algorithm, usually defined using spectral variance and object geometry. It starts with a single pixel object and then merges adjacent pairs of objects into a larger object with minimal heterogeneity until the heterogeneity exceeds a user-defined threshold (such as a scale parameter). The image segmentation in the embodiment of the present invention uses Yikang (version 9).

[0116] The segmentation algorithm requires several user-defined parameters: (1) weights associated with the input image layers; (2) a parameter defining the ratio of spectral and spatial heterogeneity contributions; (3) a parameter defining the ratio of compactness and smoothness contributions of segmented objects; and (4) a scale parameter defining the heterogeneity threshold used to merge objects.

[0117] Among them, the scale parameter affects the spatial scale of the image analysis by controlling the size of the synthetic object. At present, there is no automatic and objective method to determine the optimal scale parameter, which is usually determined by a trial and error method of visually interpreting the input image. In the embodiment of the present invention, the shape parameter is determined to be 0.3, the compactness parameter is 0.5, and the scale is 70 through trial and error.

[0118] Embodiment 2

[0119] This embodiment provides a high-precision farmland boundary extraction system, which specifically includes the following modules:

[0120] An image acquisition module is configured to: acquire remote sensing images of multiple bands of the area to be extracted;

[0121] The segmentation module is configured to: input the remote sensing image of each band into a cyclic residual convolutional neural network to obtain multiple rough farmland boundary maps;

[0122] A primary farmland boundary map acquisition module is configured to: after fusing all rough farmland boundary maps, calculate the boundary strength in the fused map to obtain a primary farmland boundary map;

[0123] The gap closing module is configured to: use a watershed transformation algorithm to close the gaps in the boundaries of the primary farmland boundary map to obtain a final farmland boundary.

[0124] It should be noted here that the various modules in this embodiment correspond one-to-one to the various steps in Example 1, and the specific implementation process is the same, which will not be repeated here.

[0125] Embodiment 3

[0126] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the method for high-precision extraction of farmland boundaries as described in the first embodiment above are implemented.

[0127] Embodiment 4

[0128] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for high-precision extraction of farmland boundaries as described in the first embodiment above are implemented.

[0129] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0130] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0131] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0132] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0133] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0134] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A high-precision method for extracting farmland boundaries, characterized in that: include: Acquire multiple bands of remote sensing images of the area to be extracted; The remote sensing image of each band is input into a cyclic residual convolutional neural network to obtain multiple rough farmland boundary maps; the construction process of the cyclic residual convolutional neural network includes the following specific steps: The R2U-Net model is improved on the basis of the U-Net model with a circular convolution layer and a residual unit, retaining the convolution encoding and decoding units of the original U-Net model, and using a circular convolution layer in the R2U-Net model; The R2U-Net model only uses concatenation operations and removes the crop and copy units used in the U-Net model; In the R2U-Net model, standardization is performed by normalizing the mean and variance; After fusing all rough farmland boundary maps, the boundary strength in the fused map is calculated to obtain a primary farmland boundary map; The watershed transformation algorithm is used to close the gaps in the boundaries of the primary farmland boundary map to obtain the final farmland boundary.

2. A high-precision farmland boundary extraction method as claimed in claim 1, characterized in that: The cyclic residual convolutional neural network includes an encoder; The encoder is used to perform down-sampling operations on the remote sensing images of each band, extract features of the remote sensing images, and generate feature maps of different scales.

3. A high-precision farmland boundary extraction method as claimed in claim 2, characterized in that: The encoder comprises a plurality of sequentially connected encoding layers; Each encoding layer consists of a convolutional layer, a batch normalization layer, and a linear rectification layer connected in sequence.

4. A high-precision farmland boundary extraction method as claimed in claim 2, characterized in that: The cyclic residual convolutional neural network also includes a decoder; The decoder performs feature fusion on the feature maps of different scales generated by the encoder to restore a rough farmland boundary map.

5. A method for extracting farmland boundaries with high precision as claimed in claim 4, characterized in that: The decoder comprises a plurality of decoding layers connected in sequence; Each decoding layer consists of a deconvolution layer, a batch normalization layer, and a linear rectification layer connected in sequence.

6. A high-precision farmland boundary extraction method as claimed in claim 1, characterized in that: The specific steps of the watershed transformation algorithm are: Find the downstream pixel for each pixel of the input image and record it in an array; For each pixel in the array, determine whether it is a local minimum. If so, assign a new label and assign the new label to the connected and locally minimum pixels until all local minimum pixels are labeled. If not, determine whether its downstream pixels have labels. If so, use the label of its downstream pixels as its own label. Otherwise, determine whether the downstream pixels of its downstream pixels have labels until a labeled pixel is found, and use the label of the labeled pixel as its own label.

7. A method for extracting farmland boundaries with high precision as claimed in claim 1, characterized in that: The fusion of all rough farmland boundary maps specifically adopts a single-band fusion method.

8. A high-precision farmland boundary extraction system, characterized in that: include: An image acquisition module is configured to: acquire remote sensing images of multiple bands of the area to be extracted; The segmentation module is configured to: input the remote sensing image of each band into a cyclic residual convolutional neural network to obtain multiple rough farmland boundary maps; the construction process of the cyclic residual convolutional neural network includes the following specific steps: The R2U-Net model is improved on the basis of the U-Net model with a circular convolution layer and a residual unit, retaining the convolution encoding and decoding units of the original U-Net model, and using a circular convolution layer in the R2U-Net model; The R2U-Net model only uses concatenation operations and removes the crop and copy units used in the U-Net model; In the R2U-Net model, standardization is performed by normalizing the mean and variance; A primary farmland boundary map acquisition module is configured to: after fusing all rough farmland boundary maps, calculate the boundary strength in the fused map to obtain a primary farmland boundary map; The gap closing module is configured to: use a watershed transformation algorithm to close the gaps in the boundaries of the primary farmland boundary map to obtain a final farmland boundary.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a method for high-precision extraction of farmland boundaries as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method for high-precision extraction of farmland boundaries as described in any one of claims 1-7 are implemented.

Images