Aiming at the complex types of plain crops, a method and system for accurately extracting plot boundaries

Abstract

The application provides a farmland boundary extraction method and system for a complex plain crop type region, and relates to the field of image processing.The method comprises the following steps: obtaining a remote sensing image;extracting time characteristics and spectral characteristics of the remote sensing image, fusing the time characteristics and the spectral characteristics, and constructing an optimal feature dataset;inputting the constructed feature dataset into a pre-trained recurrent residual convolutional neural network to obtain a plurality of rough farmland segmentation result maps;converting the plurality of rough farmland segmentation result maps into a vector boundary map, optimizing the edge contour by using an image processing algorithm, and obtaining a primary farmland boundary map;developing a rule-based decision fusion strategy, classifying crop plots in a region to be extracted by using a random forest algorithm, and obtaining a classification result map of the random forest algorithm;merging the primary farmland boundary map and the classification result map of the random forest algorithm to obtain a final farmland boundary.The application realizes high-precision demarcation of farmland in a complex plain crop type region, and greatly saves manpower and resources.

Description

Technical Field

[0001] This invention belongs to the field of farmland landscape image processing, and particularly relates to a method and system for accurate extraction of plot boundaries in plain areas with complex crop types. Background Technology

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

[0003] Accurately identifying the location and spatial distribution of farmland boundaries is fundamental and core to the digitalization and informatization of modern agriculture. As one of the main agricultural functional scenarios, accurately acquiring farmland information in plain areas is of significant scientific research value not only for crop classification, yield estimation, growth evaluation, and pest and disease assessment, but also for national land surveys, arable land area censuses, and other land resource surveys, as well as agricultural insurance loss assessments.

[0004] Currently, most research on land parcel classification and extraction by scholars both domestically and internationally is based on high-resolution imagery, focusing on object, pixel, and sub-pixel scales, and revolving around key technologies such as histogram thresholding, region-based methods, edge detection-based methods, fuzzy set-based methods, artificial neural network-based methods, and physical methods. Compared with traditional field surveys, satellite imagery-based methods can significantly reduce costs and improve efficiency. How to accurately and efficiently extract farmland parcels remains a long-standing challenge in the field of agricultural remote sensing.

[0005] Traditional methods for delineating farmland boundaries typically involve visual interpretation of remote sensing images or field surveys, which are both time-consuming and labor-intensive. With the availability of free satellite data and increased computing power, there is a growing demand for automatic or semi-automatic methods for delineating land parcel boundaries, such as edge detection, regional segmentation for farmland classification, and machine learning. While these methods are more time- and labor-saving, the optimal segmentation scale is influenced by land cover type, surrounding environmental contrast, and internal heterogeneity. They tend to perform poorly in extracting farmland boundaries in plains areas with complex crop types, exhibiting low accuracy and a mismatch with geographic entities. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, this invention provides a method and system for accurately extracting the boundaries of plots in areas with complex crop types in plains. This achieves high-precision delineation of farmland in areas with complex crop types in plains, greatly saving manpower and material resources and solving the problems in the existing technology.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0008] The first aspect of this invention provides a method for accurately extracting the boundaries of plots in plain areas with complex crop types.

[0009] A method for accurately extracting plot boundaries in plains areas with complex crop types includes the following steps:

[0010] Acquire remote sensing images of the region to be extracted in multiple bands;

[0011] Extract temporal and spectral features from remote sensing images of multiple bands, fuse the temporal and spectral features, and construct the optimal feature dataset.

[0012] The constructed feature dataset is input into a pre-trained recurrent residual convolutional neural network to obtain multiple coarse farmland segmentation result maps;

[0013] Multiple rough farmland segmentation results are converted into vector boundary maps, and image processing algorithms are used to optimize the edge contours to obtain primary farmland boundary maps.

[0014] A rule-based decision fusion strategy was developed, and the random forest algorithm was used to classify the crop plots in the area to be extracted, resulting in a classification map of the random forest algorithm. The primary farmland boundary map was then merged with the classification map of the random forest algorithm to obtain the final farmland boundary.

[0015] Preferably, the temporal and spectral features of multiple band remote sensing images are extracted, including:

[0016] A smoothing filtering algorithm is used to smooth and reconstruct time series data to obtain smoothed and reconstructed time series data.

[0017] The vector-based description method is used to extract the temporal features of the smoothed reconstructed time series data, and obtain nine temporal features: peak value, maximum growth rate, maximum decay rate and their corresponding maximum growth rate stage, peak stage, maximum decay rate stage, peak width, previous peak value and subsequent peak value.

[0018] Differential vegetation index, greenness index, red-green ratio index, ratio vegetation index, soil-adjusted vegetation index, triangular vegetation index, normalized vegetation index, and enhanced vegetation index were selected as initial spectral features. The initial spectral features were selected for importance using a feature importance assessment model based on the Gini coefficient, and the spectral features after importance selection were obtained.

[0019] By fusing temporal features and spectral features selected based on importance, an optimal feature dataset is constructed.

[0020] Preferably, before inputting the constructed feature dataset into the pre-trained recurrent residual convolutional neural network, the method further includes constructing a recurrent residual convolutional neural network containing an encoder, decoder, and predictor:

[0021] The encoder is used to downsample the remote sensing image of each band, extract the features of the remote sensing image, and generate feature maps of different scales. It includes multiple sequentially connected coding layers, each of which includes a sequentially connected convolutional layer, a batch normalization layer, and a linear rectifier layer.

[0022] The decoder fuses the feature maps of different scales generated by the encoder to reconstruct a rough farmland boundary map, which includes multiple sequentially connected decoding layers. Each decoding layer includes a sequentially connected deconvolution layer, a batch normalization processing layer, and a linear rectifier layer.

[0023] The predictor contains convolutional layers with a kernel size of 3*3 and uses a cross-entropy loss function to supervise the segmentation results.

[0024] Preferably, training the recurrent residual convolutional neural network includes:

[0025] Visually interpret remote sensing images and use manual vectorization to obtain farmland boundary sample data of the study area. Combine remote sensing images with farmland plot boundary maps to form a training sample set.

[0026] Oversample the samples in the training sample set that contain farmland, and undersample the samples containing only background that are the same number as the oversampled samples.

[0027] The oversampled and undersampled training sample sets are input into the constructed recurrent residual convolutional neural network model. The model parameters are continuously adjusted and optimized by using a step-by-step training method and the stochastic gradient descent algorithm to obtain the trained recurrent residual convolutional neural network model.

[0028] Preferably, morphological filtering and the Douglas-Puk algorithm are used to optimize the edge contour of the vector boundary map, generating smooth and refined contour lines.

[0029] The preferred rule-based decision fusion strategy comprises the following steps:

[0030] Assign an attribute value to a single land parcel based on the output of a recurrent residual convolutional neural network;

[0031] Attributes will be assigned to individual land parcels classified using the random forest algorithm.

[0032] The plotting results output by the recurrent residual convolutional neural network model are taken as a single block. The number of crop categories within a single block is counted and set as C. It is then determined whether the value is 1.

[0033] If C=1, the final result of land parcel delineation is based on the output of the recurrent residual convolutional neural network; if C>1, the land parcel delineation result based on the recurrent residual convolutional neural network is incorrect, and the classification result is used as the final boundary.

[0034] Preferably, it also includes: evaluating the accuracy of farmland plot extraction results using region-based and edge-based methods, and analyzing the applicability of plot extraction models in areas with complex crop types in plains.

[0035] The second aspect of this invention provides a system for accurately extracting the boundaries of plots in areas with complex crop types in plains.

[0036] A system for accurately extracting plot boundaries in plains areas with complex crop types includes:

[0037] The image acquisition module is configured to acquire remote sensing images of the region to be extracted in multiple bands;

[0038] The feature construction module is configured to: extract temporal and spectral features from remote sensing images of multiple bands, fuse the temporal and spectral features, and construct the optimal feature dataset.

[0039] The coarse segmentation module is configured to input the constructed feature dataset into a pre-trained recurrent residual convolutional neural network to obtain multiple coarse farmland segmentation result maps;

[0040] The edge optimization module is configured to convert multiple coarse farmland segmentation result maps into vector boundary maps, and use image processing algorithms to optimize the edge contours to obtain a primary farmland boundary map.

[0041] The decision fusion module is configured to: develop a rule-based decision fusion strategy, use the random forest algorithm to classify crop plots in the area to be extracted, and obtain the classification result map of the random forest algorithm; merge the primary farmland boundary map with the classification result map of the random forest algorithm to obtain the final farmland boundary.

[0042] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the method for accurate extraction of plot boundaries in complex plain crop types as described in the first aspect of the present invention.

[0043] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the method for accurate extraction of plot boundaries in complex plain crop types as described in the first aspect of the present invention.

[0044] The above one or more technical solutions have the following beneficial effects:

[0045] This invention provides a method for accurately extracting the boundaries of plots in areas with complex crop types in plains. Using Sentinel-2 satellite imagery as the data source, it constructs an optimal feature dataset. By improving the R2U-Net model with recurrent convolutional layers and residual units and developing a rule-based decision fusion strategy, combined with the classification results of the random forest algorithm, a plot boundary delineation model for areas with complex crop types in plains is obtained. This method effectively solves the problem of undersegmentation in plot segmentation in areas with complex crop types in plains, achieving high-precision delineation of farmland, greatly saving manpower and material resources, and providing technological support for the rapid implementation of large-scale crop area monitoring.

[0046] This invention provides a method for accurately extracting plot boundaries in complex areas with diverse crop types in plains. It uses a residual convolution module instead of a traditional forward convolution layer to continuously extract image features based on different time steps, thereby accumulating features. This helps to develop more effective deep models and ensures better and stronger feature representation. The introduction of residual units also effectively avoids network degradation caused by increasing network depth.

[0047] This invention provides a method for accurately extracting the boundaries of plots in areas with complex crop types in plains. It can determine the boundaries of plots in such areas, providing real-time information on crop conditions for developing an accurate agricultural monitoring system. By constructing an algorithm model adapted to the characteristics of agricultural planting areas, it achieves the precise delineation of fragmented plots within complex agricultural planting areas, providing a reference for the refined classification of large-scale plots. This meets the social needs of refined land management and dynamic monitoring of agricultural industrial structure adjustments, promoting the modernization of agriculture. Furthermore, it has great universality in extracting boundaries in areas with complex crop types in plains, and can serve smart agriculture, agricultural management, crop planting management, and ecosystem services, with a wide range of applications.

[0048] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0049] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0050] Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention;

[0051] Figure 2 This is a flowchart of the decision rule-based coupling strategy according to Embodiment 1 of the present invention;

[0052] Figure 3(a) is an original image of the boundary of a complex crop type area in a plain according to Embodiment 1 of the present invention.

[0053] Figure 3(b) is a map showing the results of plot boundary extraction in a plain crop type complex area according to Embodiment 1 of the present invention. Detailed Implementation

[0054] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0055] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0056] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0057] Example 1

[0058] This embodiment discloses a method for accurately extracting the boundaries of plots in plains areas with complex crop types.

[0059] like Figure 1 As shown, the method for accurately extracting plot boundaries in plains areas with complex crop types includes the following steps:

[0060] Acquire remote sensing images of the region to be extracted in multiple bands;

[0061] Extract temporal and spectral features from remote sensing images of multiple bands, fuse the temporal and spectral features, and construct the optimal feature dataset.

[0062] The constructed feature dataset is input into a pre-trained recurrent residual convolutional neural network to obtain multiple coarse farmland segmentation result maps;

[0063] Multiple rough farmland segmentation results are converted into vector boundary maps, and image processing algorithms are used to optimize the edge contours to obtain primary farmland boundary maps.

[0064] A rule-based decision fusion strategy was developed, and the random forest algorithm was used to classify the crop plots in the area to be extracted, resulting in a classification map of the random forest algorithm. The primary farmland boundary map was then merged with the classification map of the random forest algorithm to obtain the final farmland boundary.

[0065] The specific steps for constructing, training, and evaluating a recurrent residual convolutional neural network are as follows:

[0066] Step P1: Through visual interpretation, training samples of the study area are obtained using a vectorization method to obtain the training dataset of the target area.

[0067] Remote sensing images of the study area were collected, and farmland boundary sample data of the study area were obtained through visual interpretation and manual vectorization to obtain the training dataset for the target area. The specific steps included are as follows:

[0068] Step P101: The study area in this invention is Yangming County, Mudanjiang City, Heilongjiang Province. Remote sensing images with resolutions of 10m and 20m in 9 bands provided by Sentinel-2 were used. A total of 500 labeled subset images were generated in the study area, each subset being 256×256 pixels. 450 of these were used for training, and the remaining 50 were used for model testing.

[0069] Step P102: Perform preprocessing operations such as radiometric correction, atmospheric correction, and geometric registration on the acquired 500 labeled subset images.

[0070] Step P103: Use ArcGIS to vectorize the farmland plots in the preprocessed remote sensing image to obtain manually digitized field boundary vector data.

[0071] Step P104: Perform mask extraction on the digitized farmland vector boundary data to obtain corresponding farmland plot sample data. Combine the remote sensing imagery of the study area with the farmland plot boundary map to form a training sample set.

[0072] Step P2: Construct a farmland boundary delineation model based on a recurrent residual convolutional neural network (improved R2U-Net model).

[0073] Starting with the plot-scale approach for crop remote sensing identification, this invention employs a recurrent residual U-Net structure (recurrent residual convolutional neural network) originating from R2U-Net. R2U-Net initially segmented medical images by leveraging the advantages of three recently developed CNNs: the U-Net model, the deep residual model, and the recurrent CNN (RCN). In this invention, by studying the U-Net model with recurrent (recursive) convolutional layers and residual units, i.e., R2U-Net, an algorithm model for accurately extracting plot boundaries in complex plain crop types is constructed. This model achieves accurate extraction of fragmented plot boundaries within complex plain crop types, providing a reference for the refined classification of large-scale plots.

[0074] A recurrent residual convolutional neural network (RNN) comprises an encoder, a decoder, and a predictor. The encoder downsamples the input remote sensing image for each band, extracting features and generating feature maps at different scales. The encoder consists of multiple sequentially connected coding layers, each containing a residual convolutional module (convolutional layer), a batch normalization (BN) layer, and a ReLU nonlinear activation function (rectified linear layer). The remote sensing image is input into the model, and features are extracted from the target image through RCLs with residual units, thus repeatedly uncovering deeper features. To accelerate deep network training, the output is normalized, which also mitigates internal covariate shifting. In convolutional neural networks, convolution is a linear operation; to improve the network's nonlinear expressive power, activation functions are introduced to help the network learn complex abstract features from the data. Therefore, the batch-normalized feature maps are input into a rectified linear layer, where the widely used ReLU function in convolutional neural networks is used as the activation function, resulting in fused feature images at different scales.

[0075] The decoder fuses a large number of low-dimensional feature maps at different scales generated by the encoder to reconstruct the segmented image (a rough farmland boundary map). The decoder consists of multiple sequentially connected decoding layers. Each decoding layer includes a residual recurrent deconvolution block (deconvolution layer), batch normalization (BN layer), and ReLU nonlinear activation function (linear rectified layer). The feature maps output from the encoder layer are input to the deconvolution module of the decoder layer to enlarge the feature map size and restore its dimensions. To better achieve feature map fusion, auxiliary supervised data is introduced before fusing the restored feature maps. This allows the remote sensing image segmentation model to focus more on the edges of objects, making edge prediction more accurate. Batch normalization and activation functions are then applied to the output of the previous step to achieve feature fusion at different scales, fusing the supervised edge feature maps with the main feature maps to reconstruct the segmented image.

[0076] The predictor is used to supervise the segmentation results. The predictor contains convolutional layers with a kernel size of 3*3. This predictor maps the final feature map to a dimension with the same number of categories as the remote sensing image. The segmentation results are supervised using a softmax loss function (cross-entropy loss function).

[0077] The encoder encodes the input into an intermediate state, where the input is a vector, and then the decoder network decodes this intermediate state into an output form. Standardization refers to Layer Normalization, because changes in the output of one layer can significantly alter the input of the next layer, especially when using ReLU, where the output changes drastically. Therefore, standardization is used to accelerate deep network training by normalizing the mean and variance, mitigating the problem of internal covariate shift, and normalizing the data for all neurons in a layer. The recurrent convolution initially takes an image as input and outputs a feature map. Upon entering the recurrent convolution, the input is the output feature map, continuing feature extraction, and this cycle continues to uncover deeper features. The Rectified Linear Units (ReLU) layer uses the ReLU activation function f(x) = max(0,x).

[0078] The construction process of a recurrent residual convolutional neural network includes the following specific steps:

[0079] Step P201: The R2U-Net model used in this invention is an improvement on the U-Net model with recurrent convolutional layers and residual units, while retaining the convolutional encoding and decoding units of the original U-Net model. Compared with the basic forward convolutional layers in U-Net, R2U-Net uses recurrent convolutional layers, which helps to develop more efficient and deeper models.

[0080] Step P202: Only concatenation operations are used in the R2U-Net model, and the pruning and copying units used in the convolutional network (U-Net) model are removed.

[0081] Step P203: Compared to U-Net, this invention performs normalization processing in the R2U-Net model. Normalization processing is a technique that can accelerate the training of deep networks by normalizing the mean and variance, thus alleviating the problem of internal covariate shift.

[0082] The specific operational procedures for standardization are as follows:

[0083] The formula for calculating the mean of each training batch of data is:

[0084]

[0085] The variance of each training batch of data is calculated using the following formula:

[0086]

[0087] Normalize the batch of training data using the obtained mean and variance to obtain a 0-1 distribution, as shown in the formula:

[0088]

[0089] The formulas for scaling and offset are as follows:

[0090]

[0091] Where B represents a batch of input data, μ B and Let B be the mean and variance, ε be a small positive number used to avoid division by zero, γ and β be the scaling and shifting factors, respectively, and β be trainable parameters.

[0092] Standardization can significantly improve model training speed. Without standardization, the learning rate needs to be adjusted gradually. With standardization, a large initial learning rate can be used, followed by a rapid decay of the learning rate, resulting in faster convergence and significantly reducing the number of iterations required to reach convergence, thus improving final performance. Standardization effectively uses only the linear part of the sigmoid activation function, mitigating the gradient vanishing problem during backpropagation.

[0093] Step P204: Compared to the U-Net model with forward convolutional layers and residual units, a recurrent convolutional layer is added to the model. Within this added recurrent convolutional layer, an efficient feature accumulation method with different time steps (t) is employed to continuously extract image features. The time steps t are 1, 2, and 3. The input image first passes through two recurrent convolutional layers with 3×3 kernels. Figure 1 This describes the unfolding process of the circular convolution operation at a time step of t=3. At t=0, the image is input into the circular convolution layer. At t=1, the image is subjected to the first 3×3 forward convolution operation. At t=2, the input image and the output of the first forward convolution operation are used as the input for the second forward convolution operation. This process continues until the output at t=3 is the output of the circular convolution layer.

[0094] The low-level feature information obtained after the operation of the recurrent convolution module is fused with the input image to achieve residual connectivity. The feature accumulation of the recurrent convolution layer based on different time steps ensures better and stronger feature representation, which helps to extract very low-level features. These features provide more detailed features for remote sensing image segmentation tasks with complex features, and improve the accuracy of farmland segmentation.

[0095] Step P205: Since convolutional layers with increased stride are superior to max pooling in multiple image recognition benchmarks, the max pooling layer is replaced with a 3×3 convolutional layer with a stride of 2 in the R2U-Net model.

[0096] Step P3: Construct the feature dataset.

[0097] To eliminate the field boundary blurring caused by identical spectral features, we fused temporal and spectral features and selected the optimal feature dataset.

[0098] Step P301: Optimize time characteristics

[0099] NDVI (Normalized Difference Vegetation Index), also known as the normalized vegetation index or standard difference vegetation index, is one of the most significant parameters reflecting crop condition and nutrient information. In this study, multi-temporal images were used to generate an NDVI time-series dataset to extract temporal features.

[0100] Due to the influence of solar altitude angle, cloud pollution, aerosols, and soil background, outliers in time series data curves should be eliminated, and high-quality time profile curves should be reconstructed. The specific steps for optimizing time features are as follows:

[0101] (1) The Savizky-Golay (SG) smoothing filter algorithm is used to smooth and reconstruct the time series data.

[0102] (2) Use vector-based description methods to extract time features.

[0103] (3) Accurately obtain 9 characteristic parameters, such as peak value, maximum growth rate, maximum decay rate and their corresponding maximum growth rate stage, peak value stage, maximum decay rate stage, peak width, and cumulative peak value before and after.

[0104] The differences in the eigenvectors of time series curves accurately describe the changes in the shape of the curves and fully reflect the growth differences among various crop types.

[0105] Step P302: Optimize spectral characteristics

[0106] Vegetation indices (VIs) based on inter-band mathematical transformations are spectral features widely used in farmland and crop mapping. Considering the definitions and meanings of each vegetation index, eight visible light indices were selected as initial spectral features, and Table 1 shows the eight selected vegetation indices.

[0107] Table 1 List of Spectral Characteristics

[0108]

[0109] Note: B1, B2, B3, and B4 represent the reflectance of the green, red, near-infrared, and blue bands of the Sentinel-2 image.

[0110] When all features are involved in classification, information redundancy inevitably occurs, potentially leading to the "curse of dimensionality" and reduced performance. The Random Forest (RF) algorithm based on the Gini coefficient is used to evaluate the contribution of features to classification and select the optimal spectral features. The principle of the Gini coefficient is to evaluate the contribution of feature variables and normalize the importance scores of all features to compare the weights of feature importance. The optimal feature is then integrated into the deep learning model as input. The feature importance evaluation model based on the Gini coefficient is as follows:

[0111]

[0112] Gini m Let P represent the Gini index of tree node m, and k represent its position in the k-th category. mk It is the probability that tree node m belongs to the kth category.

[0113] The difference in the Gini index before and after the branch is characteristic X. i The importance of node m is called...

[0114]

[0115] Where and represent the Gini indices of new nodes h and t, respectively. The larger the S value, the more significant the feature. If feature X i If a tree appears N times in the j-th tree, then its importance score is... It can be calculated using the following formula:

[0116]

[0117] Assuming there are R trees in the random forest, Xi's importance score in RF is:

[0118]

[0119] Calculations showed that the weights of the eight feature variables varied greatly, with TVI having the highest weight and RGI the lowest. Ultimately, TVI, GI, and NDVI were selected as the three vegetation indices to improve the accuracy of crop identification and plot extraction.

[0120] Step P4: Input training samples, learn representative and discriminative features based on the model, and train the farmland boundary delineation model.

[0121] Based on the training dataset obtained after manual vectorization, the training samples are input into the constructed recurrent 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:

[0122] Step P401: Oversample the samples in the training sample set that contain farmland, and undersample the samples containing only background that are the same number as the oversampled samples.

[0123] Step P402: Input the oversampled and undersampled training sample sets into the recurrent residual convolutional neural network model, and continuously adjust and optimize the model parameters using a step-by-step training method. Utilize the stochastic gradient descent algorithm to obtain the trained recurrent residual convolutional neural network model.

[0124] Step P403: Apply the trained recurrent residual convolutional neural network model to process the feature dataset of the study area to obtain multiple rough farmland segmentation result maps.

[0125] Step P5: Accurately extract and optimize the boundaries of the land parcels.

[0126] The contour detection results based on R2U-Net are fragmented contours with insufficient smoothness and continuity of edge contours. Furthermore, the R2U-Net model is prone to undersegmentation in plains areas with complex crop types, failing to effectively separate different types of plots. Therefore, image processing and classification-aided methods are needed to address these issues following the results of the R2U-Net model, including the following specific steps:

[0127] Step P401: After obtaining the detailed farmland extraction results from R2U-Net, the plot area needs to be converted into vector boundaries, resulting in multiple coarse farmland boundary maps;

[0128] Step P402: Image processing methods are used to optimize edge contours. This study employs morphological filtering and the Douglas-Puk algorithm (DPA) to generate efficient and stable smooth contour lines.

[0129] Morphological filtering, in morphological image processing, typically uses predefined simple structuring elements to probe an image, obtain its morphological information, and then perform targeted processing and correction. Thinning can be viewed as a morphological skeleton extraction; thinning of image X using structuring element B can be expressed as:

[0130]

[0131] X⊙B represents a hit or miss match of B against X. In a complete refinement process, {B} = {B...} 1 B 2,…,B n}, with B i The image is thinned sequentially during the iteration process. As B i+1 The refined object continues until the refinement is complete, that is, with The process repeats in a loop until the result no longer changes or

[0132] The set number of iterations has been reached.

[0133] During the refinement process, a simplified version of the refinement iterative algorithm was used, which mainly uses the distribution of the eight neighbors as the condition for deciding whether to retain the current point. The entire iterative process consists of the following two loops until no foreground points are deleted.

[0134] The specific steps are as follows:

[0135] (1) Loop through the foreground points in the image and delete the current point p1 that meets the following conditions:

[0136] i.2≤N(p1)≤6

[0137] ii. A(P1) = 1

[0138] iii. P2 * P4 * P6 = 0

[0139] iv. P4 * P6 * P8 = 0

[0140] Where N(p1) represents the number of foreground points in the eight neighborhoods of p1, and A(P1) is the number of times the background-foreground swap occurs in clockwise order;

[0141] (2) Loop through the image again and delete the current point p1 that meets the following conditions:

[0142] i.2≤N(p1)≤6

[0143] ii. A(P1) = 1

[0144] iii. P2 * P4 * P8 = 0

[0145] iv. P2 * P6 * P8 = 0

[0146] After edge refinement processing, the number of detected edge patches was significantly reduced.

[0147] DPA (Distributed Persistent Analysis) is used to compress a large number of redundant graphic data points to extract the necessary data points. Its basic idea is to draw a straight line connecting the beginning and end points of each curve, calculate the distance between all points and the line, and find the maximum distance value D. max , using D max Compared to the limit D; if D max<D, all the midpoints on this curve are discarded; if D max ≥D, retain D max The corresponding coordinate points, and taking this point as the boundary, divide the curve into two parts, and repeat using this method for these two parts.

[0148] The specific operation steps are as follows:

[0149] (1) Connect a straight line between the starting point and the ending point of the curve, and draw the distances from the remaining points to the straight line.

[0150] (2) Find the maximum distance from each point to the straight line, and judge whether it is greater than the threshold. If so, retain the point with the maximum distance from the straight line; if not, eliminate all the points between the two endpoints of the straight line.

[0151] (3) According to the retained points, divide the known curve into two parts and perform iterative processing, that is, select the remaining maximum distance and compare it with the threshold, and discard them in turn until there are no points to be excluded. Finally, obtain the coordinates of the curve points that meet the given accuracy limit error, discard other points, and finally obtain a smooth boundary to obtain the primary farmland boundary map.

[0152] Step P403: For the problem of under-segmentation in the complex crop type area of the plain plot, a coupling strategy based on decision rules is applied, as Figure 2 shown below, the specific content is as follows:

[0153] (1) Unique identification of plots based on the segmentation results of the improved R2U-Net. In the present invention, an attribute value 'M' is assigned to a single plot output by the modified R2U-Net, such as M1, M2, M3.

[0154] (2) Assign attribute values to each crop plot obtained by RF classification. The classification result is obtained using the random forest (RF) algorithm. The RF model is constructed using the optimal features, and the number of required classification trees (n tree ) is set to 1000, and the number of predictor variables (m try ) is equal to the square root of the number of input variables. In the present invention, 450 training samples are used to classify 5 crop types (corn, soybean, rice, wheat, and other crops), and 250 validation samples based on the confusion matrix are used to evaluate the accuracy. To improve the discrimination of different crop types, 'n' (n = 1, 2, 3...I, I is the number of crop categories) is used to assign each crop plot. According to the actual situation that wheat, soybean, and corn are mainly planted in the study area, we assign 1 to soybean, 2 to corn, 3 to wheat, and 4 to rice.

[0155] (3) Calculate the number of crop categories within a single plot. Using the plotting results output by the modified R2U-Net model as a single plot, we count the number of classes n and set it as C;

[0156] (4) A rule-based decision fusion strategy was developed. Ideally, the c value of the target class should be 1, where the final result of farmland delineation is the modified R2U-Net; if C>1, it means that the plots incorrectly delineated by the modified R2U-Net model contain more than one crop, and the classification result of the random forest is used as the final boundary.

[0157] Figure 3(a) shows the original image, and Figure 3(b) shows the result of extracting the plot boundaries of the original image using the method for accurate extraction of plot boundaries in complex plain crop types as described in this invention.

[0158] Step P6: Accuracy assessment of the generated farmland plot results. Region-based and edge-based methods are used to evaluate the accuracy of the farmland plot extraction results, thereby analyzing the applicability of the plot extraction model for complex plain crop types. Specifically, the steps are as follows:

[0159] Step P601: The accuracy of the final farmland plot extraction results is evaluated using a region-based method. The region-based metric is calculated using a confusion matrix, which consists of 10,000 samples. 5,000 "boundary" samples are randomly selected along the reference boundary, and the other 5,000 "non-boundary" samples are randomly selected at a distance of more than 10 meters from the reference boundary.

[0160] Accuracy metrics include Overall Accuracy (OA), Kappa coefficient (K), Commission Error (CE), Omission Error (OE), and Precision Recall (PR). CE represents false boundaries of the field, often caused by over-segmentation. OE provides an accuracy metric for extracting boundaries along a reference field boundary, where high error indicates that segmentation does not accurately describe the field boundary. PR is a framework for calculating precision (P), recall (R), and F1 score (or F metric). For farmland boundary maps, the PR metric is calculated based on four terms: true positives, true negatives, false positives, and false negatives. P is the number of true positive predicted pixels divided by the total number of positive pixels returned by the network, measuring how close the detected boundary is to the reference boundary; R is the number of true positive predicted pixels divided by the total number of pixels that should be identified as positive, representing the proportion of correctly detected reference boundaries; the F1 score (F) is the harmonic mean of P and R, ranging from 0 to 1, tending towards 1 for optimal performance, and is calculated as follows:

[0161]

[0162] Step P602: The accuracy of the final farmland plot extraction results is evaluated using an edge-based method. The accuracy of the edge-based method is assessed by four indicators (oversegmentation rate, undersegmentation rate, location offset, and eccentricity rate), which reflect the shape, size, and offset of the extracted area relative to the target or reference area, respectively.

[0163] First, the overlapping area between the reference field and the delineated field is extracted, and then four indices are calculated: (1) the relative area of ​​the overlapping area and the reference field (oversegmentation rate, RAor); (2) the relative area of ​​the overlapping area and the extracted field (undersegmentation rate, RAos); (3) the positional difference between the extracted field (s) and the reference field (positional offset, Dsr), which is 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).

[0164] The calculation formula is as follows:

[0165]

[0166]

[0167]

[0168] ε=||Eccentricity or -Eccentricity os ||;

[0169] 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 related to the reference field, A r D is the area of ​​the reference field, As(i) is the area of ​​the i-th extracted field, and D sr It is the average distance between the center point of the extracted field and the center point of the reference field, representing the positional accuracy of the delineated field; X s (i) and Y s (i) are the coordinates of the center point of the i-th extraction field, X r and Y r These are the coordinates of the center point of the reference field; RA or and RA os 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 oversegmentation and undersegmentation will occur. RA or RA os and D sr The average value represents the overall segmentation quality; the eccentricity indicates the degree to which the shape of a region deviates from a circle (eccentricity = 0). Table 2 shows the measurement results based on regions and edges in the study area.

[0170] Table 2. Measurement results based on area and edge.

[0171]

[0172] Example 2

[0173] This embodiment discloses a system for accurately extracting the boundaries of plots in areas with complex crop types in plains.

[0174] like Figure 2 As shown, the system for accurately extracting plot boundaries in plains areas with complex crop types includes:

[0175] The image acquisition module is configured to acquire remote sensing images of the region to be extracted in multiple bands;

[0176] The feature construction module is configured to: extract temporal and spectral features from remote sensing images of multiple bands, fuse the temporal and spectral features, and construct the optimal feature dataset.

[0177] The coarse segmentation module is configured to input the constructed feature dataset into a pre-trained recurrent residual convolutional neural network to obtain multiple coarse farmland segmentation result maps;

[0178] The edge optimization module is configured to convert multiple coarse farmland segmentation result maps into vector boundary maps, and use image processing algorithms to optimize the edge contours to obtain a primary farmland boundary map.

[0179] The decision fusion module is configured to: develop a rule-based decision fusion strategy, use the random forest algorithm to classify crop plots in the area to be extracted, and obtain the classification result map of the random forest algorithm; merge the primary farmland boundary map with the classification result map of the random forest algorithm to obtain the final farmland boundary.

[0180] Example 3

[0181] The purpose of this embodiment is to provide a computer-readable storage medium.

[0182] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for accurate extraction of plot boundaries in complex plain crop types as described in Embodiment 1 of this disclosure.

[0183] Example 4

[0184] The purpose of this embodiment is to provide an electronic device.

[0185] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the method for accurate extraction of plot boundaries in complex plain crop-type areas as described in Embodiment 1 of this disclosure.

[0186] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0187] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0188] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for accurately extracting plot boundaries in plain areas with complex crop types, characterized in that, Includes the following steps: Acquire remote sensing images of the region to be extracted in multiple bands; Extract temporal and spectral features from remote sensing images of multiple bands, fuse the temporal and spectral features, and construct the optimal feature dataset. The constructed feature dataset is input into a pre-trained recurrent residual convolutional neural network to obtain multiple coarse farmland segmentation result maps; Multiple rough farmland segmentation results are converted into vector boundary maps, and image processing algorithms are used to optimize the edge contours to obtain primary farmland boundary maps. A rule-based decision fusion strategy was developed, and the random forest algorithm was used to classify the crop plots in the area to be extracted, resulting in a classification map of the random forest algorithm. The primary farmland boundary map was then merged with the classification map of the random forest algorithm to obtain the final farmland boundary. Extracting temporal and spectral features from multiple band remote sensing images, including: A smoothing filtering algorithm is used to smooth and reconstruct time series data to obtain smoothed and reconstructed time series data. The vector-based description method is used to extract the temporal features of the smoothed reconstructed time series data, and obtain nine temporal features: peak value, maximum growth rate, maximum decay rate and their corresponding maximum growth rate stage, peak stage, maximum decay rate stage, peak width, previous peak value and subsequent peak value. Differential vegetation index, greenness index, red-green ratio index, ratio vegetation index, soil-adjusted vegetation index, triangular vegetation index, normalized vegetation index, and enhanced vegetation index were selected as initial spectral features. The initial spectral features were selected for importance using a feature importance assessment model based on the Gini coefficient, and the spectral features after importance selection were obtained. By fusing temporal features and spectral features selected based on importance, an optimal feature dataset is constructed. Before inputting the constructed feature dataset into the pre-trained recurrent residual convolutional neural network, the process also includes constructing a recurrent residual convolutional neural network containing an encoder, decoder, and predictor: The encoder is used to downsample the remote sensing image of each band, extract the features of the remote sensing image, and generate feature maps of different scales. It includes multiple sequentially connected coding layers, each of which includes a sequentially connected convolutional layer, a batch normalization layer, and a linear rectifier layer. The decoder fuses the feature maps of different scales generated by the encoder to reconstruct a rough farmland boundary map, which includes multiple sequentially connected decoding layers. Each decoding layer includes a sequentially connected deconvolution layer, a batch normalization processing layer, and a linear rectifier layer. The predictor contains convolutional layers with a kernel size of 3*3 and uses a cross-entropy loss function to supervise the segmentation results.

2. The method for accurate extraction of plot boundaries in plains areas with complex crop types as described in claim 1, characterized in that, Training a recurrent residual convolutional neural network includes: Visual interpretation of remote sensing images was performed, and farmland boundary sample data of the study area were obtained by manual vectorization. The remote sensing images and farmland plot boundary maps were combined to form a training sample set. Oversample the samples in the training sample set that contain farmland, and undersample the samples containing only background that are the same number as the oversampled samples. The oversampled and undersampled training sample sets are input into the constructed recurrent residual convolutional neural network model. The model parameters are continuously adjusted and optimized by using a step-by-step training method and the stochastic gradient descent algorithm to obtain the trained recurrent residual convolutional neural network model.

3. The method for accurate extraction of plot boundaries in plains areas with complex crop types as described in claim 1, characterized in that, Morphological filtering and the Douglas-Puk algorithm are used to optimize the edge contours of the vector boundary map, generating smooth and refined contour lines.

4. The method for accurate extraction of plot boundaries in plains areas with complex crop types as described in claim 1, characterized in that, The rule-based decision fusion strategy involves the following steps: Assign an attribute value to a single land parcel based on the output of a recurrent residual convolutional neural network; Attributes will be assigned to individual land parcels classified using the random forest algorithm. The plotting results output by the recurrent residual convolutional neural network model are taken as a single block. The number of crop categories within a single block is counted and set as C. It is then determined whether the value is 1. If C=1, the final result of the land parcel delineation is the land parcel delineation result based on the output of the recurrent residual convolutional neural network; If C>1, then the land parcel delineation result based on the recurrent residual convolutional neural network is incorrect, and the classification result will be used as the final boundary.

5. The method for accurate extraction of plot boundaries in plain areas with complex crop types as described in claim 1, characterized in that, Also includes: The accuracy of farmland plot extraction results was evaluated using region-based and edge-based methods, and the applicability of the plot extraction model in areas with complex crop types in plains was analyzed.

6. A system for accurately extracting plot boundaries in plains areas with complex crop types, characterized by: include: The image acquisition module is configured to acquire remote sensing images of the region to be extracted in multiple bands; The feature construction module is configured to: extract temporal and spectral features from remote sensing images of multiple bands, fuse the temporal and spectral features, and construct the optimal feature dataset. The coarse segmentation module is configured to input the constructed feature dataset into a pre-trained recurrent residual convolutional neural network to obtain multiple coarse farmland segmentation result maps; The edge optimization module is configured to convert multiple coarse farmland segmentation result maps into vector boundary maps, and use image processing algorithms to optimize the edge contours to obtain a primary farmland boundary map. The decision fusion module is configured to: develop a rule-based decision fusion strategy, use the random forest algorithm to classify crop plots in the area to be extracted, and obtain the classification result map of the random forest algorithm; merge the primary farmland boundary map with the classification result map of the random forest algorithm to obtain the final farmland boundary; Extracting temporal and spectral features from multiple band remote sensing images, including: A smoothing filtering algorithm is used to smooth and reconstruct time series data to obtain smoothed and reconstructed time series data. The vector-based description method is used to extract the temporal features of the smoothed reconstructed time series data, and obtain nine temporal features: peak value, maximum growth rate, maximum decay rate and their corresponding maximum growth rate stage, peak stage, maximum decay rate stage, peak width, previous peak value and subsequent peak value. Differential vegetation index, greenness index, red-green ratio index, ratio vegetation index, soil-adjusted vegetation index, triangular vegetation index, normalized vegetation index, and enhanced vegetation index were selected as initial spectral features. The initial spectral features were selected for importance using a feature importance assessment model based on the Gini coefficient, and the spectral features after importance selection were obtained. By fusing temporal features and spectral features selected based on importance, an optimal feature dataset is constructed. Before inputting the constructed feature dataset into the pre-trained recurrent residual convolutional neural network, the process also includes constructing a recurrent residual convolutional neural network containing an encoder, decoder, and predictor: The encoder is used to downsample the remote sensing image of each band, extract the features of the remote sensing image, and generate feature maps of different scales. It includes multiple sequentially connected coding layers, each of which includes a sequentially connected convolutional layer, a batch normalization layer, and a linear rectifier layer. The decoder fuses the feature maps of different scales generated by the encoder to reconstruct a rough farmland boundary map, which includes multiple sequentially connected decoding layers. Each decoding layer includes a sequentially connected deconvolution layer, a batch normalization processing layer, and a linear rectifier layer. The predictor contains convolutional layers with a kernel size of 3*3 and uses a cross-entropy loss function to supervise the segmentation results.

7. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for accurate extraction of plot boundaries in complex areas with plain crop types as described in any one of claims 1-5.

8. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for accurate extraction of plot boundaries in complex areas with complex crop types in plains as described in any one of claims 1-5.

Images