River network intelligent simplification method under unsupervised variational self-coding

Abstract

The invention provides a river network intelligent simplification method under unsupervised variational self-encoding, which is applied to intelligent simplification of a river network in map integration. Comprising the steps of performing point sampling on target river network data to obtain a river network data sampling point set; rasterizing the river network data sampling point set to obtain a raster image, and performing image processing on the raster image to obtain an enhanced image; and inputting the enhanced image into a preset river network intelligent simplification model to obtain a simplified image, and performing raster data vectorization operation on the simplified image to obtain simplified river network point set data. By applying the river network intelligent simplification method provided by the invention, the overall structure, position, shape and local form of the river network can be effectively maintained.

Description

Technical Field

[0001] The present invention relates to the field of intelligent simplification methods for river networks in map synthesis, and in particular to an intelligent simplification method for river networks under unsupervised variational autoencoding. Background Art

[0002] In map generalization, the commonly used methods for simplifying vector lines are mainly divided into two categories, including traditional mathematical simplification methods and machine learning-based simplification methods. Traditional mathematical simplification methods mainly include node-based simplification and curvature-based simplification. Among them, node-based simplification mainly includes Douglas-Peucker (DP) algorithm, Li-Openshaw, Wang-Muller algorithm, Visvalingam-Whyatt (VW) algorithm, arc-to-chord, vertical-to-chord and their corresponding improved methods. Most of the above simplification effects depend on the determination of algorithm parameters and thresholds, and have high requirements on the experience and subjective understanding of the cartographer, and it is difficult to meet the adaptive simplification requirements of different scenarios.

[0003] Machine learning provides a paradigm for data-driven cartographic generalization. By learning the relationship between the characteristic indicators in the existing simplification cases of cartographers, a comprehensive decision-making model suitable for similar environments is constructed, which promotes the intelligent development of automatic cartographic generalization. The simplification methods based on machine learning include using support vector machines (SVM) to train and classify curve nodes and bends respectively, using case-based reasoning (CBR) ideas for analogical reasoning, designing simplification models based on genetic algorithms (GA), using self-organizing maps (SOM) for shape structure extraction in line simplification, using K-means for unsupervised learning, and integrating learning with graphics and image features for intelligent line simplification. These methods are difficult to learn the deep comprehensive knowledge implicit in the simplified cases, and the degree of intelligence still needs to be improved. Deep learning, as a deeper machine learning method, has the ability to learn complex features and patterns in data, so as to well mine the deep comprehensive knowledge hidden in the cases. In the research on line simplification using deep learning methods, there are many methods, including using deep separable convolutional neural networks to achieve vector line simplification based on regional candidate networks, Pix2Pix learning cases and automatically generating simplification results, applying U-Net to the simplification and synthesis of mountain roads, training deep neural network models based on Stacked autoencoder (SAE) to achieve line feature simplification, and introducing map synthesis knowledge into broken line simplification based on Artificial neutral network (ANN). However, it is difficult for existing deep learning methods to directly obtain vector simplification results or the simplification results are not ideal in line simplification, and the performance may be limited in different environments. Summary of the invention

[0004] In view of the fact that traditional rule-based methods usually have limitations in performance such as dependence on parameters and threshold determination, and the problem that existing machine learning-based methods are difficult to directly obtain vector simplification results or the simplification results are unsatisfactory in online simplification, the present invention provides a river network intelligent simplification method and system under unsupervised variational autoencoding, which utilizes the advantages of the variational autoencoder model's adaptive characteristics and powerful learning representation capabilities to achieve intelligent simplification of the river network.

[0005] On the one hand, the present invention provides a river network intelligent simplification method under unsupervised variational autoencoding, comprising the following steps:

[0006] Step 1: Perform point sampling on the target river network data to obtain a sampling point set of the river network data;

[0007] Step 2: rasterizing the river network data sampling point set to obtain a raster image, and performing image processing on the raster image to obtain an enhanced image;

[0008] Step 3: Input the enhanced image into a preset river network intelligent simplification model to obtain a simplified image, and perform a raster data vectorization operation on the simplified image to obtain simplified river network point set data; wherein the river network intelligent simplification model adopts a VAE model.

[0009] Furthermore, the step 1 specifically includes: classifying the target river network data in turn, extracting the grade and curvature of the benchmark river sections, clustering the benchmark river sections, and sampling the benchmark river sections.

[0010] Furthermore, the river classification of the target river network data specifically includes: classifying the target river network data according to the classification method of tree-structured rivers to obtain the mainstream and tributaries at various levels, and extracting nodes in the mainstream and tributaries at various levels, wherein the nodes include starting points, end points and intersections.

[0011] Furthermore, the extraction of the grade and curvature of the benchmark river section specifically includes: determining the shortest river section length from the mainstream and tributaries of each level, dividing the river sections of each level with the shortest river section length as the benchmark length to obtain a number of benchmark river sections, and calculating the grade and curvature of all benchmark river sections; wherein a river section refers to the river between any two adjacent nodes.

[0012] Furthermore, the clustering of the benchmark river sections specifically includes: according to the upstream and downstream order, taking the grades and curvatures of all benchmark river sections as feature vectors, clustering all benchmark river sections based on the feature vectors using the K-MEANS clustering algorithm to obtain clustering results.

[0013] Furthermore, the benchmark river section sampling specifically includes: based on the clustering result, for each cluster, sampling the benchmark river section in the cluster at a preset sampling distance; wherein, the greater the grade and / or curvature of the benchmark river section, the denser the sampling points.

[0014] Furthermore, in step 2, image processing is performed on the raster image to obtain an enhanced image, which specifically includes: image preprocessing, edge feature extraction and image enhancement.

[0015] Furthermore, in step 3, the encoder in the VAE model includes 4 convolutional layers, the convolution kernel size of each layer is 4, and the step size is 2; the decoder includes 8 deconvolutional layers, the convolution kernel size of each layer is 4, and the step size is 2, and the output of the last layer is processed using the Sigmoid activation function.

[0016] Furthermore, the river network intelligent simplification method under unsupervised variational autoencoding also includes step 4 data evaluation, which evaluates the river network simplification data obtained in step 3, and the evaluation method includes at least one of area change evaluation, Hausdorff distance evaluation, curvature coefficient evaluation and river confluence angle evaluation.

[0017] On the other hand, the present invention provides a river network intelligent simplification system under unsupervised variational autoencoding, comprising:

[0018] The sampling module is used to perform point sampling on the target river network data to obtain a sampling point set of the river network data;

[0019] An image processing module, used for rasterizing the river network data sampling point set to obtain a raster image, and performing image processing on the raster image to obtain an enhanced image;

[0020] The river network simplification module is used to input the enhanced image into a preset river network intelligent simplification model to obtain a simplified image, and perform raster data vectorization operation on the simplified image to obtain simplified river network point set data; wherein the river network intelligent simplification model adopts a VAE model.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] (1) The present invention combines the VAE model with river network data to perform cartographic synthesis, realize data dimensionality reduction and visualization, provide more flexible, diversified and comprehensive options for mapping, and enrich the expression and application methods of river network data.

[0023] (2) The present invention takes into account the topological structure and connectivity of the river network, and performs sampling on the basis of comprehensive consideration of the river network level and degree of curvature, thereby ensuring that the sampling results can more accurately capture the true morphological characteristics of the river.

[0024] (3) The present invention evaluates the overall structure, position retention, shape and local morphology of the simplified river network, and makes judgments from quantitative and qualitative perspectives to ensure that the simplification effect meets the actual application requirements.

[0025] In general, the unsupervised learning method VAE has significant advantages in solving uncertainty problems in geographic information science and generating diversified data. It can overcome the limitations of traditional rule-based methods in performance, such as dependence on parameters and threshold determination, and can also solve the problem that existing machine learning-based methods are difficult to directly obtain vectorized simplification results or the simplification results are not ideal, providing new opportunities and possibilities for the research and application of geographic science. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 A schematic diagram of a simplified method for river network intelligence under unsupervised variational autoencoding provided by an embodiment of the present invention;

[0028] Figure 2 A river network sampling case provided for an embodiment of the present invention;

[0029] Figure 3 An image processing flow chart provided by an embodiment of the present invention;

[0030] Figure 4 The overall architecture diagram of VAE provided by the embodiment of the present invention;

[0031] Figure 5 The VAE-based river network data intelligent simplified model provided by the embodiment of the present invention;

[0032] Figure 6 A schematic diagram of area calculation provided by an embodiment of the present invention;

[0033] Figure 7 A schematic diagram of Hausdorff distance calculation provided by an embodiment of the present invention;

[0034] Figure 8 A schematic diagram of the change of the river confluence angle provided by an embodiment of the present invention;

[0035] Fig. 9 The simplification results of river networks at different scales of the VAE model provided in the embodiment of the present invention;

[0036] Fig.10 A comparison chart of the area changes of the VAE model provided in the embodiment of the present invention and the other three models at different scales;

[0037] Fig.11 A comparison chart of the average values ​​of the area normalized asymmetry index of the VAE model provided by the embodiment of the present invention and the other three models at different scales;

[0038] Fig.12 The VAE model provided in the embodiment of the present invention and the other three models are simplified at different scales to show the changes in the bending coefficients of rivers at different levels of the river network;

[0039] Fig.13 The VAE model provided in the embodiment of the present invention and the other three models are simplified at different scales with respect to the average value of the river confluence angle. DETAILED DESCRIPTION

[0040] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0041] The present invention is further explained below with reference to the accompanying drawings and specific embodiments:

[0042] like Figure 1 As shown, the embodiment of the present invention provides a river network intelligent simplification method under unsupervised variational autoencoding, comprising the following steps:

[0043] S101: performing point sampling on the target river network data to obtain a sampling point set of the river network data;

[0044] S102: rasterizing the river network data sampling point set to obtain a raster image, and performing image processing on the raster image to obtain an enhanced image;

[0045] Specifically, the sampling points of the river network data are rasterized, and the vector points are mapped into the grid system. For each vector point, the corresponding pixel value is set at the corresponding grid position, corresponding to one pixel of the grid, which is calibrated into an image, and the obtained raster image is processed.

[0046] S103: Input the enhanced image into a preset river network intelligent simplified model to obtain a simplified image, and perform a raster data vectorization operation on the simplified image to obtain simplified river network point set data; wherein the river network intelligent simplified model adopts a VAE model.

[0047] Specifically, raster data vectorization is to binarize the generated image by setting a threshold, setting pixels above the threshold to white and those below the threshold to black, extracting black pixel points in the binary image for coordinate transformation, linearly scaling the coordinate points to map them back to the original data range, and obtaining a simplified sampling point data set.

[0048] VAE is a probabilistic model based on variational inference (such as Figure 4 As shown in Figure 1, it is a generative model that combines the ideas of autoencoders and probabilistic graph models. It is essentially composed of two parts: the encoder and the decoder. Figure 4 An example of an architecture is shown. Given a dataset consisting of N continuous or discrete variables x In the example, the potential latent variable is z, and the data generation process is as follows: First, the input sample x is mapped to the latent space through the encoder to obtain the distribution parameters θ of z, that is, the mean μ and standard deviation σ, and calculate the posterior distribution p θ Variational estimate of (z|x) Variational distribution for the multivariate normal distribution The formula is:

[0049]

[0050] in, represents a multivariate normal distribution; Indicates that when given data x, the parameter The mean of the latent variable z calculated by the defined function; is a diagonal matrix;

[0051] Then, sampling and decoding are performed. From a prior distribution p θ (z) to sample a hidden variable z (i) , the decoder is based on z (i) Decode, according to the conditional distribution p θ (x (i) |z) Generate a reconstructed sample x that is consistent with the original feature structure of sample i ′(i) ;

[0052] Finally, the loss between the reconstructed sample and the original sample is calculated and the model is optimized. The encoder in the VAE will try to carry all the feature information x carried by the sample. i The distribution of is converted into a Gaussian distribution d i ,This type of Gaussian distribution is encoded with Gaussian distribution as the target, which is called latent distribution. The encoder needs to output the mean μ of this type of Gaussian distribution. i and standard deviation σ iAs the encoding result, in order to ensure the mean value μ of the encoder output i and standard deviation σ i To satisfy the normal distribution, VAE sets a penalty term KL divergence (Kullback-Leibler divergence) in the loss function. KL divergence is a measure of the difference between two distributions. By minimizing KL divergence, With p θ (z|x) These two distributions are as similar as possible, and the loss function minLoss is defined VAE as follows:

[0053] minLoss VAE =D KL (q(z|x)||P(z)+E q(z|x) [logP(x|z)])

[0054] Among them, D KL (q(z|x)||P(z)) is the KL divergence, E q(z|x) [logP(x|z)] is the reconstruction error. By optimizing the loss function, the encoder and decoder can be trained simultaneously to achieve high-quality data reconstruction and generation;

[0055] In the optimization objective of the model, for the variational distribution With the prior distribution p θ The KL divergence of (z) is calculated as follows:

[0056]

[0057] Where n represents the number of elements of the latent variable z. For variational distribution and the true posterior distribution p θ The KL divergence of (z|x) is calculated as follows:

[0058]

[0059] By transformation, the left side of the equation is converted to The right side of the equation is transformed into

[0060]

[0061] Among them, logp θ (x (i) ) is the log-likelihood of generating the real data; It is the variational lower bound (EvidenceLower Bound, ELBO). For the generative model, it is necessary to maximize the sum of the marginal likelihood probabilities of each independent sample, that is, minimize the KL divergence. Since the KL divergence is non-negative, the following inequality is obtained:

[0062]

[0063] Therefore, ELBO is usually used as the optimization objective of variational inference to maximize the variational lower bound and make the variational distribution As close as possible to the true posterior distribution p θ (z|x).

[0064] The river network intelligent simplification method provided by the embodiment of the present invention combines the VAE model with the river network data to perform cartographic synthesis, realize the dimensionality reduction and visualization of the data, provide more flexible, diversified and comprehensive options for mapping, and enrich the expression and application methods of river network data.

[0065] Based on the above embodiment, an embodiment of the present invention provides a sampling method for a river network data sampling point set, which specifically includes:

[0066] The target river network data is sequentially classified into river classes, the grades and curvatures of the benchmark river sections are extracted, and the benchmark river sections are clustered and sampled.

[0067] Specifically, point sampling of river network data can better preserve the geometric characteristics of the original data, reduce data dimensions and reduce computational complexity.

[0068] The target river network data is classified into river classes, specifically including: classifying the target river network data according to the classification method of tree-structured rivers to obtain the mainstream and tributaries at various levels, and extracting nodes in the mainstream and tributaries at various levels, wherein the nodes include starting points, end points and intersections.

[0069] Specifically, the collected river network data is divided according to the classification method of tree-structured rivers, based on the selection of mainstream tributaries. The longest river in the river network is determined as the mainstream, and then the tributaries are gradually determined from the mainstream, and finally hierarchical classification is performed. The mainstream is the first-level river, and the tributaries are classified in turn. After the river network is pre-processed by topological inspection, the nodes in the mainstream and tributaries of each level are extracted, and the nodes include the starting point, the end point and the intersection point.

[0070] The grade and curvature of the benchmark river sections are extracted from the target river network data, specifically including: determining the shortest river section length from the mainstream and tributaries at all levels, dividing the river sections of each level with the shortest river section length as the benchmark length to obtain several benchmark river sections, and calculating the grade and curvature of all benchmark river sections; wherein a river section refers to the river between any two adjacent nodes.

[0071] Specifically, in order to preserve the morphological and detailed features of the river section itself, the shortest river section length is determined from the mainstream and tributaries at all levels, and the original river sections are divided into small units using the shortest river section length as the benchmark river section length. The divided river sections are called benchmark river sections, and the level of the original river sections is retained. The calculation formula for the curvature of the benchmark river section is: the curve length of the benchmark river section / the straight-line distance between the two ends of the benchmark river section.

[0072] The target river network data is clustered into benchmark river sections, specifically including: according to the upstream and downstream order, the grades and curvatures of all benchmark river sections are used as feature vectors, and all benchmark river sections are clustered based on the feature vectors using the K-MEANS clustering algorithm to obtain clustering results.

[0073] Specifically, the curve grade and curvature of the benchmark river section are associated and passed into the K-MEANS clustering algorithm as feature vectors. This embodiment selects the optimal number of clusters as 7 according to the evaluation results of the silhouette coefficient. From the classification results, the river section with grade 1 is mainly divided into two categories. This embodiment is divided into 1-7 categories according to the constraint that the higher the curve grade, the greater the curvature, and the denser the sampling points.

[0074] Sampling the benchmark river sections of the target river network data specifically includes: based on the clustering results, sampling the benchmark river sections in each cluster according to a preset sampling distance; wherein, the greater the grade and / or curvature of the benchmark river section, the denser the sampling points.

[0075] Specifically, this embodiment performs sampling according to the clustering results, and adopts equal-interval sampling to sample different types of curves. The sampling distance of categories 1-7 is between 20m-50m, and the sampling interval increases by 5m for each level to obtain a river network data point set, which is vector data.

[0076] It is understandable that Figure 2, the above embodiments can maintain the river network topology, and ensure the accuracy of the topology of the entire river network system when performing simplification operations, including the unchanged positions of the starting point, end point and intersection of each river, to ensure that the simplified data can still accurately reflect the distribution characteristics of the entire river network system; it can ensure the relationship between the main and secondary rivers. Since the relationship between the mainstream and the tributaries has an important influence on the simplification of the river, when sampling, the main rivers are densely extracted according to the river level, and the sampling interval for the tributaries with lower levels is gradually increased; it can retain the river bending characteristics. The river sections with higher curvature usually have more details and shape changes in the geographic space. By densely sampling the rivers with higher curvature, important geographic information and key features can be better preserved. A sampling method for a river network data sampling point set provided by an embodiment of the present invention takes into account the topological structure and connectivity of the river network, and performs sampling on the basis of comprehensive consideration of the river network level and curvature, which can ensure that the sampling results more accurately capture the true morphological characteristics of the river.

[0077] Based on the above embodiments, an embodiment of the present invention provides a method for processing raster images of river network data, which specifically includes: performing image processing on the raster image to obtain an enhanced image, which specifically includes: image preprocessing, edge feature extraction and image enhancement in sequence.

[0078] Specifically, the image processing process mainly includes three parts: image preprocessing, feature extraction and image enhancement. Figure 3 . Image preprocessing operations include enlarging the image to increase the resolution and detail information of the image; binarization operations convert the image into black and white, and the original data is black points, which is convenient for subsequent processing; then randomly select some black pixels and turn them into white pixels to simulate random inactivation to increase the robustness and diversity of the image. Next, edge feature extraction is performed, and the Canny edge detection algorithm is used to detect the edge of the image, and morphological processing is performed, including dilation and corrosion of the boundary, so that the boundary is clearer and more refined. Finally, in order to enhance the diversity of the image, a symmetric transformation is performed on it to further improve the generalization ability of the model.

[0079] Based on the above embodiments, an embodiment of the present invention provides an intelligent simplification model of a river network. The intelligent simplification model of the river network adopts a VAE model. The encoder in the VAE model includes 4 convolutional layers, the convolution kernel size of each layer is 4, and the step size is 2; the decoder includes 8 deconvolutional layers, the convolution kernel size of each layer is 4, and the step size is 2, and the Sigmoid activation function is used to process the output of the last layer.

[0080] Specifically, the variational autoencoder neural network model for river network simplification mainly consists of the following parts: input data, encoder, latent space, decoder and output data. Figure 5As shown in the figure, the vector data is rasterized and the processed image is input into the VAE model. The encoder extracts features from the image through the convolution layer, which specifically includes 4 convolution layers, with a convolution kernel size of 4 and a stride of 2. Then, the output of the encoder is mapped to the latent space using a fully connected layer. The model contains two fully connected layers, which respectively generate the mean μ and variance σ in the latent space. During the training process, random vectors are obtained by sampling from a normal distribution, and the latent space variables are generated by combining μ and σ. Finally, the sampled samples are obtained from the latent space, and the sampled samples are decoded by the decoder through the deconvolution layer to map the latent variables back to the dimensions and distribution of the original data. The decoder in the model includes 8 deconvolution layers, with a convolution kernel size of 4 and a stride of 2. The last layer uses the Sigmoid activation function to limit the output to the range of [0,1], thereby improving the matching degree and accuracy of the reconstructed image with the input image and obtaining a simplified image.

[0081] Based on the above embodiments, an embodiment of the present invention provides a river network simplification data evaluation method, which evaluates the obtained river network simplification data, and the evaluation method includes at least one of area change evaluation, Hausdorff distance evaluation, curvature coefficient evaluation and river confluence angle evaluation.

[0082] Specifically, according to existing literature, the length of the simplified line can be used as a metric for estimating a generalized scale. After the image is generated using the intelligent simplified model of the river network of the present invention, it is converted into coordinate points. Since the pixelated features do not have good linear features when converted into vector points, the points that are close to the original line elements are taken as the final generated points, and they are connected into line segments in the upstream and downstream order of the original river to finally form a river network. This embodiment determines the parameter size of the closest distance based on the target simplified scale length. When the point is 150mm away from the original data, the generated length meets the scale of 1:25K; when the point is 300mm away from the original data, the generated length meets the scale of 1:50K. This embodiment adjusts the parameters to obtain simplified river network data with scales of 1:25K and 1:50K, and evaluates the simplified river network data, including area change evaluation, Hausdorff distance evaluation, bending coefficient evaluation, and river intersection angle evaluation, to evaluate the simplification results. The global display after simplification in this embodiment is as follows Fig. 9 As shown, Figures 9(a) and 9(b) are the simplified results at two different scales of 1:25K and 1:50K respectively.

[0083] Specifically, the area change assessment refers to the size of the closed area enclosed by the original curve and the simplified curve, which can provide an estimate of the information loss between the two curves. The size of the area can indicate whether the details and curvature information of the original curve are preserved during the simplification process. If the area is smaller, it means that the simplified curve is close to the original curve, retaining more details and shape features, and the information loss is smaller. Conversely, the greater the difference, the greater the difference.

[0084] The river network in the study area of ​​this embodiment is an open curve, so the flow direction of the river network is used as the reference. Figure 6 As shown, the left area change S_L and the right area change S_R are calculated. The balance between the left and right area changes can represent the area preservation ability of the method. The normalized asymmetry index of its evaluation index is defined as Asym _ index=(S_L-S_R) / (S_L+S_R). The index range is (-1,1). When the value is close to 0, it means that the area changes on the left and right sides are relatively balanced, and the simplified curve changes almost equally on both sides. When the value is a positive number, it means that S_L is greater than S_R, indicating that the simplified curve has a larger deviation on the left side, and vice versa. The area change evaluation method provides a local error measure for the data set, reveals the changes in the curve in different parts, and reflects the global similarity of the entire curve.

[0085] Specifically, Hausdorff distance is a geometric metric in the matching algorithm that can be used to compare the overall morphological deviation of the river before and after simplification. It measures the similarity between two sets and represents the distance between the two most dissimilar points in the two sets. Figure 7 As shown, there are two point sets A{a0,a1,…} and B{b0,b1,…}. Take any point in A and calculate its distance to any point in B. Take the longest distance h among all the shortest distances {h0,h1,…} from A to B. n As the Hausdorff distance. The Hausdorff distance between two river network datasets reflects the maximum difference or similarity between the datasets. When the Hausdorff distance is small, the similarity between the river networks is high. Conversely, it indicates that there are large differences in shape, structure, and distribution between the two datasets.

[0086] Specifically, the curvature coefficient refers to the ratio of the actual length of a river section to the straight length of the river section, and the calculation formula is Ka = L / R, where L represents the actual length of the river and R represents the straight length of the river section. The larger the curvature coefficient of a river section, the more bends and details of the river network are preserved, reflecting the complex morphology of the natural river network.

[0087] Specifically, the river intersection angle refers to the angle of the intersection in the river network. Its change can reflect the branching and interweaving degree of the river network and evaluate the geometric changes of the river. Assume that a river network has n angles, with the intersection as the center, and the angle range is between 0° and 180° or 0° and -180°, respectively N1, N2, ... N n , the number of angles remains unchanged after simplification, and the angles become M1, M2, …M n ,like Figure 8 As shown, the formula for calculating the average value of the changing river confluence angle is: The smaller the change in the river confluence angle, the smaller the change in the geometric shape at the river node, and the higher the similarity of the rivers. Conversely, the larger the change in the confluence angle, the lower the similarity of the rivers.

[0088] Specifically, in order to intuitively reflect the effect of the VAE model on river network simplification, this embodiment uses the above four evaluation methods to conduct a comprehensive comparison with the DP algorithm, WM algorithm and VW algorithm (refer to Figures 10 to 13 ).

[0089] ① Area change: The curve formed by connecting the simplified data point set and the original curve form a closed figure, and calculate the change in the total area. Fig.10 As shown in the figure, at the scale of 1:25K and 1:50k, the area change after simplification by the method of the present invention is close to that after simplification by the WM method, but the area change of the method of the present invention is smaller, indicating that the river network simplified by the method of the present invention is closer to the original river network and the information loss is smaller. However, the DP algorithm and the VW algorithm pay more attention to the removal and retention of the turning points, and the area change of the simplified river network is larger than that of the original river network, and more details and shape features are lost;

[0090] In the simplified results at different scales, the local area changes were calculated for a single line segment, and the average values ​​of the normalized asymmetry index of the four methods for rivers of each grade were calculated, e.g. Fig.11 As shown. The higher the level of the river, the smaller the overall asymmetry index is, and the lower the level of the river, the higher the asymmetry index is. This means that the higher the level of the river, the smaller the morphological changes are, and the better the simplification effect is. The river network intelligent simplification method proposed in the present invention obtains an asymmetry index result that is closer to 0 in rivers with higher levels, and the results at different scales are relatively stable, which means that the method of the present invention has stability and consistency in dealing with the area asymmetry index, and better maintains the balance of shrinkage changes in the river network. However, the DP algorithm has a higher asymmetry index and obvious fluctuations in the two scales, the WM algorithm has a higher area asymmetry value in high-level rivers, and the VW algorithm has a larger asymmetry index at the 1:50K scale. This may be due to the partial elimination of certain characteristic bends, resulting in a large deviation from the center line of the river network, and a large loss of details and shape features.

[0091] ②Hausdorff distance: In order to make a quantitative comparison of shape similarity, the Hausdorff distance is used to measure the similarity between the two river networks, and the evaluation test index is shown in Table 1. Due to the change in scale, more details will be omitted. The Hausdorff distance at a large scale (1:25K) is lower than that at a small scale (1:50K), that is, the similarity is higher. The method of the present invention presents a smaller Hausdorff distance at different scales, indicating that the river network intelligent simplification method of the present invention has high effectiveness and stability in retaining the shape characteristics of the river network and maintains a high degree of similarity.

[0092] Table 1 Hausdorff distances between the simplified river networks and the original river networks using the four methods

[0093]

[0094] ③ Bending coefficient: The bending coefficient is the main reference factor during sampling, so the changes in bending coefficients of different levels in the 1:25k and 1:50k scales of the four methods are calculated, such as Fig.12 As shown in the figure. In the 1:25K scale map, the curvature coefficient of the DP algorithm and the WM algorithm suddenly increases at Level 5, indicating that more curvatures of low-level rivers are retained. In the two-scale simplification results, the overall curvature coefficient of the method of the present invention and the VW method shows a downward trend, but the curvature coefficient of the method of the present invention is generally higher than that of the simplified result of VW, indicating that the simplified result of the present invention retains more curvatures and details at the same scale, and better maintains the complex morphology of the natural river network.

[0095] ④ River intersection angle: Calculate the change in the intersection angle of the river network after simplification, such as Fig.13 In the river networks of the two scales, the change of intersection angle after simplification by this method is significantly smaller than that of the simplification results of other methods, indicating that the simplification results using the method of the present invention can better maintain the structure and angle of the river confluence, and pay more attention to preserving the overall structure and relative position of the river.

[0096] The simplified river network was evaluated from the overall and local perspectives. In terms of the overall structure and position maintenance of the river network, the overall area change, Hausdorff distance and river confluence angle were used for evaluation, and it was found that: ① In the overall area change, the result obtained by the river network intelligent simplification method of the present invention is similar to the simplification result of WM, and the overall area change is significantly smaller than the other two methods. ② In the 1:25K scale, the Hausdorff distance of the VW method is the smallest, followed by the river network intelligent simplification method of the present invention, and in the 1:50K scale, the Hausdorff distance obtained by the river network intelligent simplification method of the present invention is the smallest. Overall, the river network intelligent simplification method of the present invention performs better in the Hausdorff distance evaluation method. ③ In the change of the river confluence angle, the angle change of the river network intelligent simplification method of the present invention at different scales is the smallest.

[0097] In terms of maintaining the local morphology of the river network, the area normalization index and river curvature coefficient of rivers of different levels at different scales were calculated, and it was found that: ① The river network intelligent simplification method of the present invention has a smaller change in the normalized area asymmetry index in rivers with higher levels, and the area asymmetry index is the smallest among all methods, indicating that the more important the river, the more balanced the left and right area contraction changes are. ② In the river curvature coefficient, the overall curvature coefficient of the river network intelligent simplification method of the present invention is higher than that of other methods, and a higher curvature coefficient is maintained on rivers with higher levels, indicating that the curvature details of important rivers are better maintained.

[0098] Based on the above conclusions, the river network simplification data evaluation method provided in the embodiment of the present invention evaluates the overall structure, position retention, shape and local morphology of the simplified river network, and makes judgments from quantitative and qualitative perspectives to ensure that the simplification effect meets the actual application requirements.

[0099] As described in the above embodiments, the river network simplified by the river network intelligent simplification method of the present invention has advantages in overall structure, position retention, shape and local morphology. These advantages make the present invention have a wide range of application potentials in the field of river network simplification, and provide new ideas and methods for understanding and analyzing complex river network systems.

[0100] Overall, the unsupervised learning method VAE has significant advantages in solving uncertainty problems in geographic information science and generating diversified data. It can overcome the limitations of traditional methods in the case of insufficient labels and incomplete data, and provides new opportunities and possibilities for the research and application of geographic science.

[0101] It should be noted that the above embodiments are intended to deepen the understanding of the present invention and are only used to explain the present invention, and do not constitute a limitation on the scope of protection of the present invention. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. The river network intelligent simplification method based on unsupervised variational autoencoding is characterized by: The following steps are involved: Step 1: Perform point sampling on the target river network data to obtain a sampling point set of the river network data; Step 2: rasterizing the river network data sampling point set to obtain a raster image, and performing image processing on the raster image to obtain an enhanced image; Step 3: Input the enhanced image into a preset river network intelligent simplification model to obtain a simplified image, and perform a raster data vectorization operation on the simplified image to obtain simplified river network point set data; wherein the river network intelligent simplification model adopts a VAE model.

2. The river network intelligent simplification method based on unsupervised variational autoencoding according to claim 1 is characterized in that: The step 1 specifically includes: classifying the target river network data in turn, extracting the grade and curvature of the benchmark river sections, clustering the benchmark river sections, and sampling the benchmark river sections.

3. The river network intelligent simplification method under unsupervised variational autoencoding according to claim 2 is characterized in that: The river classification of the target river network data specifically includes: classifying the target river network data according to the classification method of tree-structured rivers to obtain the mainstream and tributaries at various levels, and extracting nodes in the mainstream and tributaries at various levels, wherein the nodes include starting points, end points and intersections.

4. The method for intelligent simplification of river networks under unsupervised variational autoencoding according to claim 3 is characterized in that: The method of extracting the grade and curvature of the benchmark river section specifically includes: determining the shortest river section length from the mainstream and tributaries of various levels, dividing the river sections of various levels with the shortest river section length as the benchmark length to obtain a number of benchmark river sections, and calculating the grade and curvature of all benchmark river sections; wherein a river section refers to the river between any two adjacent nodes.

5. The river network intelligent simplification method under unsupervised variational autoencoding according to claim 2 is characterized in that: The clustering of the benchmark river sections specifically includes: according to the upstream and downstream order, taking the grades and curvatures of all benchmark river sections as feature vectors, clustering all benchmark river sections based on the feature vectors using the K-MEANS clustering algorithm to obtain clustering results.

6. The method for intelligent simplification of river networks under unsupervised variational autoencoding according to claim 2, characterized in that: The benchmark river section sampling specifically includes: based on the clustering result, for each cluster, sampling the benchmark river section in the cluster at a preset sampling distance; wherein, the greater the grade and / or curvature of the benchmark river section, the denser the sampling points.

7. The method for intelligent simplification of river networks under unsupervised variational autoencoding according to claim 1, characterized in that: In step 2, image processing is performed on the raster image to obtain an enhanced image, which specifically includes: image preprocessing, edge feature extraction and image enhancement.

8. The river network intelligent simplification method based on unsupervised variational autoencoding according to claim 1 is characterized in that: In step 3, the encoder in the VAE model includes 4 convolutional layers, the convolution kernel size of each layer is 4, and the step size is 2; the decoder includes 8 deconvolutional layers, the convolution kernel size of each layer is 4, and the step size is 2, and the output of the last layer is processed using the Sigmoid activation function.

9. The method for intelligent simplification of river networks under unsupervised variational autoencoding according to claim 1, characterized in that: It also includes step 4 data evaluation, evaluating the river network simplification data obtained in step 3, and the evaluation method includes at least one of area change evaluation, Hausdorff distance evaluation, curvature coefficient evaluation and river intersection angle evaluation.

10. The river network intelligent simplification system based on unsupervised variational autoencoding is characterized by: include: The sampling module is used to perform point sampling on the target river network data to obtain a sampling point set of the river network data; An image processing module, used for rasterizing the river network data sampling point set to obtain a raster image, and performing image processing on the raster image to obtain an enhanced image; The river network simplification module is used to input the enhanced image into a preset river network intelligent simplification model to obtain a simplified image, and perform raster data vectorization operation on the simplified image to obtain simplified river network point set data; wherein the river network intelligent simplification model adopts a VAE model.

Images