Traffic element remote sensing feature extraction method and system and traffic element identification method and system

Abstract

The invention provides a traffic element remote sensing feature extraction method and system and a traffic element recognition method and system.The traffic element remote sensing feature extraction method comprises the following steps that a sample set and a feature set are formed through remote sensing data of an area to be extracted; respectively calculating the sample separability, redundancy and Gini coefficient of each remote sensing feature in the remote sensing feature set based on the sample set and the feature set; calculating a feature importance value of each remote sensing feature based on the sample separability, redundancy and Gini coefficient of each remote sensing feature; the feature importance value is used for evaluating the joint effect of the sample separability, redundancy and Gini coefficient of each remote sensing feature; and selecting the remote sensing features of which the feature importance values exceed a set threshold value as the traffic element remote sensing features of the to-be-extracted region. According to the method, comprehensive feature screening and optimization are carried out on the basis of research of samples, features and algorithms, data redundancy is further solved, and the problems of traffic element remote sensing extraction efficiency, precision optimization and the like are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning and remote sensing image traffic element extraction, and specifically relates to a traffic element remote sensing feature extraction method and system, and a traffic element recognition method and system. Background Art

[0002] Remote sensing extraction and recognition of traffic elements is to extract the characteristic information of traffic elements from remote sensing images and identify them as specific land object categories. It has a wide range of applications in national defense, agriculture and forestry, transportation, and urban planning.

[0003] In recent years, deep learning has been widely used in the research of remote sensing extraction and classification of traffic elements due to its advantage of extracting deep features of remote sensing images. For remote sensing identification of traffic elements, most research focuses on the selection and analysis of samples and models in order to improve the accuracy of traffic element extraction. However, with the increase in the types and number of remote sensing data sources, how to select appropriate remote sensing features of traffic elements to reduce data complexity, reduce the computational cost of deep learning, and improve the accuracy of traffic element extraction results is an urgent problem to be solved.

[0004] In previous studies, the methods of feature optimization can be divided into the perspective of samples, the perspective of features, and some feature optimization algorithms. The main idea of ​​the feature optimization method from the perspective of samples is to analyze the differences of different traffic elements in different features. Commonly used difference evaluation indicators include JM distance (Jeffries-MatusitaDistance, JM) and M index (M Index, MI), etc.; while from the perspective of features, the main consideration for feature optimization is the correlation between features. Correlation indicators include Pearson correlation coefficient (Pearson), Spearman rank correlation coefficient (Spearman) and Kendall rank correlation coefficient (kendall), etc.; in addition, commonly used feature optimization algorithms include genetic algorithm (Genetic Algorithm, GA), particle swarm optimization algorithm (Particle Swarm Optimization, PSO), Gini Index feature importance evaluation (GINI Index, Gini) and generalized normal distribution algorithm (Generalized Normal Distribution Optimization, GNDO), etc.

[0005] However, the feature optimization method of a single angle is often anisotropic, which will lead to inconsistent results in the extraction of traffic elements, resulting in low accuracy in the recognition of the land object category of traffic elements. Summary of the invention

[0006] The purpose of the present invention is to solve the shortcomings of the above-mentioned background technology, and to provide a method and system for extracting remote sensing features of traffic elements, and a method and system for identifying traffic elements, to perform comprehensive feature screening and optimization based on the research of samples, features and algorithms, to further solve the problems of data redundancy, and to improve the efficiency and accuracy optimization of remote sensing extraction of traffic elements.

[0007] The technical solution adopted by the present invention is: a method for extracting remote sensing features of traffic elements, comprising the following steps:

[0008] A sample set and a feature set are formed by using the remote sensing data of the area to be extracted; wherein the sample set includes the coordinate position in the area to be extracted and the corresponding traffic element and ground object category; the feature set includes the coordinate position in the area to be extracted and the corresponding remote sensing feature set;

[0009] Based on the sample set and feature set, the sample separability, redundancy and Gini coefficient of each remote sensing feature in the remote sensing feature set are calculated respectively;

[0010] The feature importance value of each remote sensing feature is calculated based on the sample separability, redundancy and Gini coefficient of each remote sensing feature; the feature importance value is used to evaluate the joint effect of the sample separability, redundancy and Gini coefficient of each remote sensing feature;

[0011] The remote sensing features whose feature importance values ​​exceed the set threshold are selected as the remote sensing features of the traffic elements in the area to be extracted.

[0012] In the above technical solution, the remote sensing features in the remote sensing feature set include: band, spectral index, spectral transformation feature, texture feature, geometric feature, backscattering coefficient, polarization index and elevation, etc.

[0013] The above technical solution also includes: obtaining Sentinel-2 optical remote sensing satellite images, Sentinel-1 radar remote sensing satellite images, and SRTM digital elevation model data of the area to be extracted;

[0014] The nearest neighbor method was used to resample the Sentinel-2 optical remote sensing satellite images, Sentinel-1 radar remote sensing satellite images and SRTM digital elevation model data;

[0015] After the Sentinel-2 optical remote sensing satellite images were subjected to radiometric calibration, atmospheric correction, terrain correction, cloud and snow removal, and mosaic cropping preprocessing, the bands, spectral index, spectral transformation characteristics, texture characteristics, and geometric characteristics were extracted;

[0016] After removing thermal noise, calibrating radiation, correcting terrain and preprocessing the Sentinel-1 radar remote sensing satellite images using the Refine Lee filter algorithm, the backscattering coefficient and polarization index were extracted.

[0017] For the SRTM digital elevation model data, after preprocessing by filling holes, the altitude, slope, aspect and mountain shadow are extracted.

[0018] In the above technical solution, the following formula is used to calculate the sample separability of any remote sensing feature:

[0019] JM=2*(1-e -B );

[0020]

[0021] Among them, JM represents the JM distance, which is used to evaluate the separability of samples; B is the Bhattacharyya distance on the remote sensing feature dimension; i and j represent different types of traffic elements, m i represents the average value of each sample in the traffic element land type i sample set on the remote sensing feature, δ i represents the variance of each sample in the traffic element land type i sample set on the remote sensing feature; m j represents the average value of each sample in the traffic element land class j sample set on the remote sensing feature, δ j Represents the variance of each sample in the traffic element j sample set on this feature.

[0022] In the above technical solution, for any remote sensing feature, the average value of the Pearson coefficients of the remote sensing feature and other remote sensing features is used as the redundancy evaluation value.

[0023] In the above technical solution, the feature importance value Y is calculated using the following formula:

[0024] Y=β1X1+β2X2+β3X3

[0025] Where, X1: [x 11 , x 12 , …, x 1m ], X2: [x 21 , x 22 , …, x 3m ], X3: [x 31 , x 32 , …, x 3m ], m represents the number of remote sensing features, xpq represents the normalized value of the pth optimization factor of the qth remote sensing feature; the optimization factors include sample separability, redundancy and Gini coefficient; p = 1, 2, 3; q∈[1, m]; β1, β2, β3 are regression coefficients, calculated using the information weight method.

[0026] In the above technical solution, the calculation process of setting the threshold value T includes:

[0027] The feature importance value set of remote sensing features is defined as M = {N1, N2, ..., N m};

[0028] T=[Max M -Min M ]×t;

[0029] Where m represents the number of remote sensing features, Max M Refers to the maximum value in the eigenvalue set, Min M Refers to the minimum value in the feature value set, and t is a custom threshold.

[0030] The present invention also provides a traffic element remote sensing feature extraction system, including: a sample set and feature set generation module, an optimization factor calculation module, a comprehensive feature evaluation module and a traffic element remote sensing feature determination module;

[0031] The sample set and feature set generation module is used to form a sample set and a feature set through remote sensing data of the area to be extracted; wherein the sample set includes the coordinate position in the area to be extracted and the corresponding traffic element and ground object category; the feature set includes the coordinate position in the area to be extracted and the corresponding remote sensing feature set;

[0032] The optimization factor calculation module is used to calculate the sample separability, redundancy and Gini coefficient of each remote sensing feature in the remote sensing feature set based on the sample set and the feature set;

[0033] The comprehensive feature evaluation module is used to calculate the feature importance value of each remote sensing feature based on the sample separability, redundancy and Gini coefficient of each remote sensing feature; the feature importance value is used to evaluate the joint effect of the sample separability, redundancy and Gini coefficient of each remote sensing feature;

[0034] The traffic element remote sensing feature determination module is used to select remote sensing features whose feature importance values ​​exceed a set threshold as the traffic element remote sensing features of the area to be extracted.

[0035] The present invention also provides a method for identifying traffic elements, comprising the following steps:

[0036] Execute the remote sensing feature extraction method of traffic elements described in the above technical solution to determine the remote sensing features of traffic elements in the area to be extracted;

[0037] Generate a training set based on the sample set and the feature set; a single sample in the training set includes a remote sensing feature of a traffic element at a certain coordinate position in the area to be extracted as a model input, and a traffic element ground object category at the coordinate position as a training label;

[0038] Use the training set to train the neural network model;

[0039] The remote sensing features of traffic elements of each pixel in the remote sensing image of the area to be extracted are input into the trained neural network model to obtain the category of traffic elements at each coordinate position of the area to be extracted.

[0040] The present invention also provides a traffic element recognition system, including a sample set and feature set generation module, an optimization factor calculation module, a comprehensive feature evaluation module, a traffic element remote sensing feature determination module, a recognition model training module and a ground object category recognition module;

[0041] The sample set and feature set generation module is used to form a sample set and a feature set through remote sensing data of the area to be extracted; wherein the sample set includes the coordinate position in the area to be extracted and the corresponding traffic element and ground object category; the feature set includes the coordinate position in the area to be extracted and the corresponding remote sensing feature set;

[0042] The optimization factor calculation module is used to calculate the sample separability, redundancy and Gini coefficient of each remote sensing feature in the remote sensing feature set based on the sample set and the feature set;

[0043] The comprehensive feature evaluation module is used to calculate the feature importance value of each remote sensing feature based on the sample separability, redundancy and Gini coefficient of each remote sensing feature; the feature importance value is used to evaluate the joint effect of the sample separability, redundancy and Gini coefficient of each remote sensing feature;

[0044] The traffic element remote sensing feature determination module is used to select remote sensing features whose feature importance values ​​exceed a set threshold as the traffic element remote sensing features of the area to be extracted.

[0045] The recognition model training module is used to generate a training set based on a sample set and a feature set; a single sample in the training set includes a remote sensing feature of a traffic element at a certain coordinate position in the area to be extracted as a model input, and a traffic element ground feature category at the coordinate position as a training label; the training set is used to train the neural network model;

[0046] The object category recognition module is used to input the remote sensing features of the traffic elements of each pixel in the remote sensing image of the area to be extracted into the trained neural network model to obtain the traffic element object category of each coordinate position in the area to be extracted.

[0047] The beneficial effects of the present invention are as follows: the present invention combines sample separability (JM), feature correlation (Pearson) and average impurity reduction (Gini) as indicators for evaluating feature importance, fully considers the influence of features on sample selection, other features and algorithms, and improves the robustness of feature analysis to a certain extent; the present invention obtains a comprehensive index value for feature importance evaluation by adopting three key factors for feature optimization, and sets a threshold value for feature optimization accordingly. This method can reduce data redundancy, reduce feature optimization calculation costs and the influence of non-critical features on the accuracy of traffic element extraction from remote sensing images.

[0048] Furthermore, based on the pre-processed multi-source data, the present invention constructs spectral features including bands, spectral indices, principal component transformation, minimum noise separation and tasseled cap transformation; constructs geometric spatial features including texture and shape; constructs radar features including backscattering coefficient and polarization index; constructs elevation features including altitude, slope and aspect, which can effectively reflect the characteristics of remote sensing data, and based on the diversified remote sensing features, can accurately locate the remote sensing features that best reflect the attributes of traffic elements in the area to be tested, thereby improving the accuracy of subsequent traffic element identification.

[0049] Furthermore, the present invention pre-processes the multi-source remote sensing data according to their characteristics. In addition, the nearest neighbor method is used to perform resampling processing to solve the problem of different spatial resolutions of multi-source data. On the basis of multimodal remote sensing big data, a multi-source and multi-type remote sensing feature set is constructed, which can better understand and interpret remote sensing images, fully expand the differences between different remote sensing traffic elements, and help improve the accuracy of remote sensing image traffic element extraction.

[0050] Furthermore, the present invention selects sample separability to evaluate the distinguishability between different types of samples on each feature. Specifically, the Jeffries-Matusita distance is used to represent the sample separability. The larger the value, the higher the contribution of the feature. The method for calculating the sample separability of remote sensing features is specifically proposed for the application scenario of remote sensing feature extraction of traffic elements, which ensures the accuracy of the sample separability evaluation results in this application scenario.

[0051] Furthermore, the present invention uses a correlation coefficient to evaluate the correlation between the features, specifically using the Pearson correlation coefficient, where a larger value represents a higher redundancy. The method for calculating the correlation coefficient of remote sensing features specifically proposed for the application scenario of remote sensing feature extraction of traffic elements ensures the accuracy of the correlation coefficient evaluation result in this application scenario.

[0052] Furthermore, the present invention calculates the weights of three key factors based on the information weight algorithm, and uses a multivariate linear regression model to evaluate the importance of each feature of the remote sensing image traffic element extraction feature set, effectively ensuring the evaluation accuracy. The present invention proposes to set a threshold according to a certain ratio to optimize the features with low importance and high redundancy. The threshold setting method proposed by the present invention ensures that the extracted remote sensing features of traffic elements can fully reflect the characteristics of traffic elements, while saving computing costs for subsequent traffic element identification.

[0053] Furthermore, traffic elements in remote sensing images are extracted based on deep learning: the sample set and feature set information are further used as the input layer of the convolutional neural network model, and the deep features are extracted using the characteristics of the convolutional neural network. Finally, the extraction of traffic elements from remote sensing images can be achieved. It can make full use of the multi-source features of remote sensing images to extract more advanced traffic element features, improve the amount of remote sensing feature information and feature expression capabilities, and thus improve the accuracy of traffic element extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0055] Figure 2 It is a schematic diagram of constructing the multiple linear regression model of the present invention.

[0056] Figure 3 It is a schematic diagram of the neural network model for traffic element recognition in remote sensing images.

[0057] Figure 4 It is a schematic diagram of the accuracy of traffic element extraction before and after the optimization of some features of the traffic element extraction framework. DETAILED DESCRIPTION

[0058] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments to facilitate a clear understanding of the present invention, but they do not constitute a limitation on the present invention.

[0059] Example 1

[0060] like Figure 1 As shown, the present invention provides a method for extracting remote sensing features of traffic elements, such as Figure 1 As shown, the following steps are included:

[0061] The first step is remote sensing data preparation and processing.

[0062] This embodiment uses satellite remote sensing data as an example, and aerial and drone remote sensing images with higher spatial, temporal and spectral resolutions can also be used as basic data sources, such as extracting spectral features, texture features and radar features. This embodiment first obtains Sentinel-2 remote sensing satellite data, radar remote sensing image satellite Sentinel-1 and SRTM elevation data of the traffic element area to be identified.

[0063] For the Sentinel-2 remote sensing satellite data source used in this content, atmospheric correction and terrain correction are first used to eliminate the interference of atmosphere and terrain. Secondly, the bit mask band "QA60" with cloud mask information in the data source is used to remove cloud from the data. Among them, MSK_CLDPRB and MSK_SNWPRB are bands that reflect cloud probability and snow probability, which can be used to remove clouds and snow.

[0064] For the radar remote sensing image satellite Sentinel-1, the Sentinel-1 toolbox is first used to implement thermal noise removal, radiometric calibration and terrain correction. Secondly, in order to retain the prominent edges, linear features and texture information of traffic elements, the improved Lee filter, namely the Refine Lee filter algorithm, is further used for processing. This algorithm is an improvement on the famous traditional Lee filter, and the improvement follows the K nearest neighbor algorithm. In this method, a sliding window of size k*k can be freely selected to obtain the best result. The processing operations are listed in pseudo code form as follows:

[0065] Table 1. SAR image Refine Lee processing flow pseudo code

[0066]

[0067] For SRTM elevation data, interpolation and fitting methods are used to fill the holes in DEM. The bilinear interpolation (Formulas 1 to 4) used are shown below.

[0068] Z=a0+a1X+a2Y (1)

[0069] The parameters a0, a1, and a2 can be calculated strictly based on three known reference points. (X, Y, and Z axes) represent the elevation value of any point. In order to avoid unstable solutions when the geometric shape formed by the three reference points approaches a straight line, a bilinear interpolation method is used based on the three known reference points (A, B, and C) to interpolate the point P (X P , Y P , Z P ) Elevation value:

[0070]

[0071] Where Y P =Y L =Y R , point L(X L , Y L , Z L ), R(X R , Y R , Z R) are located on straight lines AB and AC respectively.

[0072] In addition, since some bands of Sentinel-2 and Sentinel-1 and SRTM elevation data have different spatial resolutions, in order to ensure the consistency of spatial resolution, some bands of Sentinel-2 and SRTM elevation data are resampled to 10m spatial resolution using the nearest neighbor method. The resampling formula (Equation 5) is shown below.

[0073] Image′(x,y)=T[Image(x,y)] (5)

[0074] Where Image'(x, y) is the resampled traffic element image, Image(x, y) is the source image before resampling, and T refers to the conversion function, that is, for each pixel position of the traffic element image, the value of the nearest pixel in the source image is directly taken.

[0075] The second step is to build a sample set and feature set for remote sensing extraction of traffic elements

[0076] Taking the extraction of ground object types from remote sensing images as an example, based on the above-mentioned satellite-borne multi-source data, a sample set and feature set for identifying traffic elements of different ground object categories are constructed. Among them, the construction of the sample set is mainly interpreted according to Google Earth and domestic high-resolution remote sensing images, and verified in combination with field survey data, so as to construct a sample set containing the coordinate positions of various typical ground object categories and their corresponding label attributes. The label attribute refers to the class of traffic element ground objects. Due to the interpretation based on Google Earth and domestic high-resolution remote sensing images, as well as the combination of field survey data, only the class of traffic element ground objects at some coordinate positions of the area to be determined. Therefore, it is necessary to further extract traffic element remote sensing features that can reflect the attributes of different traffic element ground object categories, so as to be used for the subsequent traffic element ground object category identification for each coordinate position of the area (that is, each pixel in the remote sensing image of the area).

[0077] The construction of feature sets includes but is not limited to the following:

[0078] Based on the Sentinel-2 optical remote sensing satellite, spectral features such as bands, spectral indices, principal component transformation, minimum noise separation and tasseled cap transformation are constructed, as well as geometric spatial features such as texture and shape;

[0079] Based on the Sentinel-1 radar remote sensing satellite, radar features including backscatter coefficient and polarization index were constructed;

[0080] Elevation features such as altitude, slope and aspect are constructed based on SRTM data. The specific feature set is shown in Table 2.

[0081] Table 2. Multimodal remote sensing feature set

[0082]

[0083] For each sample in the sample set, the corresponding multimodal remote sensing data is located by its coordinate position, so as to calculate its corresponding multimodal remote sensing features, and then form the corresponding relationship between different multimodal remote sensing features and different traffic element and ground object categories.

[0084] The third step is to calculate and process the three key factors for feature optimization

[0085] According to the results of multimodal remote sensing data preprocessing, sample set and feature set construction, the relationship between them and the extraction of traffic elements from remote sensing images is analyzed from the perspectives of samples, features and algorithms to determine the key factors for feature optimization.

[0086] ① Sample separability is to evaluate the importance of features from the perspective of sample separability. Generally speaking, if the degree of differentiation of traffic elements in different remote sensing images is large under the same feature, it means that the feature can improve the accuracy of traffic element extraction and the feature is more important. Otherwise, it means that the feature is not important. Using JM distance to evaluate sample separability and using JM distance to calculate the differences between land types is a universal quantitative method. The JM distance calculation result value is between 0 and 2. The larger the value, the better the separability. If it is greater than 1.8, the sample separability is very good. Between 1.4 and 1.8, the sample selection is qualified. If the sample separability is less than 1.4, the separability between land types is poor. The JM distance calculation process and formula are as follows:

[0087] JM=2*(1-e -B ) (6)

[0088] In the above formula, B is the Bhattacharyya distance on a certain feature dimension. The calculation formula of the Bhattacharyya distance B of samples selected for different land types is as follows:

[0089]

[0090] In the above formula, i and j represent different traffic element land types, mi represents the average value of each sample in the traffic element land type i sample set on the remote sensing feature, δ i represents the variance of each sample in the traffic element land type i sample set on the remote sensing feature; m j represents the average value of each sample in the traffic element land class j sample set on the remote sensing feature, δ j It represents the variance of each sample in the traffic element land type j sample set on the remote sensing feature.

[0091] ② The correlation coefficient between features can indicate the degree of data redundancy between features. The larger the correlation coefficient between features, the more similar the effects of the two on traffic element extraction are, which will lead to data redundancy. The calculation process is as follows: First, the Pearson correlation coefficient is used to express the correlation between features, and the calculation formula is shown in Formula 8. Secondly, the redundancy degree of each feature is expressed by the average value of the Pearson coefficient of a feature and other features, and the calculation formula is shown in Formula 9.

[0092]

[0093] Where X and Y refer to any two remote sensing features; cov(x,y) is the covariance of X and Y; σ x σ y is the product of the standard deviations of X and Y; E(X) and E(Y) are the mathematical expectations of X and Y respectively.

[0094]

[0095] In the above formula, T refers to the correlation coefficient between two remote sensing features; n refers to the number of remote sensing features; I t It represents the remote sensing redundancy of the tth feature. The larger its value is, the stronger the redundancy is.

[0096] ③ Gini coefficient is also a commonly used feature optimization method to evaluate feature importance. It is an indicator to measure classification ability. The Gini coefficient is defined based on the concept of impurity. Impurity refers to the imbalance of the proportion of samples of different categories in a sample set. The Gini coefficient measures the contribution of a certain feature. The larger the value of the Gini coefficient, the greater the impact of the feature on the classification result and the more important it is. The specific calculation formula is as follows:

[0097]

[0098] Among them, GI is the Gini index, K is the training sample category (i.e., the category of traffic elements and features), It is the estimated probability that the sample at node m (i.e. the mth sample in the sample set) belongs to the kth type of traffic element. If the classification result is two types of traffic elements, that is, K is 2, GI = 2p(1-p). In actual application, the value of K will change according to the actual situation of the area to be extracted, and different areas contain different numbers of traffic elements.

[0099] If the number of nodes in the decision tree is n, then:

[0100]

[0101] Among them, GVIjkn is the importance of variable j (i.e., the jth remote sensing feature) to the kth category (traffic element feature category) at node n (i.e., the nth sample in the sample set); GI parent , GI daughterleft and GI daughterright They are the Gini index of node n and the Gini index of the two new nodes after the split. The importance of variable j to class k can be obtained by summing the nodes in class k that contain variable j:

[0102]

[0103] Where N is the total number of nodes. The importance of variable j in all categories can be obtained by the following formula:

[0104]

[0105] Among them, ntree is the number of decision trees.

[0106] Step 4: Build a comprehensive feature optimization model

[0107] Based on the three key factors for evaluating feature importance obtained above, a multivariate linear regression model is used to construct a comprehensive feature optimization value. The multivariate linear regression model is a statistical method used to evaluate the relationship between a dependent variable and multiple independent variables. By using this model, the joint effect between the three key factors for feature optimization can be found. In this method, the calculation process is as follows:

[0108] Assume that the feature importance values ​​of m features, that is, the dependent variable is Y; and only consider the influencing factors as three key factors, then the independent variables are X1: [x 11 , x 12 , …, x 1m ], X2: [x 21 , x 22 , …, x 3m ], X3: [x 31 , x 32 , …, x 3m ], then the intercept of the multiple linear regression model is assumed to be 0. Therefore, the multiple linear regression model is as follows:

[0109] Y=β1X1+β2X2+β3X3 (14)

[0110] Where β1, β2, and β3 are regression coefficients / weights, which indicate the degree of influence of the independent variable on the dependent variable, and their specific values ​​are calculated using the information weight method. The information weight method is also called the coefficient of variation method. The information weight method is an objective weighting method. Its idea is to use the coefficient of variation of the data to assign weights. If the coefficient of variation is larger, it means that it carries more information, and thus the weight will be larger. The calculation process is as follows:

[0111] ① Calculate the mean and standard deviation:

[0112]

[0113] Among them, m refers to the number of remote sensing features, n = 1, 2, 3 represent three optimization factors respectively, represents mean, SD n Represents standard deviation.

[0114] ② Calculate the coefficient of variation (CV coefficient), which is the ratio of the standard deviation to the mean:

[0115]

[0116] Among them, n=1, 2, 3.

[0117] ③ Calculate the weight, which is obtained by summing and normalizing the coefficient of variation:

[0118]

[0119] Among them, k=3, n=1, 2, 3.

[0120] Therefore, the obtained W1, W2, and W3 refer to β1, β2, and β3, from which the calculation formula of the feature importance value can be obtained.

[0121] On the basis of the calculation of the above comprehensive feature evaluation index, the threshold method is further used to obtain the optimized remote sensing image traffic element extraction feature set. In general, in order to reduce data redundancy as much as possible without losing the remote sensing information of different traffic elements owned by the original features, the threshold setting range is between 0 and 1. The optimization feature set screening calculation formula is as follows:

[0122] Assume that the set of m eigenvalues ​​obtained by the comprehensive feature evaluation model is M = {N1, N2, ..., N m}

[0123] T=[Max M -Min M ]×t (19)

[0124] Where Max M Refers to the maximum value in the eigenvalue set, Min MRefers to the minimum value in the feature value set, and t is a custom threshold. The features in M ​​that are greater than T are retained and grouped into an optimized feature set as the remote sensing features of traffic elements in the area to be extracted. The specific feature optimization process is as follows: Figure 2 shown.

[0125] Example 2

[0126] After executing the remote sensing feature extraction method for traffic elements described in the above technical solution, the present invention continues to execute subsequent steps.

[0127] The fifth step is to extract traffic elements from remote sensing images based on deep learning.

[0128] Generate a training set based on the sample set and the feature set; a single sample in the training set includes a remote sensing feature of a traffic element at a certain coordinate position in the area to be extracted as a model input, and a traffic element ground object category at the coordinate position as a training label;

[0129] Use the training set to train the neural network model;

[0130] The remote sensing features of traffic elements of each pixel in the remote sensing image of the area to be extracted are input into the trained neural network model to obtain the category of traffic elements at each coordinate position of the area to be extracted.

[0131] Specifically, based on the constructed sample set and optimized feature set, the sample set is divided into a training set and a validation set in a ratio of 7:3, which are used to train the model and evaluate the model accuracy respectively. The input samples and training labels of the validation set are consistent with those of the training set.

[0132] The convolutional neural network model is used to perform convolution operations on the input remote sensing feature data to learn the deep features of the input feature graphs of different traffic element categories. The calculation formula is as follows:

[0133] F i =σ(W s *X+b s ) (20)

[0134] Where F s represents the output feature map of the convolutional layer, W s refers to the convolution kernel, σ refers to the activation function, X is the input remote sensing feature, b s is the bias term, and s represents the index of the convolutional layer or feature map.

[0135] Traffic elements are extracted based on the trained convolutional neural network, and the Softmax classifier is used to classify the remote sensing features of each pixel in the image to determine which traffic element category each pixel belongs to.

[0136] P(Ci )=e zi / ∑ j e zi ) (twenty one)

[0137] In the above formula, P(C i ) represents the probability of traffic feature category i, e zi It refers to the score of image e in traffic feature category i, and j is the loss function.

[0138] like Figure 3 As shown, a training set is generated based on the multi-source remote sensing images of the area to be measured and the traffic element object categories corresponding to the known coordinate positions to train the neural network model. Then, the multi-source remote sensing images of the area to be measured are obtained to obtain the remote sensing features of the traffic elements at each pixel position (i.e., the optimized feature set), and the input values ​​are used as the training neural network model. The neural network model interprets the image traffic element features based on the activated areas in the extracted deep feature map, thereby realizing the recognition of the traffic element areas on the remote sensing image.

[0139] In this embodiment, the following indicators are used to evaluate the accuracy of the neural network model in identifying traffic elements and the model performance before and after feature optimization. The evaluation indicators are mainly shown in Table 3:

[0140] Table 3. Performance evaluation indicators of traffic element extraction model

[0141]

[0142] Note: TP is the positive sample predicted by the model as positive; TN is the negative sample predicted by the model as negative; FP is the negative sample predicted by the model as positive, that is, the number of false positive negative samples; FN is the positive sample predicted by the model as negative, that is, the number of false positive positive samples. Figure 4 As shown in FIG. 1 , the remote sensing features selected when the threshold t is defined as 0.6 are taken as the remote sensing features of the traffic elements in the area to be measured. The accuracy of traffic element extraction reaches a stable high value at the remote sensing features of the traffic elements in the selected area to be measured.

[0143] Example 3

[0144] The present invention also provides a traffic element remote sensing feature extraction system, including: a sample set and feature set generation module, an optimization factor calculation module, a comprehensive feature evaluation module and a traffic element remote sensing feature determination module;

[0145] The sample set and feature set generation module is used to form a sample set and a feature set through remote sensing data of the area to be extracted; wherein the sample set includes the coordinate position in the area to be extracted and the corresponding traffic element and ground object category; the feature set includes the coordinate position in the area to be extracted and the corresponding remote sensing feature set;

[0146] The optimization factor calculation module is used to calculate the sample separability, redundancy and Gini coefficient of each remote sensing feature in the remote sensing feature set based on the sample set and the feature set;

[0147] The comprehensive feature evaluation module is used to calculate the feature importance value of each remote sensing feature based on the sample separability, redundancy and Gini coefficient of each remote sensing feature; the feature importance value is used to evaluate the joint effect of the sample separability, redundancy and Gini coefficient of each remote sensing feature;

[0148] The traffic element remote sensing feature determination module is used to select remote sensing features whose feature importance values ​​exceed a set threshold as the traffic element remote sensing features of the area to be extracted.

[0149] Example 4

[0150] The present invention also provides a traffic element recognition system, including a sample set and feature set generation module, an optimization factor calculation module, a comprehensive feature evaluation module, a traffic element remote sensing feature determination module, a recognition model training module and a ground object category recognition module;

[0151] The sample set and feature set generation module is used to form a sample set and a feature set by acquiring remote sensing data of the area to be extracted; wherein the sample set includes the coordinate position in the area to be extracted and the corresponding traffic element and ground object category; the feature set includes the coordinate position in the area to be extracted and the corresponding remote sensing feature set;

[0152] The optimization factor calculation module is used to calculate the sample separability, redundancy and Gini coefficient of each remote sensing feature in the remote sensing feature set based on the sample set and the feature set;

[0153] The comprehensive feature evaluation module is used to calculate the feature importance value of each remote sensing feature based on the sample separability, redundancy and Gini coefficient of each remote sensing feature; the feature importance value is used to evaluate the joint effect of the sample separability, redundancy and Gini coefficient of each remote sensing feature;

[0154] The traffic element remote sensing feature determination module is used to select remote sensing features whose feature importance values ​​exceed a set threshold as the traffic element remote sensing features of the area to be extracted.

[0155] The recognition model training module is used to generate a training set based on a sample set and a feature set; a single sample in the training set includes a remote sensing feature of a traffic element at a certain coordinate position in the area to be extracted as a model input, and a traffic element ground feature category at the coordinate position as a training label; the training set is used to train the neural network model;

[0156] The object category recognition module is used to input the remote sensing features of the traffic elements of each pixel in the remote sensing image of the area to be extracted into the trained neural network model to obtain the traffic element object category of each coordinate position in the area to be extracted.

[0157] Example 5

[0158] The present invention provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method for extracting remote sensing features of traffic elements described in the above technical solution is implemented.

[0159] Example 6

[0160] The present invention provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the traffic element identification method described in the above technical solution is implemented.

[0161] Example 7

[0162] The present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method for extracting remote sensing features of traffic elements described in the above technical solution by executing the computer instructions.

[0163] Example 8

[0164] The present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the traffic element identification method described in the above technical solution by executing the computer instructions.

[0165] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0167] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0168] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0169] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

[0170] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A method for extracting remote sensing features of traffic elements, characterized in that: The following steps are involved: A sample set and a feature set are formed by using the remote sensing data of the area to be extracted; wherein the sample set includes the coordinate position in the area to be extracted and the corresponding traffic element and ground object category; the feature set includes the coordinate position in the area to be extracted and the corresponding remote sensing feature set; Based on the sample set and feature set, the sample separability, redundancy and Gini coefficient of each remote sensing feature in the remote sensing feature set are calculated respectively; The feature importance value of each remote sensing feature is calculated based on the sample separability, redundancy and Gini coefficient of each remote sensing feature; the feature importance value is used to evaluate the joint effect of the sample separability, redundancy and Gini coefficient of each remote sensing feature; The remote sensing features whose feature importance values ​​exceed the set threshold are selected as the remote sensing features of the traffic elements in the area to be extracted.

2. A method for extracting remote sensing features of traffic elements according to claim 1, characterized in that: The remote sensing features in the remote sensing feature set include: band, spectral index, spectral transformation feature, texture feature, geometric feature, backscattering coefficient, polarization index and elevation.

3. A method for extracting remote sensing features of traffic elements according to claim 2, characterized in that: Also includes: Obtain Sentinel-2 optical remote sensing satellite images, Sentinel-1 radar remote sensing satellite images, and SRTM digital elevation model data of the area to be extracted; The nearest neighbor method was used to resample the Sentinel-2 optical remote sensing satellite images, Sentinel-1 radar remote sensing satellite images and SRTM digital elevation model data; After the Sentinel-2 optical remote sensing satellite images were subjected to radiometric calibration, atmospheric correction, terrain correction, cloud and snow removal, and mosaic cropping preprocessing, the bands, spectral index, spectral transformation characteristics, texture characteristics, and geometric characteristics were extracted; After removing thermal noise, calibrating radiation, correcting terrain and preprocessing with the RefineLee filtering algorithm, the backscattering coefficient and polarization index of the Sentinel-1 radar remote sensing satellite images were extracted. For the SRTM digital elevation model data, after preprocessing by filling holes, the altitude, slope, aspect and mountain shadow are extracted.

4. A method for extracting remote sensing features of traffic elements according to claim 1, characterized in that: The sample separability of any remote sensing feature is calculated using the following formula: JM=2*(1-e -B ); Among them, JM represents the JM distance, which is used to evaluate the separability of samples; B is the Bhattacharyya distance on the remote sensing feature dimension; i and j represent different types of traffic elements, m i represents the average value of each sample in the traffic element land type i sample set on the remote sensing feature, δ i represents the variance of each sample in the traffic element land type i sample set on the remote sensing feature; m j represents the average value of each sample in the traffic element land class j sample set on the remote sensing feature, δ j Represents the variance of each sample in the traffic element j sample set on this feature.

5. The method for extracting remote sensing features of traffic elements according to claim 1, characterized in that: For any remote sensing feature, the average value of the Pearson coefficient between the remote sensing feature and other remote sensing features is used as the redundancy evaluation value.

6. A method for extracting remote sensing features of traffic elements according to claim 1, characterized in that: The feature importance value Y is calculated using the following formula: Y = β1X1 + β2X2 + β3X3; Where, X1: [x 11 , x 12 , …, x 1m ], X2: [x 21 , x 22 , …, x 3m ], X3: [x 31 , x 32 , …, x 3m ], m represents the number of remote sensing features, x pq It represents the normalized value of the pth optimization factor of the qth remote sensing feature; the optimization factors include sample separability, redundancy and Gini coefficient; p = 1, 2, 3; q∈[1, m]; β1, β2, β3 are regression coefficients, calculated using the information weight method.

7. A method for extracting remote sensing features of traffic elements according to claim 1, characterized in that: The calculation process of setting the threshold T includes: The feature importance value set of remote sensing features is defined as M = {N1, N2, ..., N m }; T=[Max M -Min M ]×t; Where m represents the number of remote sensing features, Max M Refers to the maximum value in the eigenvalue set, Min M Refers to the minimum value in the feature value set, and t is a custom threshold.

8. A remote sensing feature extraction system for traffic elements, characterized in that include: Sample set and feature set generation module, optimization factor calculation module, comprehensive feature evaluation module and traffic element remote sensing feature determination module; The sample set and feature set generation module is used to form a sample set and a feature set through remote sensing data of the area to be extracted; wherein the sample set includes the coordinate position in the area to be extracted and the corresponding traffic element and ground object category; the feature set includes the coordinate position in the area to be extracted and the corresponding remote sensing feature set; The optimization factor calculation module is used to calculate the sample separability, redundancy and Gini coefficient of each remote sensing feature in the remote sensing feature set based on the sample set and the feature set; The comprehensive feature evaluation module is used to calculate the feature importance value of each remote sensing feature based on the sample separability, redundancy and Gini coefficient of each remote sensing feature; the feature importance value is used to evaluate the joint effect of the sample separability, redundancy and Gini coefficient of each remote sensing feature; The traffic element remote sensing feature determination module is used to select remote sensing features whose feature importance values ​​exceed a set threshold as the traffic element remote sensing features of the area to be extracted.

9. A method for identifying traffic elements, characterized in that: The following steps are involved: Execute the remote sensing feature extraction method of traffic elements as described in any one of claims 1 to 8 to determine the remote sensing features of traffic elements in the area to be extracted; Generate a training set based on the sample set and the feature set; a single sample in the training set includes a remote sensing feature of a traffic element at a certain coordinate position in the area to be extracted as a model input, and a traffic element ground object category at the coordinate position as a training label; Use the training set to train the neural network model; The remote sensing features of traffic elements of each pixel in the remote sensing image of the area to be extracted are input into the trained neural network model to obtain the category of traffic elements at each coordinate position of the area to be extracted.

10. A traffic element recognition system, characterized in that: It includes sample set and feature set generation module, optimization factor calculation module, comprehensive feature evaluation module, traffic element remote sensing feature determination module, recognition model training module and ground object category recognition module; The sample set and feature set generation module is used to form a sample set and a feature set through remote sensing data of the area to be extracted; wherein the sample set includes the coordinate position in the area to be extracted and the corresponding traffic element and ground object category; the feature set includes the coordinate position in the area to be extracted and the corresponding remote sensing feature set; The optimization factor calculation module is used to calculate the sample separability, redundancy and Gini coefficient of each remote sensing feature in the remote sensing feature set based on the sample set and the feature set; The comprehensive feature evaluation module is used to calculate the feature importance value of each remote sensing feature based on the sample separability, redundancy and Gini coefficient of each remote sensing feature; the feature importance value is used to evaluate the joint effect of the sample separability, redundancy and Gini coefficient of each remote sensing feature; The traffic element remote sensing feature determination module is used to select remote sensing features whose feature importance values ​​exceed a set threshold as the traffic element remote sensing features of the area to be extracted. The recognition model training module is used to generate a training set based on a sample set and a feature set; a single sample in the training set includes a remote sensing feature of a traffic element at a certain coordinate position in the area to be extracted as a model input, and a traffic element ground feature category at the coordinate position as a training label; the training set is used to train the neural network model; The object category recognition module is used to input the remote sensing features of the traffic elements of each pixel in the remote sensing image of the area to be extracted into the trained neural network model to obtain the traffic element object category of each coordinate position in the area to be extracted.

Images