A supervised classification method for remote sensing images based on GA optimized PPLN

By using genetic algorithms to optimize the projection tracking learning network in remote sensing image classification, the problem of easily falling into local extreme values ​​in the existing technology is solved, the classification accuracy and training speed are improved, and global convergence and robustness are achieved.

CN118675027BActive Publication Date: 2024-12-03XINGYUN BAOTAI (BEIJING) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410729378.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-06
Publication Date
2024-12-03
Estimated Expiration
2044-06-06

AI Technical Summary

Technical Problem

The existing remote sensing image classification methods are prone to falling into local extreme values ​​and difficult to converge to the optimal value, resulting in low classification accuracy, especially when processing high-dimensional data, lacking self-learning ability.

Method used

Using the projection tracking learning network (PPLN) based on genetic algorithm (GA) optimization, the initial value of the input weight and smoothing function is calculated through the genetic algorithm, combined with the Gauss-Newton method optimization, the hidden layer nodes are gradually added to achieve global convergence and rapidity. train.

Benefits of technology

It overcomes the problem that traditional neural networks are prone to fall into local extreme values, improves the accuracy and training speed of remote sensing image classification, has robustness and rapid convergence capabilities, and enhances learning capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118675027B_ABST
    Figure CN118675027B_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of remote sensing satellite measurement, and specifically relates to a remote sensing image supervised classification method based on GA optimized PPLN, comprising the following steps: collecting remote sensing image samples; selecting samples of different modes from the collected samples, and respectively constructing sample data sets; determining the network structure of a projection pursuit learning network; training the projection pursuit learning network by using input and output sample data sets, and determining the state of the projection pursuit learning network, wherein a genetic algorithm is used to calculate input weights, coefficients of a smoothing function, and initial values ​​of output weights; inputting an image to be identified into the trained projection pursuit learning network for identification and classification; the invention has the dimensionality reduction capability of projection pursuit, the extensive mapping capability of a neural network, and the performance of GA fast, global convergence, and enhanced learning capability, and can improve the image classification accuracy and overcome the problem of being easily trapped in a local minimum.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing, and in particular relates to a remote sensing image supervised classification method based on GA optimized PPLN, which is applied to supervised classification of multispectral or hyperspectral remote sensing images. Background Art

[0002] Remote sensing image classification is to classify each pixel in the image into different categories according to its spectral brightness, spatial structure characteristics or other information in different bands according to certain rules and algorithms. It is the basis for applications such as environmental and disaster monitoring, agriculture, forestry, and land resource utilization. In the current remote sensing classification applications, traditional pattern recognition classification methods include minimum distance method, parallelepiped method, maximum likelihood method, ISODATA, chain method and other classification methods. Due to the spatial resolution of the remote sensing image itself and the existence of "same object different spectrum" and "different objects same spectrum" phenomena, the classification results often have more misclassification and omission phenomena, resulting in low classification accuracy. With the development of remote sensing application technology, artificial neural networks have been widely used in remote sensing image classification. Its main idea is to use the extracted features of the remote sensing image as the input signal of the neural network. After the neural network is trained according to certain rules, the output signal can be classified at the output end. The neural network has good parallel processing performance, can adapt to the needs of multiple bands, has good generalization ability, and does not require the prior probability distribution of the data. Compared with the statistical method, the neural network method has great superiority. However, most neural network remote sensing image classifiers use a feedforward neural network structure and implement supervised image classification using the BP learning algorithm. The main drawback is that they are prone to fall into local extreme values ​​and are not easy to converge to the optimal value, which affects the classification accuracy. Projection Pursuit (PP) is a new statistical method for processing high-dimensional data, but it has the disadvantage of no self-learning function. Summary of the invention

[0003] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a supervised classification method of remote sensing images based on GA optimized PPLN (Classification of Remote Sensing Image based on PPLN, PPLN-CRSI).

[0004] In order to achieve the purpose of the present invention, the present invention will be implemented by adopting the technical solution described below.

[0005] A remote sensing image supervised classification method based on GA optimized PPLN includes the following steps:

[0006] S1, collect remote sensing image samples;

[0007] S2. Select samples of different modes from the collected samples and construct sample data sets respectively. The number of the sample data sets is the number of image classifications.

[0008] S3, determining the network structure of the projection pursuit learning network, wherein the number of network input layer nodes of the network structure is the same as the number of bands of the remote sensing image, each node corresponds to one band, the number of network output layer nodes is the same as the number of image categories, and the hidden layers all use Hermite polynomial functions as smoothers, wherein the highest power of the polynomial is determined according to the number of image categories;

[0009] S4. Training the projection pursuit learning network using the input and output sample data sets and determining the state of the projection pursuit learning network to achieve a given input-output mapping relationship; wherein,

[0010] The training of all parameters in the projection pursuit learning network is that the projection pursuit learning network groups all parameters with the hidden layer nodes as the center, starts from the input weight and output weight of the first hidden layer node, and sequentially trains the input weight and output weight of each hidden layer node;

[0011] In the process of training all parameters, when a hidden layer node is added, a genetic algorithm is used to calculate the input weight, the coefficient of the smoothing function and the initial value of the output weight. The calculation process of the initial value includes the following steps:

[0012] S41, determining the distribution range of the current node subgroup parameter, the value interval of the input weight and the output weight is between [-1, 1];

[0013] S42, selecting m global evolutionary solutions as initial values ​​of subgroup parameters through a genetic algorithm, the specific selection process includes the following steps:

[0014] S421. Randomly generate an initial population in the encoded solution space;

[0015] S422, evaluate the fitness of each individual in the initial population, and the fitness objective function takes the inverse of the loss function L2;

[0016] S423, performing selection, crossover and mutation genetic operations on individuals in the initial population to generate a new generation population;

[0017] S424, repeat the evolution until the termination condition is met;

[0018] S425. Sort the individuals by quality, select m global evolutionary solutions, and avoid similar individuals when selecting;

[0019] S43, using these evolutionary solutions as initial solutions for parameters related to newly added nodes, and further optimizing subgroup parameters using a projection pursuit learning network;

[0020] S431, optimize input weights using Gauss-Newton method;

[0021] S432, fitting and approximating the hidden layer smoothing function to obtain optimized polynomial coefficients;

[0022] S433, repeat steps S431-S432 for several iterations;

[0023] S434, using the latest input weights and polynomial coefficients, the output weights are estimated using the following formula:

[0024]

[0025] S435, repeating steps S431-S434 until the relative error of the loss function converges;

[0026] S44, comparing the network loss function L2 obtained by further optimizing the m groups of initial solutions through PPLN, and selecting the optimization solution with the smallest L2 as the optimal solution for the node subgroup parameters;

[0027] S45, if the network loss function L2 is less than the predetermined target value, or the number of hidden layer nodes reaches the predetermined number, the network training ends; otherwise, a hidden layer node is added and the process goes to step S41;

[0028] S5. Input the image to be identified into the projection pursuit learning network trained in step S4 for identification and classification.

[0029] As a preferred solution of the present invention, the state in S4 includes the number of hidden layer nodes, polynomial functions, input weights and output weights.

[0030] As a preferred embodiment of the present invention, the specific selection process of the evolutionary solution includes the following steps:

[0031] S421. To solve the minimum value of the input weight, establish a minimization problem expression:

[0032]

[0033] S422, Real number encoding of optimization variables:

[0034] The following linear transformation formula is used:

[0035] α(j)=a(j)+y(j)(b(j)-a(j)),j=1,2,…,p,, (2)

[0036] Among them, L2 is the optimization objective function, and p is the number of optimization variables. Formula (2) corresponds the jth variable to be optimized α(j) on the initial variable interval [a(j), b(j)] to the real number y(j) on the interval [-1,1], where y(j) is the genetic gene. At this time, the genes corresponding to all the variables in the optimization problem are sequentially connected to form the encoding form of the solution to the problem (y(1), y(2), ..., y(p)), which is called a chromosome. After encoding, the value range of all optimization variables becomes the interval [-1,1].

[0037] S423. Generation of initial parent population:

[0038] Assume that the population size is n, generate n groups of uniform random numbers in the interval [-1,1], each group has p numbers, that is, {y(i,j)|i=1,2,…,n;j=1,2,…,p}, substitute y(i,j) into equation (2) to get the optimized variable value α(i,j), and replace the p-dimensional vector α i Normalize to make ||α i ||=1; Substitute α(i,j) into formula (1) to obtain the corresponding objective function value L2(i), sort {L2(i)|i=1,2,…,n} in ascending order, and sort the corresponding individuals {y(i,j)} accordingly. The first few individuals after sorting are called excellent individuals and are allowed to enter the next generation directly;

[0039] S424, fitness evaluation:

[0040] The smaller the objective function value L2(i), the higher the fitness of the individual, and vice versa. Based on this, the fitness function value F(i) of the i-th parent individual after sorting is defined as:

[0041]

[0042] In the formula, "0.0001" in the denominator is set empirically to avoid the situation where the F(i) value is 0;

[0043] S425, Select:

[0044] The selection operation generates the first offspring population {y1(i,j)|i=1,2,…,n;j=1,2,…,p}. According to the fitness function value, a probability is set for each chromosome y(i,j) in the population, so that the possibility of the chromosome being selected is proportional to the fitness of other chromosomes in the population; the stronger the fitness of the chromosome, the greater the possibility of being selected; using the proportional selection method, the selection probability ps(i) of the parent individual y(i,j) is:

[0045]

[0046] make The sequence {p(i)|i = 1, 2, …, n} divides the interval [0, 1] into n sub-intervals, and these sub-intervals correspond one-to-one with n parent individuals;

[0047] Generate n - 4 random numbers {u(k)|k = 1, 2, …, n - 4}. If u(k) is in (p(i - 1), p(i)], then the i-th individual y(i, j) is selected. In this way, the i-th individual is selected from the parent population y(i, j) with probability ps(i), and a total of n - 4 individuals are selected. Then perform the immigration operation, and directly add the 4 best parent individuals into the offspring population, y1(i + n - 4, j) = y(i, j), i = 1, 2, …, 4;

[0048] S426. Crossover:

[0049] Perform the crossover operation to generate the second offspring population {y2(i, j)|i = 1, 2, …, n; j = 1, 2, …, p}. First, define the crossover parameter Pc as the probability of the crossover operation. This probability indicates that there are an expected number of PcN chromosomes in the population that will perform the crossover operation. The specific crossover operation is to randomly select a pair of parent individuals y(i1, j) and y(i2, j) as parents according to the selection probability in Equation (4), generate a random number c from (0, 1), and perform the crossover operation between y(i1, j) and y(i2, j) in the following form to generate the following two offspring individuals y2(i1, j) and y2(i2, j):

[0050]

[0051] S427. Mutation:

[0052] Perform the mutation operation to generate the third offspring population {y3(i, j)|i = 1, 2, …, n; j = 1, 2, …, p}. Define the mutation parameter Pm as the mutation probability in the genetic system. This probability indicates that there are an expected number of PmN chromosomes in the population used for the mutation operation. First, sort the parents according to the fitness function value. Then, from i = 1 to N, repeat the following process: Generate a random number r from the interval [0, 1]. If r < Pm, then select the chromosome y(i, j) as the mutated parent, and mutate each selected parent according to the following method;

[0053]

[0054] In the formula, u(j) (j = 1, 2, …, p) is a random number on the interval [a(j), b(j)], x is a random number on (0.8, 1.2). If the generated y3(i, j) exceeds the upper and lower limits of the value range, then directly assign the parent chromosome to the offspring. That is, for individuals with poor fitness, large-scale mutations are adopted; while for individuals with good fitness, small-scale mutations are adopted;

[0055] S428, Evolution Iteration:

[0056] The 3n offspring individuals generated previously are sorted from large to small according to their fitness function values, and the first n offspring individuals are taken as the new parent population; the algorithm enters S424 to enter the next round of selection, crossover and mutation calculations, and stops the optimization calculation when the given end condition is met, and outputs the optimal input weight α.

[0057] As a preferred solution of the present invention, the initial value is a global minimum value.

[0058] As a preferred solution of the present invention, the recognition classification in S5 is to standardize the grayscale values ​​of each band of each pixel in the image to be recognized, and use them as the input vector of the input node of the projection pursuit learning network, obtain the output vector by calculation, compare the output result of the projection pursuit learning network with each category of expected output values, and classify the pixels into a category with the smallest error.

[0059] Beneficial Effects

[0060] Projection Pursuit Learning Network (PPLN) is a new type of network that combines artificial neural network and projection pursuit. It not only overcomes some defects of using only artificial neural network BP algorithm, but also makes up for the lack of self-learning ability of using only PP method. It has the characteristics of simple structure and strong robustness.

[0061] Genetic Algorithm (GA) uses an overall search strategy, and the optimization calculation does not rely on the traditional gradient method, so it is particularly suitable for dealing with highly complex nonlinear problems that are difficult to solve with traditional methods. Its advantages are: first, GA is not easy to fall into local optimality during the search process; second, GA also uses parallel methods to deal with problems, and is more suitable for dealing with large-scale parallel problems, which is consistent with the actual situation; in addition, it is also very easy to combine with other technologies to form a better solution method.

[0062] Combining GA and PPLN can combine the dimensionality reduction ability of projection pursuit, the extensive mapping ability of neural network, and the fast, global convergence and enhanced learning ability of GA. To some extent, it can avoid the shortcomings of slow convergence and local minima of traditional PPLN learning algorithm. Therefore, GA can be used to optimize the connection weights of PPLN model to improve its training speed, overcome its easy to fall into local minima, and establish a remote sensing image supervised classification method based on GA optimized projection pursuit learning network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1The PPLN training flow chart for optimizing the initial value of input weights in the present invention;

[0064] Figure 2 is the connection weight space of multiple minimum points in the present invention;

[0065] Figure 3 This is the training flow chart of PPLN based on GA optimization;

[0066] Figure 4 This is the process structure diagram of supervised classification of remote sensing images based on PPLN;

[0067] Figure 5 It is the RGB image of AVIRIS image;

[0068] Figure 6 This is a video data map;

[0069] Figure 7 Spectral curves of eight types of ground objects;

[0070] Figure 8 Figure 2 is a supervised classification diagram of hyperspectral images, where (a) is the classification result diagram of PPLN-CRSI, (b) is the classification result diagram of the parallelepiped method, (c) is the classification result diagram of the minimum distance method, (d) is the classification result diagram of the maximum likelihood method, (e) is the classification result diagram of the BP neural network method, and (f) is the experimental sample data diagram. DETAILED DESCRIPTION

[0071] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0072] As an embodiment of the present invention, Figures 1 to 4 As shown in FIG. 1 , a remote sensing image supervised classification method based on GA optimized PPLN includes the following steps:

[0073] S1, collect remote sensing image samples;

[0074] S2. Select samples of different modes from the collected samples and construct sample data sets respectively. The number of the sample data sets is the number of image classifications.

[0075] S3, determining the network structure of the projection pursuit learning network, wherein the number of network input layer nodes of the network structure is the same as the number of bands of the remote sensing image, each node corresponds to one band, the number of network output layer nodes is the same as the number of image categories, and the hidden layers all use Hermite polynomial functions as smoothers, wherein the highest power of the polynomial is determined according to the number of image categories;

[0076] S4. Training the projection pursuit learning network using the input and output sample data sets and determining the state of the projection pursuit learning network to achieve a given input-output mapping relationship; wherein,

[0077] The training of all parameters in the projection pursuit learning network is that the projection pursuit learning network groups all parameters with the hidden layer nodes as the center, starts from the relevant variables of the first hidden layer node, and sequentially trains the relevant variables of each hidden layer node, wherein the relevant variables include input weights and output weights;

[0078] In the process of training all parameters, when a hidden layer node is added, a genetic algorithm is used to calculate the input weight, the coefficient of the smoothing function and the initial value of the output weight. The calculation process of the initial value includes the following steps:

[0079] S41, determining the distribution range of the current node subgroup parameter, the value interval of the input weight and the output weight is between [-1, 1];

[0080] S42, selecting a number of global evolutionary solutions as initial values ​​of subgroup parameters through a genetic algorithm, the specific selection process includes the following steps:

[0081] S421. Randomly generate an initial population in the encoded solution space;

[0082] S422, evaluate the fitness of each individual in the initial population, and the fitness objective function takes the inverse of the loss function L2;

[0083] S423, performing selection, crossover and mutation genetic operations on individuals in the initial population to generate a new generation population;

[0084] S424, repeat the evolution until the termination condition is met;

[0085] S425. Sort the individuals by quality, select m global evolutionary solutions, and avoid similar individuals when selecting;

[0086] S43, using these evolutionary solutions as initial solutions for parameters related to newly added nodes, and further optimizing subgroup parameters using a projection pursuit learning network;

[0087] S431, optimize input weights using Gauss-Newton method;

[0088] S432, fitting and approximating the hidden layer smoothing function to obtain optimized polynomial coefficients;

[0089] S433, repeat steps S431-S432 for several iterations;

[0090] S434, estimating the output weights using the following formula based on the latest input weights and polynomial coefficients;

[0091]

[0092] S435, repeating steps S431-S434 until the relative error of the loss function converges;

[0093] S44, comparing the network loss function L2 obtained by further optimizing the m groups of initial solutions through PPLN, and selecting the optimization solution with the smallest L2 as the optimal solution for the node subgroup parameters;

[0094] S45, if the network loss function L2 is less than the predetermined target value, or the number of hidden layer nodes reaches the predetermined number, the network training ends; otherwise, a hidden layer node is added and the process goes to step S41;

[0095] S5. Input the image to be identified into the projection pursuit learning network trained in step S4 for identification and classification. The identification and classification is to standardize the grayscale value of each band of each pixel in the image to be identified, and use it as the input vector of the input node of the projection pursuit learning network, obtain the output vector by calculation, compare the output result of the projection pursuit learning network with each category of expected output value, and classify the pixel into a category with the smallest error.

[0096] As a specific embodiment of the present invention, a specific implementation process of a remote sensing image supervised classification method based on GA optimized PPLN is as follows:

[0097] 1. Select training sample area

[0098] Traditional supervised classification must meet the requirements of calculating variance and covariance matrices in the classification algorithm. Therefore, if the image has N bands, each category generally requires at least 10N training samples. For the PPLN-CRSI method, there is no specific requirement for the number of samples, and traditional small-number group sampling can be used.

[0099] The number of training samples, that is, the number of image classifications, determines the number of output nodes of PPLN. One class corresponds to one node. When the input pattern belongs to a certain class, the output of the node corresponding to this class is 1, otherwise it is 0. Assuming that the pattern is finally divided into 8 classes, the expected output value of class 1 is in the form of: O1(1,0,0,0,0,0,0,0); the expected output value of class 2 is in the form of: O2(0,1,0,0,0,0,0,0,0), and so on. Figure 4 shown.

[0100] 2. Network Training

[0101] The key issues in applying PPLN classification involve network structure design and network learning. First, we need to determine the network structure. The number of nodes in the network input layer is the same as the number of bands of the remote sensing image. Each node corresponds to a band. The number of output nodes of the network is the same as the number of image classifications. The hidden layer uses the Hermite polynomial function as a smoother. The highest power of the polynomial can be determined according to the image classification situation.

[0102] After the network structure is determined, the network is trained using the input and output sample sets to determine the state of the projection pursuit learning network, including the number of hidden layer nodes, polynomial functions, input weights, and output weights, so that the network can achieve a given input-output mapping relationship.

[0103] As an embodiment of the present invention, Figure 1 As shown, here we take the algorithm based on GA to optimize PPLN as an example, and the implementation algorithm is as follows:

[0104] The network groups all parameters around the hidden layer nodes, starts training from the variables related to the first hidden layer node, and trains the variables of each hidden layer node in turn.

[0105] 1. The distribution range of the current node subgroup parameters is determined based on experience. The value range of the input weight and the output weight is between [-1, 1]. The value range of each coefficient of the smoothing function is determined based on experience.

[0106] 2. Select m global evolutionary solutions as the initial values ​​of subgroup parameters through genetic algorithm;

[0107] 1) Randomly generate an initial population in the encoded solution space;

[0108] 2) Evaluate the fitness of each individual in the group, and the fitness objective function is the inverse of the loss function L2;

[0109] 3) Perform genetic operations such as selection, crossover, and mutation on individuals in the group to generate a new generation of groups;

[0110] 4) Evolve repeatedly until the termination condition is met;

[0111] 5) Sort the individuals by quality and select m global evolutionary solutions, avoiding similar individuals when selecting;

[0112] 3. These evolved solutions are used as the initial solutions of the parameters related to the newly added nodes, and the projection pursuit learning network is used to further optimize the subgroup parameters;

[0113] 1) Use Gauss-Newton method to optimize input weights;

[0114] 2) Fitting and approximating the hidden layer smoothing function to obtain optimized polynomial coefficients;

[0115] 3) Repeat steps 1-2 for several iterations;

[0116] 4) Estimate the output weights using the following formula based on the latest input weights and polynomial coefficients;

[0117]

[0118] 5) Repeat steps 1-4 until the relative error of the loss function converges;

[0119] 4. Compare the network loss function L2 obtained by further optimizing the initial solutions of m groups through PPLN, and select the optimized solution with the smallest L2 as the optimal solution for the parameters of the node subgroup.

[0120] 5. If the network's loss function L2 is less than the predetermined target value, or the number of hidden layer nodes reaches the predetermined number, the network training ends; otherwise, add a new hidden layer node and go to step 1.

[0121] Among them, in the process of training all parameters, when adding hidden layer nodes, the genetic algorithm is used to select m global evolutionary solutions, such as Figure 3 As shown, the following steps are included:

[0122] 1. GA Optimization of Initial Values ​​of Input Weights

[0123] The optimization of all parameters of PPLN is centered on the hidden layer nodes and optimized in groups. In each group, the input weights, smoothing functions and output weights are optimized iteratively. The initial values ​​of the three types of parameters are initialized with random numbers. Due to the advantages of genetic algorithms, the basic steps of using genetic algorithms to solve the optimization problem of the initial value of input weights when adding hidden layer nodes are given here.

[0124] In order to solve the minimum value of the input weight, the following expression of the minimization problem is established:

[0125]

[0126] (1) Real number encoding of optimization variables:

[0127] The following linear transformation is used:

[0128] α(j)=a(j)+y(j)(b(j)-a(j)),j=1,2,…,p, (2)

[0129] Where L2 is the optimization objective function, and p is the number of optimization variables. Formula (2) corresponds the jth variable to be optimized α(j) on the initial variable interval [a(j), b(j)] to the real number y(j) on the interval [-1, 1], where y(j) is the genetic gene. At this time, the genes corresponding to all the variables in the optimization problem are sequentially connected to form the encoding form of the solution (y(1), y(2), ..., y(p)), which is called a chromosome. After encoding, the value range of all optimization variables becomes the interval [-1, 1].

[0130] (2) Generation of the initial parent population:

[0131] Assume that the population size is n, generate n groups of uniform random numbers in the interval [-1,1], each group has p numbers, that is, {y(i,j)|i=1,2,…,n;j=1,2,…,p}, substitute y(i,j) into equation (2) to get the optimized variable value α(i,j), and replace the p-dimensional vector α i Normalize to make ||α i ||=1; Substitute α(i,j) into formula (1) to obtain the corresponding objective function value L2(i), sort {L2(i)|i=1,2,…,n} in ascending order, and sort the corresponding individuals {y(i,j)} accordingly. The first few individuals after sorting are called excellent individuals and are allowed to enter the next generation directly;

[0132] (3) Fitness evaluation:

[0133] The smaller the objective function value L2(i), the higher the fitness of the individual, and vice versa. Based on this, the fitness function value F(i) of the i-th parent individual after sorting is defined as:

[0134]

[0135] In the formula, “0.0001” in the denominator is set empirically to avoid the situation where the F(i) value is 0.

[0136] (4) Select:

[0137] The selection operation generates the first offspring population {y1(i,j)|i=1,2,…,n;j=1,2,…,p}. According to the fitness function value, a probability is set for each chromosome y(i,j) in the population, so that the possibility of the chromosome being selected is proportional to the fitness of other chromosomes in the population; the stronger the fitness of the chromosome, the greater the possibility of being selected; using the proportional selection method, the selection probability ps(i) of the parent individual y(i,j) is:

[0138]

[0139] make The sequence {p(i)|i = 1, 2, …, n} divides the interval [0, 1] into n sub-intervals, and these sub-intervals correspond one-to-one with n parental individuals;

[0140] Generate n - 4 random numbers {u(k)|k = 1, 2, …, n - 4}. If u(k) is in (p(i - 1), p(i)], then the i-th individual y(i, j) is selected. In this way, the i-th individual is selected from the parental population y(i, j) with probability ps(i), and a total of n - 4 individuals are selected, and an immigration operation is performed. The four most excellent parental individuals are directly added to the offspring population, y1(i + n - 4, j) = y(i, j), i = 1, 2, …, 4;

[0141] (5) Crossover:

[0142] Perform the crossover operation to generate the second offspring population {y2(i, j)|i = 1, 2, …, n; j = 1, 2, …, p}. First, define the crossover parameter Pc as the probability of the crossover operation. This probability indicates that there are an expected value of PcN chromosomes in the population that will perform the crossover operation. The specific crossover operation is to randomly select a pair of parental individuals y(i1, j) and y(i2, j) as parents according to the selection probability of Equation (4), generate a random number c from (0, 1), and perform the crossover operation between y(i1, j) and y(i2, j) in the following form to generate the following two offspring individuals y2(i1, j) and y2(i2, j):

[0143]

[0144] (6) Mutation:

[0145] Perform the mutation operation to generate the third offspring population {y3(i, j)|i = 1, 2, …, n; j = 1, 2, …, p}. Define the mutation parameter Pm as the mutation probability in the genetic system. This probability indicates that there are an expected value of PmN chromosomes in the population used for the mutation operation. First, sort the parental individuals according to the fitness function value. Then, from i = 1 to N, repeat the following process: Generate a random number r from the interval [0, 1]. If r < Pm, then select the chromosome y(i, j) as the mutated parent, and mutate each selected parent according to the following method;

[0146]

[0147] In the formula, u(j) (j = 1, 2, …, p) is a random number on the interval [a(j), b(j)], x is a random number on (0.8, 1.2). If the generated y3(i, j) exceeds the upper and lower limits of the value range, then directly assign the parental chromosome to the offspring; that is, for individuals with poor fitness, large-scale mutation is adopted; while for individuals with good fitness, small-scale mutation is adopted;

[0148] (7) Evolutionary iteration:

[0149] Sort the 3n offspring individuals generated previously by their fitness function values ​​from large to small, and take the first n offspring individuals as the new parent population; the algorithm goes to step (3) and enters the next round of selection, crossover and mutation calculations until the given end condition is met, stopping the optimization calculation and outputting the optimal input weight α.

[0150] The output input weight α is used as the initial value and substituted into formula (4), and the iterative optimization of the other two subgroup parameters is continued.

[0151] When a hidden layer node is added, the parameters of the two subgroups are initialized with random numbers, and the input weight α is initialized with a genetic algorithm; at the same time, the other two subgroups are optimized with a genetic algorithm. For this purpose, the minimization problem can be modified as follows:

[0152]

[0153] That is, when optimizing the input weight α, the optimization problem of the polynomial coefficient c and the output weight β is considered.

[0154] The steps of optimizing the parameters of the three subgroups using genetic algorithms are the same as above. When real-number encoding is performed on the optimization variables, the genes corresponding to the three types of variables are sequentially connected to form the encoding form of the problem solution (y(1), y(2), ..., y(p)), where p is the number of optimization variables, which is the sum of the dimension of the input weights, the number of polynomial terms, and the output dimension. It is particularly important to note that the three types of variables have different value intervals, the value intervals of the input weights and the output weights are in [-1, 1], and the value interval of the polynomial coefficient c depends on different problems, and the value intervals of each coefficient are different.

[0155] From formula (1), we can see that the solution of the loss function requires the participation of all training samples, that is, the calculation amount of the fitness function of individuals in the genetic algorithm is relatively large, and the number of individuals in the population will increase with the increase of optimization variables, so the population evolution is bound to be slow. Comprehensively balancing the operation time and the target optimization effect, the three sets of parameters obtained by the genetic algorithm may not be the optimal solution to the problem. If the problem has multiple local minima, such as Figure 2 As shown, after adding hidden layer nodes, the genetic algorithm is used to optimize the input weights, polynomial coefficients and output weights. Although the global minimum may not be obtained, the area where the global minimum is located can generally be searched, that is, there will be better individuals falling in the area where the global minimum is located. Then, using it as the initial value, the global optimal solution can be obtained.

[0156] 2. Classification and Identification

[0157] After the network learning is completed, the category of each pixel of the image to be identified is judged. The grayscale value of each band of each pixel in the image is standardized and used as the input vector of the network input node. The output vector can be obtained by calculation. The network output result is compared with the expected output value of each category, and the pixel judgment is classified into the category with the smallest error.

[0158] The classification accuracy of the remote sensing image supervised classification method based on GA optimized PPLN described in the present invention is verified through experiments.

[0159] The hyperspectral remote sensing image used in the experiment is the AVIRIS data that comes with envi, the image name is cup95eff.int, and the image is as follows Figure 5 shown.

[0160] According to the field survey results of the United States Geological Survey (USGS), Figure 6 As shown in the figure, the hyperspectral image is divided into 8 categories, namely (a) alum stone; (b) kaolinite; (c) chalcedony; (d) calcite; (e) muscovite; (f) halloysite; (g) mountain shadow; (h) sodium montmorillonite. The spectral curve is shown in Figure 7 The number of samples in each category is listed in Table 1, and the classification distribution is shown in Figure 8 As shown in Figure (f) in Figure 1. In order to compare the classification results, both the training samples and the evaluation samples were randomly extracted from the total samples using the stratified random sampling tool of Envi. Various classification methods use the same training data, and the sample data for evaluating the accuracy is also consistent.

[0161] Table 1 Categories and sample numbers in hyperspectral experiments

[0162]

[0163]

[0164] The PPLN input layer uses 50 nodes, corresponding to 50 bands, and the output layer uses 8 nodes, corresponding to 8 types of objects. The number of hidden layer nodes is determined by the training situation and is determined according to the threshold setting method. The threshold is set to 0.003, and the number of hidden layer nodes after training is 21. Other main parameters used are: the relative error of the loss function is 0.2, and the polynomial order of the hidden layer smoothing function is 5. The BP algorithm uses 1 hidden layer, the momentum term is 0.9, the initial learning rate is 0.25, the initial weights and offsets are random values ​​within [-0.5, 0.5], the target error is 0.05, and the classification accuracy is highest when the number of hidden nodes is 15. The parameters of the parallelepiped method, the minimum distance method, and the maximum likelihood method all use the default parameters of env i.

[0165] The experiment shows that the overall accuracy and Kappa coefficient of the PPLN-CRSI method are higher than those of the other four classification methods, reaching 96.01% and 0.9543, as shown in Table 2. The error matrix comparison results of the PPLN-CRSI method, the maximum likelihood method and the BP algorithm are shown in Table 3.

[0166] Table 2 Comparison results of five hyperspectral classification methods

[0167]

[0168] Table 3 Error matrix comparison of hyperspectral image classification methods

[0169]

[0170]

[0171] The implementation modes of the present invention are described in detail above in conjunction with the embodiments / drawings, but the present invention is not limited to the above-mentioned implementation modes. For ordinary technicians in this technical field, after knowing the contents recorded in the present invention, they can make several equivalent changes and substitutions thereto without departing from the principle of the present invention, and these equivalent changes and substitutions should also be regarded as belonging to the protection scope of the present invention.

Claims

1. A supervised classification method for remote sensing images based on GA optimized PPLN, characterized by: The steps include: S1, collect remote sensing image samples; S2. Select samples of different modes from the collected samples and construct sample data sets respectively. The number of the sample data sets is the number of image classifications. S3, determining the network structure of the projection pursuit learning network, wherein the number of network input layer nodes of the network structure is the same as the number of bands of the remote sensing image, each node corresponds to one band, the number of network output layer nodes is the same as the number of image categories, and the hidden layers all use Hermite polynomial functions as smoothers, wherein the highest power of the polynomial is determined according to the number of image categories; S4. Training the projection pursuit learning network using the input and output sample data sets and determining the state of the projection pursuit learning network to achieve a given input-output mapping relationship; wherein, The training of all parameters in the projection pursuit learning network is that the projection pursuit learning network groups all parameters with the hidden layer nodes as the center, starts from the input weight and output weight of the first hidden layer node, and sequentially trains the input weight and output weight of each hidden layer node; In the process of training all parameters, when a hidden layer node is added, a genetic algorithm is used to calculate the input weight, the coefficient of the smoothing function and the initial value of the output weight. The calculation process of the initial value includes the following steps: S41, determining the distribution range of the current node subgroup parameter, the value interval of the input weight and the output weight is between [-1, 1]; S42, selecting a number of global evolutionary solutions as initial values ​​of subgroup parameters through a genetic algorithm, the specific selection process includes the following steps: S421. To solve the minimum value of the input weight, establish the minimization problem expression: S422, Real number encoding of optimization variables: The following linear transformation formula is used: α(j)=a(j)+y(j)(b(j)-a(j)),j=1,2,…,p,, (2) Where L2 is the optimization objective function, and p is the number of optimization variables. Formula (2) corresponds the jth variable to be optimized α(j) on the initial variable interval [a(j), b(j)] to the real number y(j) on the interval [-1, 1], where y(j) is the genetic gene. At this time, the genes corresponding to all the variables in the optimization problem are sequentially connected to form the encoding form of the solution (y(1), y(2), ..., y(p)), which is called a chromosome. After encoding, the value range of all optimization variables becomes the interval [-1, 1]. S423. Generation of initial parent population: Assume that the population size is n, generate n groups of uniform random numbers in the interval [-1,1], each group has p numbers, that is, {y(i,j)|i=1,2,…,n;j=1,2,…,p}, substitute y(i,j) into equation (2) to get the optimized variable value α(i,j), and replace the p-dimensional vector α i Normalize to make ||α i ||=1; Substitute α(i,j) into formula (1) to obtain the corresponding objective function value L2(i), sort {L2(i)|i=1,2,…,n} in ascending order, and sort the corresponding individuals {y(i,j)} accordingly. The first few individuals after sorting are called excellent individuals and are allowed to enter the next generation directly; S424, fitness evaluation: The smaller the objective function value L2(i), the higher the fitness of the individual, and vice versa. Based on this definition, the fitness function value F(i) of the i-th parent individual after sorting is: In the formula, "0.0001" in the denominator is set empirically to avoid the situation where the F(i) value is 0; S425, Select: Perform a selection operation to generate the first offspring population {y1(i,j)|i=1,2,…,n;j=1,2,…,p}; According to the fitness function value, a probability is set for each chromosome y(i,j) in the population, such that the possibility of this chromosome being selected is proportional to the fitness of other chromosomes in the population; the stronger the fitness of the chromosome, the greater the possibility of being selected; adopting the proportional selection method, the selection probability ps(i) of the parental individual y(i,j) is as follows: make The sequence {p(i)|i=1,2,…,n} divides the interval [0,1] into n subintervals, which correspond one-to-one to the n parent individuals; Generate n - 4 random numbers {u(k)|k = 1, 2, …, n - 4}. If u(k) is within (p(i - 1), p(i)), then the i-th individual y(i,j) is selected; in this way, the i-th individual is selected from the parental population y(i,j) with probability ps(i), a total of n - 4 individuals are selected, and an immigration operation is performed. The four most excellent parental individuals are directly added to the offspring population, y1(i + n - 4,j) = y(i,j), i = 1, 2, …, 4; S426. Crossover: Perform the crossover operation to generate the second offspring population {y2(i,j)|i = 1, 2, …, n; j = 1, 2, …, p}; first, define the crossover parameter Pc as the probability of the crossover operation. This probability indicates that there are an expected value of PcN chromosomes in the population that will perform the crossover operation; the specific crossover operation is to randomly select a pair of parental individuals y(i1,j) and y(i2,j) as parents according to the selection probability of formula (4), generate a random number c from (0, 1), and perform the crossover operation between y(i1,j) and y(i2,j) in the following form to generate the following two offspring individuals y2(i1,j) and y2(i2,j): S427. Mutation: Perform the mutation operation to generate the third offspring population {y3(i,j)|i = 1, 2, …, n; j = 1, 2, …, p}; define the mutation parameter Pm as the mutation probability in the genetic system. This probability indicates that there are an expected value of PmN chromosomes in the population used for the mutation operation; first, sort the parental individuals according to the fitness function value, and then, from i = 1 to N, repeat the following process: generate a random number r from the interval [0, 1]. If r < Pm, then select the chromosome y(i,j) as the mutated parent, and mutate each selected parent according to the following method; In the formula, u(j)(j = 1, 2, …, p) is a random number on the interval [a(j), b(j)], x is a random number on (0.8, 1.2). If the generated y3(i,j) exceeds the upper and lower limits of the value range, then directly assign the parental chromosome to the offspring; that is, for individuals with poorer fitness, large - scale mutations are adopted; while for individuals with better fitness, small - scale mutations are adopted; S428. Evolutionary iteration: Sort the 3n offspring individuals generated above in descending order according to their fitness function values, and select the first n offspring individuals as the new parental population; the algorithm transfers to S424 to enter the next round of selection, crossover, and mutation calculations, and stops the optimization calculation until the given end condition is met, and outputs the optimal input weight α; S429. Sort the individuals according to their quality, select m globally optimal solutions for evolution, and avoid similar individuals during selection; S43, taking the m evolutionary solutions as initial solutions for the parameters related to the newly added nodes, and further optimizing the subgroup parameters using a projection pursuit learning network; the method for optimizing the subgroup parameters comprises the following steps: S431, optimize input weights using Gauss-Newton method; S432, fitting and approximating the hidden layer smoothing function to obtain optimized polynomial coefficients; S433, repeat steps S431-S432 for several iterations; S434, using the latest input weights and polynomial coefficients, the output weights are estimated using the following formula: S435, repeating steps S431-S434 until the relative error of the loss function converges; S44, comparing the network loss function L2 obtained by further optimizing the m groups of initial solutions through PPLN, and selecting the optimization solution with the smallest L2 as the optimal solution for the node subgroup parameters; S45, if the network loss function L2 is less than the predetermined target value, or the number of hidden layer nodes reaches the predetermined number, the network training ends; otherwise, a hidden layer node is added and the process goes to step S41; S5. Input the image to be identified into the projection pursuit learning network trained in step S4 for identification and classification.

2. According to claim 1, a remote sensing image supervised classification method based on GA optimized PPLN is characterized in that: The states in S4 include the number of hidden layer nodes, polynomial functions, input weights, and output weights.

3. According to claim 1, a remote sensing image supervised classification method based on GA optimized PPLN is characterized in that: The initial value is the global minimum value.

4. The remote sensing image supervised classification method based on GA optimized PPLN according to claim 1 is characterized in that: The recognition classification in S5 is to standardize the grayscale value of each band of each pixel in the image to be recognized, and use it as the input vector of the input node of the projection pursuit learning network, obtain the output vector by calculation, compare the output result of the projection pursuit learning network with each category of expected output values, and classify the pixels into a category with the smallest error.