Hyperspectral Image Band Selection Method Based on Multi-Agent Feature Selection Model

Abstract

The present invention proposes a hyperspectral image band selection method based on a multi-agent feature selection model, which mainly solves the problems of a large action space for hyperspectral band selection, random exploration directions, and long running times. The implementation steps are as follows: generating a training set; constructing and training an evaluation classification network; introducing a teacher for guidance during the construction and training of the multi-agent feature selection model; and performing band selection using the trained multi-agent band selection model. The present invention uses multi-agents for band selection, which can reduce the action space to two dimensions, introduces a teacher to guide the agents' selections, guides the agents' exploration towards a better direction, and designs a pre-trained evaluation classification network to evaluate different band sets, thereby reducing the overall running time of the algorithm.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and further relates to a hyperspectral image band selection method based on a multi-agent feature selection model in the technical fields of feature extraction and analysis. By selecting bands in the hyperspectral image, the present invention provides necessary information support for hyperspectral target classification in fields such as resource investigation and application, environmental monitoring and evaluation, and regional analysis and planning. Background Art

[0002] Hyperspectral imaging can obtain the luminance values of the electromagnetic waves emitted or reflected by an object in each band, and the obtained more refined and specific spectral data provides great advantages for target positioning and recognition. The hyperspectral sensor simultaneously images the target area with dozens to hundreds of continuous and subdivided spectral bands, forming a hyperspectral image containing dozens to hundreds of bands. Due to the rich spectral and spatial information, the hyperspectral target detection technology can break through the morphological features of the target. Therefore, hyperspectral images are widely used in various tasks, such as fine classification of ground objects, target detection, and vegetation area estimation in complex environments.

[0003] Beihang University proposed a hyperspectral image band selection method based on an unsupervised reinforcement learning band selection model in its patent document "A Hyperspectral Image Band Selection Method Based on Reinforcement Learning" (Patent Application No.: 202111268014.8, Publication No.: CN 113936219 A). This method uses the trial-and-error learning and the ability to consider long-term benefits of reinforcement learning to solve the problem of limited spatial search and decision optimization in existing band selection methods, and introduces a greedy strategy to achieve the balance between exploration and application in the decision-making process of the agent. The model is trained in an unsupervised manner using unlabeled samples to solve the problem of insufficient labeled samples. However, the still existing deficiencies of this method are that a single agent is used to select all bands, resulting in a problem of too large action space due to the large number of selectable combinations for the agent. In addition, due to the use of a large number of unlabeled samples for training, the exploration direction of the agent is random, resulting in a poor classification effect of the finally selected band set.

[0004] Northwestern Polytechnical University proposed a hyperspectral image classification method based on a deep neural network in its patent document "A Hyperspectral Image Classification Method Based on a Deep Neural Network" (Patent Application No.: 202211095575.7, Publication No.: CN 115424138 A). This method proposed a strategy for expanding labeled hyperspectral training data. For the input hyperspectral dataset, a collaborative selection strategy and a competitive selection strategy were used to collaboratively analyze the confidence evaluation method to provide more available recognition training data for training, so as to train a better deep neural network to improve the classification performance of the hyperspectral data to be classified. However, the deficiency of this method is that for the hyperspectral dataset, if different bands are selected as the input, the deep neural network model needs to be retrained, which greatly increases the running time. Summary of the Invention

[0005] The purpose of the present invention is to propose a hyperspectral image band selection method based on a multi-agent feature selection model in view of the deficiencies of the above-mentioned existing technologies, so as to solve the technical problems of large agent action space, random exploration direction and long running time existing in the existing methods of hyperspectral image band selection and classification.

[0006] The idea of implementing the present invention is that the present invention constructs a multi-agent feature selection model to generate agents equal in number to the number of bands. Each agent only determines the selection of its corresponding band, thus greatly reducing the action space that the agent can select and overcoming the problem of large action space in the prior art. In the process of training multi-agents, according to the classification accuracy of the selected bands for the classification task, three models with better classification accuracy of the selected bands are introduced as teachers to guide the bands selected by the agents. This way introduces better external knowledge and can guide the agents to explore in a better direction, overcoming the problem of random exploration direction of agents in the prior art. The present invention uses a pre-trained evaluation classification network to evaluate the bands selected by the agents, and uses a training method of multiple random samplings, so that the network can effectively evaluate different band sets without adjusting parameters, overcoming the problem that the existing evaluation classification network needs to be retrained for different band sets, resulting in an increase in the overall running time of the model.

[0007] The technical solution adopted by the present invention includes the following steps:

[0008] Step 1, generate a training set:

[0009] Step 1.1, input a hyperspectral image with at least 3 bands, each band having at least 25*25 pixels, and labeled with a true label vector;

[0010] Step 1.2: Taking each pixel in the hyperspectral image as the center, define a spatial window of 25×25 pixels; form a data cube with all the pixels within each spatial window; and form a sample set of the hyperspectral image with all the data cubes in the hyperspectral image.

[0011] Step 1.3: Randomly select 5% of the samples in the sample set of the hyperspectral image to form the training set of the hyperspectral image.

[0012] Step 2: Construct a multi-agent feature selection model:

[0013] Generate a multi-agent feature selection model composed of N neural networks with the same structure. Each neural network represents an agent, and each agent represents a band in the hyperspectral image. The structure of the neural network is: the first fully connected layer, the second fully connected layer, and the third fully connected layer. Set the number of nodes in the first to third fully connected layers to 16, 100, and 2 respectively. Here, the value of N is equal to the number of bands in the input hyperspectral image.

[0014] Step 3: Train and evaluate the classification network:

[0015] Input the training set into the evaluation classification network to obtain the predicted label vectors of all samples in the training set; use the cross-entropy formula to calculate the cross-entropy between the predicted label vectors and the true label vectors; adopt the gradient descent method to iteratively update the parameters of the evaluation classification network until the network loss function converges, and obtain the trained evaluation classification network.

[0016] Step 4: Train each agent band selection model:

[0017] Step 4.1: Generate a vector with the same length as the number of bands in the input hyperspectral image. Each element in the vector corresponds to a band, and the element values in the initial vector are all 0. After the initial vector is input into each agent respectively, the elements corresponding to the bands selected by the agent are set to 1, thus forming the state vector s at the current moment.

[0018] Step 4.2: Input the state vector s at the current moment into each agent respectively to obtain the agent action a at the current moment. Combine the state vector s and the agent action a at the current moment to form the current agent band set, and calculate the classification accuracy of the current agent band set.

[0019] Step 4.3, using the classification accuracy of the current agent band set as the standard, select the three models with high classification accuracy selected by the band selection model as teachers, and perform interactive learning between the teachers and the agent. The teachers will optimize the band set selected by the agent to obtain the optimized current agent band set, and update the state vector before optimization with the optimized current agent band set to obtain the state vector s′ at the next moment;

[0020] Step 4.4, input the updated agent band set into the trained evaluation classification network, and output the predicted label vector of this set;

[0021] Step 4.5, calculate the classification accuracy of the predicted label vector;

[0022] Step 4.6, use the Q-learning formula to update each agent parameter in the multi-agent feature selection model;

[0023] Step 4.6, repeat Steps 4.2 to 4.6 a total of 90 times to obtain a trained multi-agent feature selection model;

[0024] Step 5, select the bands of the hyperspectral image:

[0025] Step 5.1, use the same method as in Step 1.2 to process the hyperspectral image of the bands to be selected to obtain a sample set;

[0026] Step 5.2, input the sample set into the trained multi-agent feature selection model, and output the selected band set.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] First, since the present invention constructs a multi-agent feature selection model and introduces three models with good band selection effects to guide the bands selected by the agent, it overcomes the problem of random exploration directions caused by large action spaces and random model initialization in the prior art, resulting in poor classification results, so that the present invention can select a band combination with higher classification accuracy.

[0029] Second, since the present invention uses a pre-trained evaluation classification network, without adjusting the network parameters when evaluating different band sets on the premise of ensuring the classification effect, it overcomes the problem in the prior art that the evaluation classification network needs to be retrained for different band sets, resulting in an increase in the overall running time of the model. Therefore, the present invention can evaluate any band set in a short time, reducing the calculation amount and accelerating the model running time. Description of the Drawings

[0030] Figure 1 is the flowchart of the present invention;

[0031] Figure 2 This is the simulation diagram of the present invention. Specific implementation manners

[0032] The present invention will be further described below with reference to the accompanying drawings.

[0033] Refer to Figure 1 to further describe the implementation steps of the embodiments of the present invention.

[0034] Step 1. An input hyperspectral image with a size of 144×144×200 and containing a true label vector is input in the embodiment of the present invention. This hyperspectral image contains 200 bands.

[0035] Step 2. Generate a sample set.

[0036] Taking each pixel in the input hyperspectral image as the center, a spatial window with a size of 27×27 pixels is delimited. All the pixels within each spatial window are combined to form a data cube. All the data cubes are combined to form the sample set of the hyperspectral image.

[0037] Step 3. Generate a training sample set and a test sample set.

[0038] In the sample set of the hyperspectral image, 5% of the samples are randomly selected to form the training sample set of the hyperspectral image. The remaining 95% of the samples are combined to form the test sample set of the hyperspectral image.

[0039] Step 4. Construct an evaluation classification network.

[0040] In the embodiment of the present invention, a 7-layer evaluation classification network of the existing technology is adopted, and its structure is successively: the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the first fully connected layer, the second fully connected layer, and the third fully connected layer.

[0041] Set the parameters of each layer;

[0042] Set the convolutional kernel size of the first convolutional layer to 4×4, the number to 64, and the convolutional stride to 1.

[0043] The convolutional kernel size of the first convolutional layer is set to 7x3x3, the number is set to 8, and the convolutional stride is set to 1.

[0044] The convolutional kernel size of the second convolutional layer is set to 5x3x3, the number is set to 16, and the convolutional stride is set to 1.

[0045] The convolutional kernel size of the third convolutional layer is set to 3x3x3, the number is set to 32, and the convolutional stride is set to 1.

[0046] The convolutional kernel size of the fourth convolutional layer is set to 3x3, the number is set to 64, and the convolutional stride is set to 1.

[0047] The nodes of the first, second, and third fully connected layers are set to 256, 128, and 16 respectively.

[0048] Step 5. Train and evaluate the classification network.

[0049] Step 1, generate a set of candidate bands uniformly distributed in the interval [0, 200]. The set of candidate bands contains 2000 elements

[0050] Step 2, select the first element n from the set of candidate band numbers. The value of element n is b. Randomly select b bands from the training sample set to form a new training sample set.

[0051] Step 3, input the new training sample set into the convolutional neural network, perform convolutional and fully connected operations in sequence, and output a predicted label vector of size 27×27×16.

[0052] Step 4, use the following cross-entropy formula to calculate the cross-entropy between the predicted label vector and its corresponding true label vector, and obtain the loss value of the evaluation classification network.

[0053] The cross-entropy formula is as follows:

[0054]

[0055] Among them, L represents the cross-entropy between the predicted label vector and the true label vector, Σ represents the summation operation, y i represents the i-th element in the predicted label vector, ln represents the logarithmic operation with the natural constant e as the base, represents the m-th element in the predicted label vector.

[0056] Step 5, use the gradient descent method to optimize the network parameters with the loss function of the evaluation classification network until the network parameters converge, and remove element n from the set of candidate band numbers.

[0057] Step 5, continue to execute the second step until the set of candidate band numbers is an empty set, and obtain the trained evaluation classification network.

[0058] Step 6. Establish a multi-agent feature selection model.

[0059] Generate a multi-agent feature selection model composed of 200 neural networks with the same structure. Each neural network represents an agent, and each agent represents a band in the hyperspectral image. The structure of each neural network is: the first fully connected layer, the second fully connected layer, and the third fully connected layer; the number of nodes in the first to third fully connected layers are set to 16, 100, and 2 respectively.

[0060] The mathematical expression of the multi-agent feature selection model is: a Markov decision process <S, A, R, P, γ>, and the meaning of each parameter is as follows:

[0061] State space S: It is defined as a 200-dimensional binary encoding vector of the currently selected band set, and each dimension represents a band. Among them, each band has two states, 1 means selected, and 0 means not selected.

[0062] Action space A: The action a in it n represents the action that the nth intelligent agent can choose, and there are two possibilities, namely 1 to select a band and 0 not to select a band.

[0063] State transition function P: It represents the distribution of the state transferred from the current moment to the next moment.

[0064] Reward R: The reward function r is set to the accuracy of the classification of the selected bands.

[0065] Reward discount factor γ: γ ∈ [0, 1] represents the discount coefficient for future benefits. The closer it is to 1, the more important the future benefits are.

[0066] Step 7. Train the multi-agent feature selection model.

[0067] Step 1, generate a vector with the same length as the number of bands in the input hyperspectral image. Each element in the vector corresponds to a band, and the element values in the initial vector are all 0; after the initial vector is input into each intelligent agent respectively, the elements corresponding to the bands selected by the intelligent agent are set to 1, so as to form the state vector s.

[0068] Step 2, input the state vector s into the constructed multi-agent model (the initial state is 0) to obtain the set of bands selected by the intelligent agent.

[0069] Step 3, select three models with better band selection effects, namely the self-improving convolutional neural network SICNN (self-improving convolutional neural network), the band selection network BS-Net (band selection network), and the ternary weight convolutional neural network TWCNN (ternary weight convolutional neural network) in the existing technology as teachers, and interactively learn (Interactive learning) with each intelligent agent in three different stages.

[0070] The way of interactive learning is as follows: Prepare the band set A1 selected by the agent last time, where a total of n1 bands are selected. The band set A2 selected in this round, where n2 bands are selected. Input the bands selected in the band set A1 but not selected in the band set A2 into the teacher, and the teacher selects k bands from them, where Compare the k bands selected by the teacher with the band set selected by the agent. If there is a band selected by the teacher but not selected by the agent, change the state of this band to the selected state. In this way, the optimized band set selected by the current agent is obtained, where b bands are selected. Update the state vector before optimization with the optimized current agent band set to obtain the state vector s' at the next moment.

[0071] Step 4, evaluate the band combination selected by the current agent. Form a new hyperspectral data aggregation according to the selected bands, and generate a sample set according to the method in Step 2. Input the sample generated from the hyperspectral image with a size of 27×27×b pixels into the convolutional neural network, and pass through the convolutional and fully connected layers in turn to obtain a prediction label vector with a size of 27×27×16. By comparing the real label vector, the classification accuracy is obtained, and this is used as the reward r.

[0072] The classification accuracy is obtained by the following formula:

[0073]

[0074] Step 5, use the following Q-learning formula to update each agent parameter in the multi-agent feature selection model:

[0075] L(θ i )=E[(r+γmax(Q(s',a';θ i )-Q(s,a;θ i ))) 2

[0076] where L(θ i ) represents the updated parameters of the three fully connected layers in the i-th agent of the multi-agent feature selection model, E() represents the calculation of expectation, r represents the classification accuracy of the current agent band set, γ represents a hyperparameter value selected in the range of [0,1], Q represents the state value function, s and a respectively represent the state vector and action at the current moment, and s' and a' represent the state vector and action at the next moment.

[0077] Step 6, repeat Steps 2 to 5 of this step 90 times to obtain a trained multi-agent feature selection model

[0078] Step 8. Select the bands of the hyperspectral image

[0079] ​Input the test sample set into the trained multi-agent feature selection model, and output the selected band set.

[0080] The following further illustrates the effect of the present invention in combination with simulation experiments:

[0081] 1. Simulation experiment conditions:

[0082] The hardware platform for the simulation experiment of the present invention is: the processor is an Intel i7 7820X CPU with a main frequency of 3.6 GHz and a memory of 64 GB.

[0083] The software platform for the simulation experiment of the present invention is: Windows 10 operating system and python 3.8

[0084] The input image used in the simulation experiment of the present invention is the Indian Pines hyperspectral image. The hyperspectral data was collected from the Indian Remote Sensing Test Area in northwestern Indiana, USA. The imaging time was in June 1992. The image size is 145×145×200 pixels. The image contains 200 bands and 16 types of ground objects in total, and the image format is mat.

[0085] 2. Simulation content and its result analysis:

[0086] In the simulation experiment of the present invention, the present invention and three existing technologies (attention mechanism ABCNN band selection method, self-improved SICNN band selection method, support vector machine SVM band selection method) are respectively used to select the bands in the input Indian Pines hyperspectral image, and the obtained selected band results are as Figure 2 shown.

[0087] In the simulation experiment, the three existing technologies adopted refer to:

[0088] The attention mechanism ABCNN band selection method refers to the hyperspectral image band selection method proposed by Lorenzo et al. in "Hyperspectral Band Selection Using Attention-Based Convolutional Neural Networks, IEEE Access. 8(2020)42384-42403.", abbreviated as the attention mechanism ABCNN band selection method.

[0089] The self-improving SICNN band selection method refers to the hyperspectral band selection method proposed by Ghamisi et al. in "A Self-Improving Convolution Neural Network for the Classification of Hyperspectral Data, IEEE Geosci. Remote. Sens. Lett. 13(10)(2016)1537-1541.", which is abbreviated as the self-improving SICNN band selection method.

[0090] The support vector machine SVM band selection method refers to the hyperspectral band selection method proposed by Guyon et al. in "Gene Selection for Cancer Classification using Support Vector Machines, Mach. Learn. 46(2002)389-422.", which is abbreviated as the support vector machine SVM band selection method.

[0091] The following combines Figure 2 the simulation diagrams to further describe the effect of the present invention.

[0092] Figure 2 (a) is a false color image composed of the 50th, 27th, and 17th bands among the 200 bands of the hyperspectral image selected for the simulation experiment of the present invention. Figure 2 (b) is the true ground object distribution map of the Indian Pines hyperspectral image selected for the simulation experiment of the present invention, and the size of this distribution map is 145×145 pixels. Figure 2 (c) is the result diagram of classifying the Indian Pines hyperspectral image by using the attention mechanism ABCNN band selection method of the prior art. Figure 2 (d) is the result diagram of classifying the Indian Pines hyperspectral image by using the self-improving SICNN band selection method of the prior art. Figure 2 (e) is the result diagram of classifying the Indian Pines hyperspectral image by using the support vector machine SVM band selection method of the prior art. Figure 2 (f) is the result diagram of classifying the Indian Pines hyperspectral image by using the method of the present invention.

[0093] From Figure 2(c) It can be seen that compared with the classification results of the attention mechanism ABCNN band selection method of the prior art, more objects are misclassified by the self-improved SICNN band selection method of the prior art. This is mainly because the band selection and classification processes of this method are separated, which results in the selected bands not being able to complete the classification task well.

[0094] From Figure 2 (d) It can be seen that for the self-improved SICNN band selection method of the prior art, compared with the classification results of the support vector machine SVM, it has fewer noise points. However, for this classification method when combined with different bands, the network needs to be continuously trained.

[0095] From Figure 2 (e) It can be seen that compared with the classification results of the three prior arts, the classification results of the present invention have fewer misclassified and undetected samples, and have better regional consistency and edge smoothness. This proves that the classification effect of the present invention is better than the classification methods of the first three prior arts, and the classification effect of the selected bands is relatively ideal.

[0096] The classification results of the four methods are evaluated respectively using three evaluation indicators (classification accuracy for each class, overall accuracy OA, average accuracy AA).

[0097] Using the following formulas, calculate the overall accuracy OA, average accuracy AA, and classification accuracy of 16 types of ground objects. All calculation results are plotted in Table 1.

[0098]

[0099]

[0100]

[0101] Table 1. Quantitative analysis table of the classification results of the present invention and each prior art in the simulation experiment

[0102]

[0103] Combined with Table 1, it can be seen that the overall classification accuracy OA of the present invention is 97.8%, and the average classification accuracy AA is 97.5%. These two indicators are higher than the three prior art methods, which proves that the present invention can select a better band set.

[0104] The above simulation experiments show that:

[0105] The method of the present invention uses a hyperspectral image band selection method based on a multi-agent feature selection model, which can select a set of bands with better classification effect in a limited time, solves the problems in the prior art methods that the search space is large, it is difficult for the agents generated by random initialization to search for the optimal iterative direction, and at the same time, the network used to evaluate the band set usually needs to adjust the network model for different band sets, resulting in poor classification effect and high time consumption of the finally selected band set.

Claims

1. A hyperspectral image band selection method based on a multi-agent feature selection model, characterized in that Construct a multi-agent feature selection model, and use the teacher and the trained evaluation classification network to train the selection model respectively; the specific steps of this method are as follows: Step 1, generate a training set: Step 1.1, input a hyperspectral image with at least 3 bands, each band containing at least 25*25 pixels, and labeled with a true label vector; Step 1.2, centered on each pixel in the hyperspectral image, delimit a 25×25 pixel-sized spatial window; form a data cube by all the pixels within each spatial window; form a sample set of the hyperspectral image by all the data cubes in the hyperspectral image; Step 1.3, randomly select 5% of the samples in the sample set of the hyperspectral image to form the training set of the hyperspectral image; Step 2, construct a multi-agent feature selection model: Generate a multi-agent feature selection model composed of N neural networks with the same structure. Each neural network represents an agent, and each agent represents a band in the hyperspectral image; the structure of each neural network is: the first fully connected layer, the second fully connected layer, and the third fully connected layer; set the number of nodes in the first to third fully connected layers to 16, 100, and 2 respectively; where the value of N is equal to the number of bands in the input hyperspectral image; Step 3, train the evaluation classification network: Input the training set into the evaluation classification network to obtain the predicted label vectors of all samples in the training set; use the cross-entropy formula to calculate the cross-entropy between the predicted label vectors and the true label vectors; adopt the gradient descent method to iteratively update the parameters of the evaluation classification network until the network loss function converges, and obtain the trained evaluation classification network; Step 4, train each agent band selection model: Step 4.1, generate a vector with the same length as the number of bands in the input hyperspectral image. Each element in the vector corresponds to a band, and the element values in the initial vector are all 0; after the initial vector is input into each agent respectively, the element corresponding to the band selected by the agent is set to 1, so as to form the state vector s at the current moment; Step 4.2, input the state vector s at the current moment into each agent respectively to obtain the agent action a at the current moment. Combine the state vector s and the agent action a at the current moment to form the current agent band set, and calculate the classification accuracy of the current agent band set; Step 4.3, taking the classification accuracy of the current agent band set as the standard, use the models with high classification accuracy of the three bands selected by the band selection model as teachers, and perform interactive learning between the teachers and the agents. The teachers will optimize the band set selected by the agents to obtain the optimized current agent band set, and update the state vector before optimization with the optimized current agent band set to obtain the state vector s' at the next moment; Step 4.4, input the updated agent band set into the trained evaluation classification network to output the predicted label vector of this set; Step 4.5, calculate the classification accuracy of the predicted label vector; Step 4.6, use the Q-learning formula to update the parameters of each agent in the multi-agent feature selection model; Step 4.7: Repeat steps 4.2 to 4.6 90 times to obtain the trained multi-agent feature selection model. Step 5: Select the band of the hyperspectral image: Step 5.1: Process the hyperspectral image of the selected band in the same way as step 1.2 to obtain a sample set; In step 5.2, the sample set is input into the trained multi-agent feature selection model and the selected band set is output.

2. The hyperspectral image band selection method based on the multi-agent feature selection model according to claim 1, wherein The cross entropy formula described in step 3 is as follows: Among them, L represents the cross-entropy between the predicted label vector and the true label vector, Σ represents the summation operation, y i represents the i-th element in the predicted label vector, ln represents the logarithmic operation with the natural constant e as the base, represents the m-th element in the predicted label vector.

3. The hyperspectral image band selection method based on the multi-agent feature selection model according to claim 1, characterized in that The Q-learning formula described in step 4.6 is as follows: L(θ i )=E[(r+γmax(Q(s',a';θ i )-Q(s,a;θ i ))) 2 ] Among them, L(θ i ) represents the updated parameters of the three fully connected layers in the i-th agent of the multi-agent feature selection model, E() represents the expected calculation, r represents the classification accuracy of the current agent band set, γ represents a hyperparameter value selected in the range of [0,1], Q represents the state value function, s, a represent the state vector and action at the current moment, respectively, and s', a' represent the state vector and action at the next moment.

Images