A Hyperspectral Image Classification Method Based on Evolutionary Multi-Task Ensemble Learning

Abstract

This invention discloses a hyperspectral image classification method based on evolutionary multi-task ensemble learning, comprising: pre-training each base classifier using labeled samples as a training set; inputting unlabeled samples to be classified into the pre-trained base classifiers for classification; expanding the samples using a sample augmentation strategy based on the classification results of the unlabeled samples to be classified in the base classifiers, and applying weighted mean filtering to the remaining unlabeled samples to be classified; establishing feature selection tasks for each base classifier based on the expanded training set, and selecting the target optimal feature subspace for each feature selection task using a designed multi-task evolutionary optimization method; obtaining a low-dimensional training set and low-dimensional unlabeled samples to be classified based on the target optimal feature subspace set; retraining the pre-trained base classifiers based on the low-dimensional training set, using the trained base classifiers to classify the low-dimensional unlabeled samples to be classified, and ensemble the classification results. This invention improves the classification performance of hyperspectral images.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral image processing technology, specifically relating to a hyperspectral image classification method based on evolutionary multi-task ensemble learning. Background Technology

[0002] In the early 1980s, the development of a new generation of sensor imaging spectrometers officially began. Since then, hyperspectral imaging and sensor technology have entered a period of rapid development, and the concept of spectral imaging has come into the researchers' view. Because hyperspectral images have high spectral resolution and can provide detailed spectral features, this has brought new opportunities for analyzing and studying land cover categories.

[0003] In practical applications, the high-dimensional data structure of hyperspectral images poses challenges to remote sensing image processing. With a limited number of labeled samples, as the data becomes more complex, the accuracy of land cover classification decreases with the increase in the number of features, resulting in the Hughes phenomenon. Considering this phenomenon, many feature selection-based hyperspectral image classification algorithms have been widely applied in recent years for land cover classification research in hyperspectral images, achieving good classification results. When processing hyperspectral remote sensing images using feature selection-based classification methods, it is first necessary to select some excellent features from high-dimensional candidate features to form a feature subspace, i.e., solving the problem of combinatorial optimization of spectral features in hyperspectral images. Then, the feature subspace is used for training and classification of base classifiers. For example, Zhang et al. designed a multi-objective optimization model based on the relationship between the number of selected features and the amount of information retained to select the optimal feature subspace set with the least redundancy and the highest information content suitable for the classification task. Wan et al. proposed a multi-objective feature selection method for hyperspectral images based on the discrete sine-cosine algorithm to enhance the selection of the ideal feature subspace set.

[0004] However, when using traditional feature selection-based classification methods to solve many practical problems, a single base classifier exhibits bias and poor generalization ability. To address the issue of varying biases among different base classifiers during feature selection, feature selection-based hyperspectral image classification methods employ ensemble learning algorithms to seek classification performance with better generalization and robustness. For example, ensemble classification methods based on random feature subspaces train multiple base classifiers using randomly generated feature subspaces. By integrating the dominant base classifiers, more efficient and robust classification results can be obtained than with a single base classifier. Although these methods strive to maintain important spectral information, remove redundant information, and reduce computational burden, the correlation between multiple search processes is rarely noticed because the search for the optimal feature subspace by different base classifiers is independent. Therefore, when ensemble-based hyperspectral image classification methods perform the task of optimizing multiple spectral feature subspaces, they are inevitably susceptible to interference from redundant features and repeated searches for important features. This makes it difficult for the search for the optimal spectral feature subspace to converge to the global optimum, thus affecting the final classification result of the hyperspectral remote sensing image.

[0005] Therefore, the ability to search for the globally optimal spectral feature subspace is a crucial issue for hyperspectral image classification methods based on ensemble learning. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides a hyperspectral image classification method based on evolutionary multi-task ensemble learning. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] This invention provides a hyperspectral image classification method based on evolutionary multi-task ensemble learning, comprising:

[0008] Input hyperspectral image data; the hyperspectral image data contains a small number of labeled samples, and the remaining samples are unlabeled samples to be classified.

[0009] The labeled samples are used as the training set to pre-train each base classifier, and the unlabeled samples to be classified are sequentially input into the pre-trained base classifiers for classification.

[0010] Based on the classification results of the unlabeled samples to be classified in each base classifier, a sample augmentation strategy is used to select a portion of the unlabeled samples to be classified with high confidence to augment the training samples of the training set. Spatial location information is incorporated into the remaining unlabeled samples to be classified through weighted mean filtering.

[0011] Based on the training set expanded with training samples, feature selection tasks are established for each base classifier. The designed multi-task evolutionary optimization method is used to select the target optimal feature subspace for each feature selection task.

[0012] Based on the target optimal feature subspace selected for each feature selection task, a low-dimensional training set is obtained after mapping the training set expanded by the training samples, and a low-dimensional unlabeled sample to be classified is obtained after mapping the unlabeled sample to be classified by weighted mean filtering. Based on the low-dimensional training set, each pre-trained base classifier is trained again, and the trained base classifiers are used to classify the low-dimensional unlabeled sample to be classified, resulting in multiple classification results.

[0013] The final classification result of the hyperspectral image data is obtained by integrating the multiple classification results.

[0014] In one embodiment of the present invention, the process of selecting a portion of the unlabeled samples to be classified with high confidence to expand the training samples of the training set according to the classification results of the unlabeled samples to be classified in each base classifier includes:

[0015] For each unlabeled sample to be classified, the following process is performed: the unlabeled sample to be classified is sequentially input into each pre-trained base classifier to obtain the classification result; the unlabeled sample to be classified that obtains a consistent classification result in all base classifiers is assigned a pseudo-label and added to the training set; wherein, before the next unlabeled sample to be classified is input into each pre-trained base classifier, each base classifier is pre-trained again using a new training set.

[0016] Once all unlabeled samples to be classified have gone through the above process, the training sample expansion process is complete.

[0017] In one embodiment of the present invention, a feature selection task for each base classifier is established based on a training set augmented with training samples. A designed multi-task evolutionary optimization method is used to select the optimal target feature subspace for each feature selection task, including:

[0018] For each base classifier, the optimal feature subspace is selected based on the training set expanded by the training samples, and the process of selecting the optimal feature subspace is modeled as the corresponding feature selection task.

[0019] Based on the modeling-based feature selection task, a multi-task evolutionary optimization method is used to select the optimal feature subspace for each feature selection task, including:

[0020] The designed multi-task evolutionary optimization method is used to optimize the population for tasks, thereby selecting the optimal feature subspace for each feature selection task; wherein, each individual in the population represents a feature subspace.

[0021] In one embodiment of the present invention, a designed multi-task evolutionary optimization method is used to perform task optimization on the population, thereby selecting the optimal feature subspace for each feature selection task, including:

[0022] Several feature subspaces are randomly generated to form a population of individuals;

[0023] The fitness value of each individual in the population is calculated on all feature selection tasks. The skill factor of each individual in the population is calculated based on the fitness value of the population to obtain the parent population and the initial optimal feature subspace set.

[0024] The offspring population is obtained by performing crossover and mutation operations on the parent population.

[0025] For each individual in the offspring population, calculate the fitness value of that individual on the task of selecting the corresponding feature based on that individual's skill factor;

[0026] The parent population and the offspring population are merged to obtain an intermediate population. The initial optimal feature subspace set is updated according to the skill factor and fitness value of the intermediate population to obtain the intermediate optimal feature subspace set for all feature selection tasks.

[0027] The target population is generated from the intermediate population using an elite selection strategy. The target population is used as the parent population, and the intermediate optimal feature subspace set is used as the initial optimal feature subspace set. The above update process continues until the maximum number of evolutions is reached. The intermediate optimal feature subspace set selected by all feature selection tasks is used as the target optimal feature subspace set.

[0028] In one embodiment of the present invention, the formula for calculating the fitness value is as follows:

[0029] ;

[0030] in, Represents an individual In feature selection tasks The fitness value of the population is used in the fitness value calculation. Values ​​range from 1 to T , T This indicates the number of feature selection tasks, and is used in the calculation of the fitness value of the offspring population. Value is an individual The corresponding skill factors, and They represent In feature selection tasks The number of samples correctly classified on the corresponding base classifier and Total number of samples This represents the subset obtained by partitioning the training set, which is expanded from the training samples, and used to calculate the fitness value of each individual.

[0031] In one embodiment of the present invention, calculating the skill factor of each individual in the population based on the fitness value of the population includes:

[0032] The fitness values ​​of each individual in the population on all feature selection tasks are sorted, and the index value corresponding to the minimum value in the sorting results is taken as the skill factor of that individual.

[0033] In one embodiment of the present invention, crossover and mutation operations are performed on the parent population to obtain the offspring population, including:

[0034] When two randomly selected parent individuals in the parent population have the same skill factor, the crossover operation condition is met, and the crossover operation is performed on the two parent individuals using a difference-based crossover operator.

[0035] When two randomly selected parent individuals from the parent population possess different skill factors, a determination is made based on a predefined crossover probability to determine whether to perform a crossover operation on the two parent individuals. This includes: generating a random number; if the random number is less than or equal to the predefined crossover probability, satisfying the crossover operation condition, and performing a crossover operation on the two parent individuals using a difference-based crossover operator; if the random number is greater than the predefined crossover probability, not satisfying the crossover operation condition, and performing a mutation operation on the individuals in the parent population using a constrained single-point mutation operator.

[0036] In one embodiment of the present invention, a crossover operation is performed on two parent individuals using a difference-based crossover operator, including:

[0037] Calculate the feature difference between two randomly selected parent individuals from the parent population.

[0038] Crossing is performed on two randomly selected parent individuals using the crossover operator and the calculated feature difference set.

[0039] In one embodiment of the present invention, a constrained single-point mutation operator is used to perform mutation operations on individuals in the parent population, including:

[0040] The individuals in the parent population are mutated according to the mutation probability, and the number of chromosomes with gene mutations in each mutated individual is counted.

[0041] For each individual after mutation, select chromosome positions equal to the number of chromosome positions from the remaining chromosome positions that have not undergone gene mutation and perform reverse mutation.

[0042] In one embodiment of the present invention, the final classification result of the hyperspectral image data is obtained by integrating multiple classification results, including:

[0043] The final classification result of the hyperspectral image data is obtained by integrating the multiple classification results using an ensemble strategy that combines majority voting and a high-precision classifier.

[0044] The beneficial effects of this invention are:

[0045] The hyperspectral image classification method proposed in this invention, based on evolutionary multi-task ensemble learning, is a classification method that combines multi-task ensemble and sample augmentation strategies. Specifically: Input hyperspectral image data; the hyperspectral image data contains a small number of labeled samples, with the remaining samples being unlabeled samples to be classified; the labeled samples are used as the training set to pre-train each base classifier; the unlabeled samples to be classified are sequentially input into each pre-trained base classifier for classification; based on the classification results of the unlabeled samples to be classified in each base classifier, a sample augmentation strategy is used to select a portion of the unlabeled samples with high confidence to augment the training set; the remaining unlabeled samples to be classified are then incorporated into the spatial classification using a weighted mean filter. Location information; based on the training set expanded with training samples, feature selection tasks are established for each base classifier. A designed multi-task evolutionary optimization method is used to select the target optimal feature subspace for each feature selection task. Based on the target optimal feature subspace selected for each feature selection task, a low-dimensional training set is obtained after mapping the training set expanded with training samples, and a low-dimensional unlabeled sample to be classified is obtained after mapping the unlabeled sample to be classified using weighted mean filtering. The pre-trained base classifiers are then trained again based on the low-dimensional training set. The trained base classifiers are then used to classify the low-dimensional unlabeled sample to be classified, resulting in multiple classification results. The final classification result of the hyperspectral image data is obtained by integrating these multiple classification results. It can be seen that this invention, based on a multi-task integration and sample expansion strategy, first expands the limited number of labeled samples. Then, based on the expanded labeled samples, multiple base classifiers are used in parallel to generate a target optimal feature subspace set, achieving feature sharing among the base classifiers. They also assist each other in the optimal feature search process to find the globally optimal feature subspace. Finally, by integrating the classification results of multiple base classifiers, the performance of hyperspectral image classification is improved. In addition, this classification method minimizes noise in the classification map and has a certain degree of robustness.

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a hyperspectral image classification method based on evolutionary multi-task ensemble learning provided in an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of the framework of the hyperspectral image classification method based on evolutionary multi-task ensemble learning provided in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the tagged sample augmentation strategy provided in an embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of a WMF feature filtering operation example provided in an embodiment of the present invention;

[0051] Figure 5 This is a flowchart illustrating the multi-task evolutionary optimization method provided in an embodiment of the present invention;

[0052] Figure 6 This is a schematic diagram illustrating an example of the population initialization mechanism provided in this embodiment of the invention;

[0053] Figure 7 This is a schematic diagram illustrating an example of the cross-operation mechanism provided in an embodiment of the present invention;

[0054] Figure 8 This is a schematic diagram illustrating an example of the mutation operation mechanism provided in an embodiment of the present invention;

[0055] Figure 9 This is a schematic diagram illustrating an example of an integrated strategy for generating the final classification result provided in an embodiment of the present invention;

[0056] Figure 10 This is a schematic diagram of the truth images covering 16 real-world features in the Salinas dataset provided in this embodiment of the invention;

[0057] Figure 11 These are schematic diagrams illustrating the classification results of the truth images of the Salinas dataset provided in this embodiment of the invention using SVM, SSMR, ERW, SCSSR, R-VCANet, and the algorithm of this invention. Detailed Implementation

[0058] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0059] To achieve high-accuracy classification results with limited labeled samples, please refer to [link / reference needed]. Figure 1 This invention provides a hyperspectral image classification method based on evolutionary multi-task ensemble learning, specifically including the following steps:

[0060] S10. Input hyperspectral image data; the hyperspectral image data contains a small number of labeled samples, and the remaining samples are unlabeled samples to be classified.

[0061] Assuming hyperspectral image data Include indivual Samples with 3D spectral characteristics It contains a small number of labeled sample sets. A large number of unlabeled sample sets to be classified , .

[0062] Labeled sample set One of the samples It can be represented as:

[0063] (1);

[0064] in, Indicates the number of land cover categories. This indicates the number of labeled samples. express Having Spectral feature vectors of spectral features, express Category tags.

[0065] Unlabeled sample set to be classified One of the samples It can be represented as:

[0066] (2);

[0067] in, This represents the number of unlabeled samples to be classified. express Having The spectral feature vectors of each spectral feature, where unlabeled samples do not have category labels. .

[0068] S20. Use labeled samples as the training set to pre-train each base classifier. Then, input the unlabeled samples to be classified into each pre-trained base classifier for classification.

[0069] This method employs existing base classifiers and training methods. Labeled samples are used as the training set to pre-train each base classifier. Then, the pre-trained base classifiers are used to classify the unlabeled samples to be classified. The number of base classifiers can be designed as needed; for example, in this invention, the number of base classifiers can be designed to be 2.

[0070] S30. Based on the classification results of the unlabeled samples to be classified in each base classifier, a sample augmentation strategy is used to select a portion of the unlabeled samples to be classified with high confidence to augment the training samples of the training set. Spatial location information is incorporated into the remaining unlabeled samples to be classified through weighted mean filtering.

[0071] Specifically, such as Figure 2 As shown in Part I, the embodiments of the present invention first perform sample augmentation and feature filtering operations on the hyperspectral image data.

[0072] To expand the training samples, this invention proposes a sample expansion strategy. This strategy measures the consistency of the classification results of the unlabeled samples to be classified on each base classifier, selects some high-confidence samples from the unlabeled sample set to be classified, assigns pseudo-labels to them, and adds them to the training set, thereby expanding the training samples. Specifically:

[0073] When expanding the sample set, there will be a labeled sample set. As a training set To train each base classifier, i.e. Then, the unlabeled sample set to be classified... The unlabeled samples to be classified are sequentially input into each base classifier. When different base classifiers classify the currently input unlabeled samples... If the classification results are consistent, the unlabeled sample to be classified is considered to be unlabeled. There is a very high probability that it will be correctly classified. Subsequently, the unlabeled sample to be classified... They are assigned pseudo-labels and added to the training set, while also being processed from the unlabeled sample set to be classified. Removal, that is:

[0074] (3);

[0075] in, express It has spectral eigenvectors, This represents the unlabeled samples to be classified by each base classifier. The pseudo-labels assigned to categories that match the classification criteria.

[0076] At this time, the training set This completes one update. Since the newly added pseudo-labeled samples are instructive for classifying other unlabeled samples, the base classifiers need to be retrained after the training set is updated. Then, the newly pre-trained base classifiers are used to classify the next unlabeled sample. The training sample expansion process ends when all samples in the unlabeled sample set have been classified by the base classifiers. For example, assuming the number of base classifiers is 2... Figure 3 As shown, the unlabeled sample set to be classified Each unlabeled sample to be classified is sequentially input into base classifier 1 and base classifier 2. Based on the discrimination results of the two base classifiers, it can be seen that: the unlabeled samples to be classified with black numbers have consistent classification results on both base classifiers. Therefore, these unlabeled samples to be classified are assigned pseudo-labels sequentially after classification and then used for the next training. The remaining unlabeled samples to be classified with gray numbers have inconsistent classification results on both base classifiers. Therefore, these unlabeled samples to be classified need to undergo weighted mean filtering processing in the subsequent process.

[0077] As can be seen, the sample augmentation strategy for hyperspectral image data in this embodiment of the invention expands the training set of classifiers by using a simple consistency evaluation strategy when there are few labeled samples. The expanded training samples are then added back to the pre-training process of each base classifier, thereby improving the final classification result and reducing noise points in the final classification image.

[0078] Furthermore, after sample augmentation, pseudo-labeled samples (samples with consistent classification results across all base classifiers) added to the training set are removed. The remaining unlabeled samples to be classified are then processed using a weighted mean filter to incorporate spatial location information, eliminating noise and background interference. Specifically:

[0079] This invention considers the unlabeled sample set to be classified The remaining samples are those that are difficult to classify. To eliminate the interference of noise and background points, this invention provides a weighted mean filter (WMF) strategy for this unlabeled sample set to be classified. The raw hyperspectral data is then integrated with spatial location information. The specific implementation for this operation is as follows:

[0080] Each sample in a hyperspectral image exhibits spectral and spatial correlation with its spatially adjacent samples, suggesting they likely belong to the same category of land cover. Based on this, embodiments of the present invention utilize spatial location information to perform feature filtering on samples that are difficult to classify. Figure 4 Provided for the sample Example of weighted mean filtering operation:

[0081] Hypothesis Sample The coordinates in the hyperspectral image data to be classified are: , then With the center point, The nearest neighbor space of a square with side length . It can be defined as:

[0082] (4);

[0083] Among them, the neighboring space Includes 1 pixel, excluding the center point The rest Each sample can be represented as follows: , ... ,for example Figure 4 The middle indicates In this case, for samples located at the edge of the hyperspectral image, the sample itself is used for sample filling.

[0084] center point By using a weighted summation method, utilizing the nearest neighbor space Reconstructing samples by considering their neighborhood points and themselves yields reconstructed samples. Then reconstruct the sample It can be represented as:

[0085] (5);

[0086] in, Representing the nearest neighbor space Middle neighborhood samples, Representing neighborhood samples The weights in the weighted summation, Representing the nearest neighbor space The Middle One sample, Representing the nearest neighbor space The Middle Sample The weights in the weighted summation can be solved using the heat kernel function:

[0087] (6);

[0088] in, Representing the nearest neighbor space All samples and the center sample The average distance can be expressed as:

[0089] (7);

[0090] The WMF feature filtering method uses parameters to... The value is used to adjust the filter window, which is essentially a measure of the nearest neighbor space. Middle neighbor samples and center samples The similarity of the samples is used to reconstruct the weighted mean of the center samples. The higher the similarity, the greater the weight; the greater the difference, the smaller the weight. Therefore, the WMF feature filtering method can effectively eliminate the interference of background points and noise, and obtain images with smoother edges.

[0091] (8);

[0092] S40. Based on the training set expanded by the training samples, establish feature selection tasks for each base classifier, and use the designed multi-task evolutionary optimization method to select the target optimal feature subspace for each feature selection task.

[0093] Traditional feature selection classification methods using ensemble learning do not take into account the correlation between multiple search processes when each base classifier searches for the optimal spectral feature subspace. This invention proposes a multi-task evolutionary optimization method to perform the search for the target optimal spectral feature subspace of each base classifier in parallel.

[0094] like Figure 2 As shown in Part II, for each base classifier, the optimal feature subspace is selected based on the training set expanded with training samples, and the process of selecting the optimal feature subspace is modeled as a corresponding feature selection task, such as... Modeling a base classifier T Multiple feature selection tasks are modeled and optimized simultaneously using a designed multi-task evolutionary optimization method. The ultimate goal of the evolution is to find an optimal solution that describes the performance of each base classifier in its respective feature selection task. The optimal spectral feature subspace selected above , represented as:

[0095] (9);

[0096] in, Represents an individual in an evolutionary population. The corresponding selected spectral feature subspace, The number of individuals in the evolutionary population. Represents an individual In feature selection tasks The objective function value on.

[0097] Therefore, based on the modeled multiple feature selection tasks, the designed multi-task evolutionary algorithm can provide... Generate a base classifier Different optimal spectral feature subspaces , .

[0098] Furthermore, based on the established multiple feature selection tasks, a designed multi-task evolutionary optimization method is used to optimize the population for each task, thereby selecting the optimal feature subspace for each feature selection task. Please see [link to relevant documentation]. Figure 5The specific process of selecting the optimal feature subspace for the target includes the following steps:

[0099] S401. Randomly generate several feature subspaces as individuals to form a population.

[0100] Before multi-task parallel search, first generate a collection of... A population of individuals, where each individual represents a feature subspace. Since the spectral feature dimension of each sample in a hyperspectral image is the same, the dimension can be defined as... A unified search space. When performing feature selection on hyperspectral images, binary 0-1 vector representations are typically used to describe the solution. Therefore, during population initialization, each individual is assigned a... A vector of binary variables. This vector contains all the chromosomes of an individual, and for each individual, the first... The binary value of each chromosome bit determines whether to select a hyperspectral image. The first of all samples Dimensional features.

[0101] Figure 6 A containing An instance of population initialization operation for each individual. Specifically, for the first... Each individual is assigned a randomly generated solution, which can be represented as a vector:

[0102] (10);

[0103] In formula (10), if a certain feature is selected, the value of the corresponding chromosome position is... The value is 1 if the binary vector contains 1s, otherwise it is 0. It is important to note that the number of features selected for each feature selection task is fixed; that is, the number of 1s in the binary vector corresponding to each individual is equal to the predetermined number of features. This is to ensure the validity of feasible solutions. Figure 6 In the given example, the first The feature subset selected by each individual includes the second feature and the fourth-to-last feature.

[0104] S402. Calculate the fitness value of each individual in the population on all feature selection tasks, and calculate the skill factor of each individual in the population based on the fitness value of the population to obtain the parent population and the initial optimal feature subspace set.

[0105] After obtaining the population, in order to generate new individuals through subsequent crossover and mutation operations, it is necessary to calculate and initialize the skill factor of each individual in the population before evolution begins. This embodiment of the invention calculates the skill factor of each individual in the population based on the population's fitness value, including: sorting the fitness values ​​of each individual in the population across all feature selection tasks to form a ranking value for each individual across all feature selection tasks, and taking the index value corresponding to the minimum value among the ranking values ​​of each individual as the skill factor of that individual. The specific content of the population fitness value evaluation strategy is as follows:

[0106] Each base classifier is used to maximize its value in the selected spectral feature subspace set. The classification accuracy is the goal of population evolution. In this embodiment of the invention, only the training set... The classification performance of each base classifier is evaluated. Initially, a training set is formed consisting of a small number of labeled samples and unlabeled samples to be classified, with pseudo-labels. For example, they are randomly divided into two subsets in an 8:2 ratio. and In each task: Used for training the classifier, and for calculating the fitness value of an individual. Assume... In the feature selection task The fitness value of an individual, individual The fitness value is calculated for all feature selection tasks as follows:

[0107] (11);

[0108] in, Individuals in a population In feature selection tasks The fitness value is used in the calculation of the fitness value of a population. Values ​​range from 1 to T , T This indicates the number of feature selection tasks. and Each represents an individual In feature selection tasks On the corresponding base classifier The number of correctly classified samples and Total number of samples This represents a subset of samples obtained by partitioning the training set, which is expanded with training samples, and used to calculate the fitness value of each individual in the population.

[0109] After calculating the fitness values ​​of individuals in the population, the individual exist The skill factor in each task can be calculated as follows:

[0110] (12);

[0111] in, Individuals are sorted in ascending order of fitness value. In the mission The sorting values ​​are used to obtain the skill factors for each individual. Specifically, assuming the number of tasks... All individuals were evaluated on three tasks, and their fitness scores on the three tasks were used to rank them. The three values ​​in the vector represent the individual's ranking for the three tasks. Since a smaller fitness value indicates better performance on the task, a smaller ranking value after sorting the fitness values ​​in ascending order indicates that the individual performs better on the current task; that is, the individual is better at handling tasks with lower ranking values, and is ultimately selected... The index of the smallest sorted value in the vector is used as the individual's skill factor. Specifically, suppose the first individual in the population... individual Vector is The smallest sorted value in the vector is This value is in The index value in the vector is 2, therefore the first... individual skill factors Based on the calculated fitness value and skill factor, we can obtain individuals with a fitness value... and skill factors The parent population of the attribute.

[0112] After obtaining the parent population, the set of individuals for each task can be determined based on the skill factors of the individuals in the parent population. For each task, the initial optimal subspace for that task can be determined by comparing the fitness values ​​of each individual in its respective set. Furthermore, the initial optimal subspaces for all tasks constitute the initial optimal feature subspace set. Specifically, assuming the number of tasks... There were a total of 6 individuals in the parent population. Their skill factor values ​​are {1, 2, 3, 1, 2, 3}. Based on the preceding text, the set of individuals belonging to Task 1 is... The fitness values ​​of these two individuals are respectively When satisfied At that time, the initial optimal feature subspace of Task 1 is the individual The selected feature subspace. Similarly, for Task 2 and Task 3, when the fitness values ​​of individuals in the individual sets to which the two tasks belong respectively satisfy... as well as At that time, the initial optimal feature subspace set is the individual. , and The set of selected feature subspaces.

[0113] S403. Perform crossover and mutation operations on the parent population to obtain the offspring population.

[0114] To generate new individuals, genetic operations such as crossover and mutation are typically performed on the parent population. In a multi-task evolutionary framework, a key characteristic of generating a new spectral feature subspace through crossover is that two randomly selected parent individuals must meet certain conditions before crossover can occur. To ensure that each individual does not repeatedly select the same spectral feature, this invention proposes the following: for cases where the crossover conditions are met, a difference-based crossover operator is used to perform crossover on individuals in the parent population; for cases where the crossover conditions are not met, a constrained single-point mutation operator is used to perform mutation on individuals in the parent population to obtain the offspring population. Specifically, a certain population in the evolutionary process is selected as the current population for analysis.

[0115] Two individuals are randomly selected from the current population to be paired as parents. These two parents are defined as... When the skill factors of two parent individuals satisfy... When using the crossover operator pair Perform a crossover operation to generate offspring individuals. Offspring individuals are generated when the skill factors of the two parent individuals satisfy... When using the generated random number With a predefined crossover probability Comparison: When When using the crossover operator pair Perform a crossover operation to generate offspring individuals; when At that time, mutation operations are performed on both parent individuals. After crossover, the offspring individuals generated by crossover inherit the skill factor of either parent individual with a 50% probability.

[0116] Among them, the crossover operator is the crossover operator based on the difference set, specifically including:

[0117] Calculate the feature difference between two randomly selected parent individuals from the parent population; then perform crossover on the two randomly selected parent individuals using the crossover operator and the calculated feature difference. Specifically:

[0118] First, by analyzing the two parent individuals To create two sets, iterate through the chromosome positions with a value of 1. M and N ,in M Representative of individuals Selected features NRepresentative of individuals Selected features. Further, from the set M and N Two difference sets can be obtained. and .in Represents an individual But the individual No selected feature difference set Representative of individuals Not selected but individual The selected feature difference set. Then, from... N Random selection h Replace each element with Randomly selected from h One element, at the same time, from M Another selection h Replace each element with Randomly selected from h Each element. h for A random number within a given range. Figure 7 An example of a crossover operation is given: Individual The selected feature set is M= {1,2,4,5,6}, Individual The selected feature set is N= {2,3,4,6,7}. Based on this, the feature difference set can be obtained. , .from M Randomly select elements 2 and 5 and replace them with Elements 3 and 7 in the text, from N Randomly select elements 3 and 7 and replace them with N \ M Elements 1 and 5 in the original text. After replacement, the offspring individual will be obtained. and .

[0119] Mutation occurs when the conditions for crossover are not met. This invention proposes using a constrained single-point mutation operator to mutate individuals in the parent population, thereby ensuring the effectiveness of the number of spectral features selected by each individual. Specifically, this includes:

[0120] Based on the mutation probability, individuals in the parent population are mutated, and the number of chromosomes with gene mutations in each mutated individual is counted. For each mutated individual, chromosome positions equal to the number of chromosome positions are selected from the remaining chromosome positions that have not undergone gene mutations and then reverse-mutated. Figure 8 An example of a single-point mutation operation is given, with a single individual For example, suppose an individual If a mutation occurs at one chromosome locus, the corresponding mutation operator will be generated from the individual. One chromosome from the remaining chromosomes that have not undergone gene mutation is randomly selected and mutated in the opposite direction to enhance the resulting individual. Randomness and guaranteeing individual The effectiveness. Individual The dimensions representing solutions with values ​​of 1 and 0 remain unchanged, i.e., individual... The total dimension is In this case, the number of values ​​1 is Offspring individuals generated through mutation directly inherit the skill factors of their corresponding parent individuals.

[0121] S404. For each individual in the offspring population, calculate the fitness value of that individual on the task of selecting the corresponding feature based on that individual's skill factor.

[0122] From S403, the skill factor of each individual in the offspring population can be known, and thus the corresponding trait selection task for each individual in the offspring population can be determined. The fitness value of each individual in the offspring population on its corresponding trait selection task can be calculated using formula (11). At this point, the calculation of the fitness value of the offspring population... t The value is taken from the individual's corresponding skill factor. Assume the individual... skill factors Its fitness value can be determined by Calculated.

[0123] S405. Merge the parent and offspring populations to obtain an intermediate population. Update the initial optimal feature subspace set based on the skill factor and fitness value of the intermediate population to obtain the intermediate optimal feature subspace for each feature selection task.

[0124] From S403 and S404, we know the skill factors and fitness values ​​of all individuals in the intermediate population. Based on the skill factors of the individuals in the intermediate population, we can determine the corresponding set of individuals for each feature selection task. Based on the individual fitness values, each task can select an individual with the optimal fitness value from the corresponding set of individuals. Specifically, when a task is selected as the current task, the individuals corresponding to the initial optimal feature subspace for the current task are... The individual with the optimal fitness value in the intermediate population for the current task is... ,Compare and fitness value and :like Then update the intermediate optimal feature subspace of the current task to the individual. The corresponding feature subspace; otherwise, the intermediate optimal feature subspace of the current task remains unchanged. The corresponding feature subspace. Based on this, the initial optimal feature subspace for each feature selection task is updated to obtain the intermediate optimal feature subspace set for all feature selection tasks.

[0125] S406. The intermediate population generates the target population through an elite selection strategy. The target population is used as the parent population, and the above update process continues until the evolutionary stopping condition is met. The optimal feature subspace set is selected for each feature selection task.

[0126] To ensure that the number of individuals in the parent population remains constant during each evolutionary iteration... In this embodiment of the invention, an elite strategy is used to select the best-performing individuals from the evaluated intermediate population. Each individual forms the next parent population, and this parent population repeats steps S403~S406 above until the maximum number of evolutions is satisfied, i.e., when... At this point, the current evolutionary round represents the intermediate optimal feature subspace selected by the feature selection task corresponding to each base classifier, which is also the target optimal feature subspace for the feature selection task corresponding to each base classifier. Here, the initial evolutionary iteration count is defined. .

[0127] S50. Based on the target optimal feature subspace selected by each feature selection task, a low-dimensional training set is obtained after mapping the training set expanded by the training samples, and a low-dimensional unlabeled sample to be classified is obtained after mapping the unlabeled sample to be classified by weighted mean filtering. Based on the low-dimensional training set, each pre-trained base classifier is trained again, and the trained base classifiers are used to classify the low-dimensional unlabeled sample to be classified, resulting in multiple classification results.

[0128] The multi-task optimization method designed using S40 outputs the target optimal feature subspace for each base classifier. Then, on the low-dimensional mapping dataset of the corresponding target optimal feature subspace, each base classifier is trained and classified, yielding multiple classification results. Specifically, for a certain task... The corresponding base classifier, assuming the target optimal feature subspace output by the feature selection task corresponding to this base classifier. :

[0129] Task The corresponding low-dimensional training set after mapping One of the samples It can be represented as

[0130] (13);

[0131] in, Indicates the number of land cover categories. Indicates task-based The sample of the low-dimensional mapping obtained from the target optimal feature subspace. spectral eigenvectors, Indicates sample Category tags, This indicates the number of training set samples after sample expansion.

[0132] Task The corresponding low-dimensional unlabeled sample set to be classified after mapping One of the samples It can be represented as

[0133] (14);

[0134] in, Indicates task-based The sample of the low-dimensional mapping obtained from the target optimal feature subspace. spectral eigenvectors, This represents the number of unlabeled samples to be classified after weighted mean filtering.

[0135] The base classifier is used on the corresponding low-dimensional training set. Training is performed on the low-dimensional unlabeled sample set to be classified after mapping. The classification is performed to obtain the classification result of the base classifier.

[0136] Based on the classification results of the base classifier, S30-S50 are re-executed. Using the expanded training set and the unlabeled sample set to be classified after weighted mean filtering, the next round of evolutionary optimization selection of the target optimal feature subspace set and the acquisition of classification results are performed until the maximum number of sample expansion iterations is met. Or the updated unlabeled sample set to be classified is an empty set. This represents the maximum number of sample augmentation iterations. The classification results output at this point can then be used for S60 ensemble classification. Specifically, this initializes the number of sample augmentation iterations. .

[0137] S60. Integrate multiple classification results to obtain the final classification result of the hyperspectral image data.

[0138] Multiple classification results of the hyperspectral image are obtained through the S50 process described above. Integrating these results yields the final classification result for the hyperspectral image to be classified. During the integration process, each sample may be classified into different categories by different base classifiers. Majority voting is a commonly used integration strategy. It involves multiple base classifiers judging the samples, with each classifier voting on its predicted category. The category receiving the most votes becomes the final classification result. When several categories receive the same number of votes, a category is usually randomly selected as the final classification result. However, because the classification characteristics of base classifiers often differ, the "one person, one vote" principle cannot reflect the classification advantages of high-performance classifiers.

[0139] Therefore, such as Figure 2 As shown in Part III, this embodiment of the invention designs a comprehensive integration strategy that combines majority voting and a high-precision classifier. It integrates the majority voting method with the high-precision classifier's decision, combining multiple classification results to obtain the classification result of the hyperspectral image to be classified. Specifically: when a sample to be classified has a majority voting category, the sample is directly assigned the majority voting category as its pseudo-label. When a sample to be classified does not have a majority voting category, i.e., when the classification results of multiple base classifiers are inconsistent, the sample is assigned the category corresponding to the high-precision classifier. The integration strategy designed in this embodiment of the invention addresses the problem of sample misclassification caused by inconsistent discriminator classification results in the majority voting method. By combining the majority voting method with the high-precision classifier's decision, it can solve the problem of random sample misclassification to a certain extent. Figure 9 A concrete example of a comprehensive integration strategy is given:

[0140] The samples to be classified are input into base classifiers 1, 2, and 3 respectively. Base classifier 3 has high classification accuracy on the training samples. For the sample to be classified in the upper left corner, the classification results of each base classifier are category 1, category 2, and category 1, respectively. In this case, the majority voting strategy is used to label the sample in the upper left corner as category 1. When the classification results of the three base classifiers are inconsistent, such as the sample to be classified in the lower right corner, the classification results of each base classifier are category 1, category 2, and category 3, respectively. In this case, the high-precision classifier ensemble strategy is used to label the sample in the lower right corner as category 3.

[0141] It should be noted that, in the embodiments of the present invention Figure 2 , Figure 3 , Figure 9 The colored portion represents the input hyperspectral image data and the processed hyperspectral image data. The coloring does not affect the understanding of the method of the present invention.

[0142] To verify the effectiveness of the hyperspectral image classification method based on evolutionary multi-task ensemble learning provided in this embodiment of the invention, the following experiments were conducted.

[0143] To obtain the statistical properties of the algorithm, this embodiment of the invention uses different initial random feature subspaces, repeats each experiment 30 times, and simultaneously calculates the overall accuracy (OA), average accuracy (AA), and the average of the Kappa coefficients. In sample augmentation, the WMF filter window size is adjusted. In hyperspectral image classification, the number of tasks is set. The ensemble classifier uses a multi-class version of Support Vector Machines (SVM) with a Gaussian kernel, K-Nearest Neighbor (KNN) with a neighborhood value of 3, and a discriminator as its three base classifiers. The ensemble result is obtained through a relative voting mechanism among these classifiers. In the evolutionary experiment, the number of individuals in the population... Evolutionary algebra Number of sample expansion iterations To allow for sufficient communication between tasks, a random crossover probability is set. The value is 0.8, and the mutation probability is 0.1. Each feature subspace is initialized using 10%–50% of the features. 10% of the samples are used to construct the labeled training set. The remaining samples serve as the unlabeled sample set to be classified. In the process of individualization, there is a labeled training set. The evaluations were randomly divided into two groups, with 80% used as the training set for individual evaluations and 20% used as the test set for individual evaluations.

[0144] By comparing the proposed algorithm with several hyperspectral image classification algorithms on different datasets, including SVM, Superpixel Spectral Metrics Representation (SSMR), Extended Random Walker (ERW), Spatial and Class Structure Constrained Sparse Representation (SCSSR), and Rolling Guidance Filter and Vertex Component Analysis Network (R-VCANet), the detection results and detection quantization metrics of the proposed algorithm on the Indian Pines, PaviaU, and Salinas datasets are compared, thus verifying the practicality of the proposed model.

[0145] The first dataset consists of images of Indian Pines from AVIRIS, with an image size of [size missing]. Spatial resolution is The spectral bands in the original image are within 0.2. ~2.4 The dataset contains 220 bands. After removing 20 bands due to atmospheric and moisture absorption and low noise ratio, the remaining 200 bands were used for classification. The dataset includes 16 different land cover classes: Alfalfa, Corn-notill, Corn-min, Corn, Grass-mowed, Grass-trees, Grass-pasture, Hay-windrowed, Oats, Soybeans-notill, Soybeans-min, Soybeans-clean, Wheat, Woods, Buildings-grass, and Stones-towers. Table 1 shows the comparison results of OA, AA, and Kappa for six algorithms trained on the Indian Pines dataset using 1%–4% of the labeled samples.

[0146] Table 1. Comparison of classification accuracy results for the Indian Pines dataset.

[0147]

[0148] As shown in Table 1, when the training samples are small, the classification metrics of the algorithm proposed in this invention are higher than those of the comparison algorithms. This is mainly because the proposed algorithm expands the labeled samples through a sample augmentation strategy, and the expanded samples are then reintroduced into the training process of the classification model, thereby improving the final classification results. When training with 1% of the samples, compared with the other five algorithms, the algorithm proposed in this invention improves the OA (Objectives and Quotient) of the Indian Pine dataset by 38.33%, 24.43%, 8.11%, 7.19%, and 6.75%, respectively.

[0149] The second dataset is ROSIS hyperspectral imagery of the University of Pavia (PaviaU). This image has a spatial resolution of 1.3m per pixel. It contains... Each pixel. After removing 12 noise bands and water vapor bands, the remaining 103 ranges are in arrive The bands were used for classification. The image includes nine land cover samples: Asphalt, Meadows, Gravel, Trees, Metal sheet, Bare soil, Bitumen, Brick, and Shadows. When applying the PaviaU dataset, 20 samples from each class were selected as the labeled sample set. Table 2 shows the comparison results of classification accuracy on the PaviaU dataset.

[0150] Table 2 Comparison of classification accuracy results for the Pavia University dataset

[0151]

[0152] The first nine rows of Table 2 correspond to the classification results for each category, while the last three rows show the OA, AA, and kappa results for all categories. It can be seen that the algorithm of this invention has relatively high overall classification accuracy and achieves high statistical results for most categories. For example, in the Asphalt and Shadows categories, the OA value for Asphalt is 97.11%, and the OA value for Shadows is 99.78%. The algorithm of this invention achieves the highest statistical results in OA, AA, and Kappa metrics. When encountering samples that are difficult to classify, the algorithm of this invention performs feature filtering on the original hyperspectral data, incorporating spatial information, and selects a subset of features with high information content and low redundancy to train additional base classifiers. The purpose is to provide differentiated feature subsets for different base classifiers, thereby improving classification accuracy.

[0153] The third dataset consists of hyperspectral images of the Salinas, acquired by the 224-band AVIRIS sensor in the Salinas Valley, California. This dataset contains truth images covering 16 real-world features, such as... Figure 10 As shown, the Salinas scene dataset is 512 rows and 217 columns in size, characterized by its high spatial resolution (3.7m). Similar to the Indian pine scene, after discarding 20 noise and water vapor bands, the remaining 204 bands were used for classification. It contains 16 land cover species: Brocoli1, Brocoli2, Fallow, Fallow_r, Fallow_s, Stubble, Celery, Grapes, Soil, Corn-s, Lettuce-romaine-4wk, Lettuce-romaine-5wk, Lettuce-romaine-6wk, Lettuce-romaine-7wk, Vinyard-u, and Vinyard-v. When applying the Salinas dataset, three samples from each class were selected to construct a labeled sample set. Table 3 presents the classification accuracy comparison results of the Salinas dataset.

[0154] Table 3 Comparison results of classification accuracy in the Salinas dataset

[0155]

[0156] Table 3 shows the classification accuracy of different classification algorithms on the Salinas dataset when using 3 training samples per class. It can be seen that the algorithm of this invention achieves the highest statistical results in terms of OA, AA, and Kappa. The R-VCANet algorithm has an OA of 81.81%, and its poor classification performance is mainly due to the limited number of available training samples. The algorithm of this invention achieves an OA of 98.66%, which is 16.85% higher than R-VCANet.

[0157] Figure 11 Subgraphs (a) to (g) are used to compare the classification mapping diagrams of various traditional algorithms and the algorithm of this invention. Figure 11 Subgraphs (a) to (g) illustrate the classification results of SVM, SSMR, ERW, SCSSR, R-VCANet, and the Salinas classification results obtained by the algorithm of this invention, respectively. It can be seen that the classification map generated by the algorithm of this invention contains the least noise and has the clearest boundaries.

[0158] In summary, the hyperspectral image classification method based on evolutionary multi-task ensemble learning proposed in this invention is a classification method that combines multi-task ensemble and sample augmentation strategies. Specifically: inputting hyperspectral image data; the hyperspectral image data contains a small number of labeled samples, and the remaining samples are unlabeled samples to be classified; using the labeled samples as the training set to pre-train each base classifier, and sequentially inputting the unlabeled samples to be classified into each pre-trained base classifier for classification; based on the classification results of the unlabeled samples to be classified in each base classifier, using a sample augmentation strategy to select a portion of the unlabeled samples to be classified with high confidence to augment the training set, and applying weighted mean filtering to the remaining unlabeled samples to be classified. This invention incorporates spatial location information; based on a training set expanded with training samples, feature selection tasks are established for each base classifier. A designed multi-task evolutionary optimization method is used to select the target optimal feature subspace for each feature selection task. Based on the target optimal feature subspace selected for each feature selection task, a low-dimensional training set mapped from the expanded training set and a low-dimensional unlabeled sample to be classified mapped from the unlabeled sample to be classified after weighted mean filtering are obtained. The pre-trained base classifiers are then trained again based on the low-dimensional training set, and the trained base classifiers are used to classify the low-dimensional unlabeled sample to be classified, resulting in multiple classification results. The final classification result of the hyperspectral image data is obtained by integrating multiple classification results. Therefore, this invention, based on a multi-task integration and sample expansion strategy, first expands the limited labeled samples, then uses multiple base classifiers in parallel to generate a target optimal feature subspace set based on the expanded labeled samples. This achieves feature sharing among the base classifiers and mutual assistance during the optimal feature search process to find the globally optimal feature subspace. Finally, the performance of hyperspectral image classification is improved by integrating the classification results of multiple base classifiers. In addition, this classification method minimizes noise in the classification map and has a certain degree of robustness.

[0159] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0160] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the specification and accompanying drawings, will understand and implement other variations of the disclosed embodiments in carrying out the claimed invention. In the specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0161] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A hyperspectral image classification method based on evolutionary multi-task ensemble learning, characterized in that, The method comprises the following steps: inputting hyperspectral image data; the hyperspectral image data comprises a small number of labeled samples and the remaining samples are all unlabeled samples to be classified; pre-training each base classifier by taking the labeled samples as a training set, and inputting the unlabeled samples to be classified into each pre-trained base classifier in sequence for classification; according to the classification results of the unlabeled samples to be classified in each base classifier, selecting high-confidence parts of the unlabeled samples to be classified by using a sample expansion strategy to expand the training samples of the training set, and fusing the remaining unlabeled samples to be classified by using weighted mean filtering to incorporate spatial position information; based on the training set expanded by the training samples, establishing a feature selection task for each base classifier, and selecting a target optimal feature subspace for each feature selection task by using a designed multi-task evolutionary optimization method; based on the target optimal feature subspace selected by each feature selection task, obtaining a low-dimensional training set mapped by the training set expanded by the training samples, and a low-dimensional unlabeled sample to be classified mapped by the unlabeled sample to be classified subjected to the weighted mean filtering, re-training each pre-trained base classifier based on the low-dimensional training set, classifying the low-dimensional unlabeled sample to be classified by using the trained base classifiers, and obtaining multiple classification results; integrating the multiple classification results to obtain a final classification result of the hyperspectral image data; wherein, based on the training set expanded by the training samples, establishing a feature selection task for each base classifier, and selecting a target optimal feature subspace for each feature selection task by using a designed multi-task evolutionary optimization method, comprises: for each base classifier, selecting an optimal feature subspace based on the training set expanded by the training samples, and modeling the process of selecting the optimal feature subspace as a corresponding feature selection task; based on the modeled feature selection task, selecting a target optimal feature subspace for each feature selection task by using a designed multi-task evolutionary optimization method, comprises: using the designed multi-task evolutionary optimization method to perform task optimization on a population, so as to select a target optimal feature subspace for each feature selection task; wherein each individual in the population represents a feature subspace.

2. The method of claim 1, wherein the method is based on evolutionary multitask ensemble learning. According to the classification results of the unlabeled samples to be classified in each base classifier, the process of selecting high-confidence parts of the unlabeled samples to be classified by using a sample expansion strategy to expand the training samples of the training set, comprises: for each unlabeled sample to be classified, the following process is performed: inputting the unlabeled sample to be classified into each pre-trained base classifier in sequence to obtain a classification result; assigning a pseudo-label to the unlabeled sample to be classified which obtains consistent classification results in all base classifiers, and adding it to the training set; wherein, before the next unlabeled sample to be classified is input into each pre-trained base classifier, each base classifier is re-trained by using the new training set; after all the unlabeled samples to be classified have undergone the above process, the training sample expansion process ends.

3. The method of claim 1, wherein the method is based on evolutionary multitask ensemble learning. using the designed multi-task evolutionary optimization method to perform task optimization on a population, so as to select a target optimal feature subspace for each feature selection task, comprises: S401, randomly generate several feature subspaces as individuals to form a population; S402, calculate the fitness value of each individual in the population on all feature selection tasks, calculate the skill factor of each individual in the population according to the fitness value of the population, obtain a parent population and an initial optimal feature subspace set; S403, perform cross and mutation operations on the parent population to obtain a child population; S404, for each individual in the child population, calculate the fitness value of the individual on the feature selection task to which the individual belongs according to the skill factor of the individual; S405, merge the parent population and the child population to obtain an intermediate population, update the initial optimal feature subspace set according to the skill factor and the fitness value of the intermediate population to obtain an intermediate optimal feature subspace set of all feature selection tasks; S406, the intermediate population generates a target population through an elite selection strategy, takes the target population as the parent population, takes the intermediate optimal feature subspace set as the initial optimal feature subspace set, continues to repeatedly execute S403-S406 until the maximum evolution number is reached, and takes the intermediate optimal feature subspace set selected by all feature selection tasks as the target optimal feature subspace set.

4. The method of claim 3, wherein the method is based on evolutionary multitask ensemble learning. The calculation formula of the fitness value is: ； wherein, denotes an individual The fitness value of an individual is calculated as the sum of the fitness values of the base classifiers of the individual on the feature selection task t The fitness value of a population is calculated as T , T denotes the number of feature selection tasks, the fitness value of a population is calculated as t The fitness value of an individual is calculated as the sum of the fitness values of the base classifiers of the individual on the feature selection task and respectively denote the number of samples correctly classified by the base classifier of the individual on the feature selection task and the total number of samples, denotes the subset of the training set obtained by dividing the training set augmented with the training samples for calculating the fitness value of each individual.

5. The method of claim 3, wherein the method is based on evolutionary multitask ensemble learning. The calculation formula of the fitness value is: The calculation formula of the fitness value is:

6. The method of claim 3, wherein the method is based on evolutionary multitask ensemble learning. Sort the fitness values of each individual in the population on all feature selection tasks, and take the index value corresponding to the minimum value in the sorting result as the skill factor of the individual. The calculation formula of the fitness value is: When the two parent individuals randomly selected from the parent population have the same skill factor, the cross operation condition is met, and the difference set-based crossover operator is used to perform cross operation on the two parent individuals; 7. The method of claim 6, wherein the method is based on evolutionary multitask ensemble learning. When the two parent individuals randomly selected from the parent population have different skill factors, it is determined whether to perform cross operation on the two parent individuals according to a predetermined cross probability, including: generating a random number; if the random number is less than or equal to the predetermined cross probability, the cross operation condition is met, and the difference set-based crossover operator is used to perform cross operation on the two parent individuals; if the random number is greater than the predetermined cross probability, the cross operation condition is not met, and the single-point-based mutation operator with constraints is used to perform mutation operation on the individuals in the parent population. The calculation formula of the fitness value is: Calculate the feature difference set of the two parent individuals randomly selected from the parent population with respect to each other; 8. The method of claim 6, wherein the method is based on evolutionary multitask ensemble learning. Cross the two parent individuals randomly selected by using the crossover operator and the calculated feature difference set. The calculation formula of the fitness value is: According to the mutation probability, the individuals in the parent population are mutated, and the number of chromosome positions in which gene mutation occurs in each individual after mutation is counted; For each individual after mutation, a number of chromosome positions equal to the number of chromosome positions in which gene mutation occurs are selected from the remaining chromosome positions of the individual in which gene mutation does not occur for reverse mutation.

9. The method of claim 1, wherein the method is based on evolutionary multitask ensemble learning. The integrated multiple classification results obtain a final classification result of the hyperspectral image data, comprising: The integrated multiple classification results obtain a final classification result of the hyperspectral image data by using an integrated strategy of comprehensive majority voting and a high-precision classifier.

Images