A band-selective hyperspectral classification method based on heterogeneous cuckoo search

Abstract

The present invention discloses a band selection hyperspectral classification algorithm based on heterogeneous cuckoo search to obtain high-quality hyperspectral image classification results. The method comprises: first, adopting a heterogeneous search strategy to improve the exploration and development capabilities of the cuckoo search algorithm. Then, a band selection model based on a matched filter is designed to select a band subset with a high signal-to-noise ratio. The model removes redundant bands by reducing the dimension of the hyperspectral data, while providing a band subset that can retain high information content and low noise. A mapping method based on a neighborhood grouping strategy is designed to reduce the similarity between the selected bands. Subsequently, an automatic recommendation strategy for the number of bands based on sliding spectrum decomposition is proposed. Experimental results on three well-known datasets show that compared with the most advanced methods such as marginalized graph self-representation and neighborhood grouping normalized matched filter, the proposed algorithm has the advantages of robustness, superiority and regional multi-graph fusion.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a band selection hyperspectral classification method based on heterogeneous cuckoo search. Background Art

[0002] In recent years, with the advancement of spectral imaging technology, hyperspectral images captured by sensors have rich spectral information and can easily distinguish subtle differences between land cover materials. Therefore, hyperspectral images are applied in a variety of fields, such as medicine, agriculture, and construction. As the number of spectral channels increases, the number of frequency bands also increases accordingly, which places a heavy burden on the research tasks of hyperspectral image classification, target detection, and change detection. In order to alleviate the high computing and storage requirements brought about by the explosive growth of hyperspectral data, it is crucial to study dimensionality reduction technology. Feature selection and feature extraction are the main research topics of hyperspectral image data dimensionality reduction technology. The former is to select a set of suitable bands from the dataset through pre-set criteria, also known as band selection; the latter is to generate new features by performing a series of transformations on the original spectral bands. Generally speaking, hyperspectral image band selection includes supervised methods and unsupervised methods.

[0003] Taking into account prior knowledge, supervised band selection methods can achieve better dimensionality reduction performance while providing more accurate features. Unsupervised band selection methods rely on assumptions about the data to complete dimensionality reduction, but their performance is usually inferior to the former. However, supervised methods often require data to be labeled for dimensionality reduction, which limits their application. Due to the high complexity and large-scale nature of hyperspectral images, performing labeling tasks is extremely time-consuming and labor-intensive, and is also prone to errors. Therefore, unsupervised band selection methods have wider applicability, and this application will focus on unsupervised band selection methods. Summary of the Invention

[0004] Hyperspectral image classification involves assigning each pixel in a hyperspectral image to a specific category or set of categories. This application addresses the aforementioned issues with existing methods by providing an unsupervised band-selective hyperspectral image classification algorithm based on heterogeneous cuckoo search, aiming to achieve more accurate hyperspectral image classification results.

[0005] In order to achieve the purpose of the present invention, the technical solution adopted by the present invention is:

[0006] A band selection hyperspectral classification method based on heterogeneous cuckoo search is characterized in that the algorithm includes the following steps:

[0007] S1. Initialize the weight parameters of the algorithm and determine the search space for the input hyperspectral image data;

[0008] S2. Using the cuckoo search algorithm to perform global exploration and local utilization within the determined search space to obtain population individuals;

[0009] S3. Perform heterogeneous search on the obtained population individuals to obtain individuals with weight labels;

[0010] S4. Filter out the bands with densely packed individuals with weighted labels to obtain a band subset with a high signal-to-noise ratio;

[0011] S5. performing group mapping of neighboring bands on the subset of bands with high signal-to-noise ratio, and grouping the bands;

[0012] S6. Recommending the number of bands for the grouped bands using sliding spectrum decomposition, and obtaining the bands for the final classification task from each group through recommendation;

[0013] S7. Apply the classification algorithm to the recommended bands to obtain classification results, thereby completing the classification.

[0014] Furthermore, the global exploration in S2 is completed by updating the population through Lévy flight, which is defined as follows:

[0015]

[0016] in: is the i-th population individual in the t-th iteration, is the i-th population individual in the t+1th iteration, α is the weight scale parameter, which is 0.01 (supplement), and Levy(λ) represents the Levy flight algorithm formula, which is defined as follows:

[0017] Levy(λ)=μ / |v| 1 / β

[0018] Where: μ~N(0,σ 2 ),v~N(0,1),

[0019] Where: N represents the normal distribution, μ and v represent random variables that obey the normal distribution, β is the weight scale parameter of the random variable v, σ 2 is the variance.

[0020] Furthermore, the local utilization in S2 is completed by performing population update through preference random walk, which is defined as follows:

[0021]

[0022] Among them: rand represents the weight scale parameter of the population individual (which needs to be distinguished from other scale parameters), which is a random real number between 0 and 1. represents the jth individual in the population at the tth iteration, represents the kth individual in the population at the tth iteration;

[0023] In order to ensure that the algorithm achieves better performance in later iterations, a quantum mechanism is introduced, which is defined as follows:

[0024]

[0025] where δ is the difference weight parameter that controls the differences between individuals in the population.

[0026] Furthermore, the definition of heterogeneous search in S3 is as follows:

[0027]

[0028] in: is the i-th individual in the t+1-th iteration, ε=δexp(η), where and ε are both difference parameters that control the differences between individuals in the population. η is a function variable that is mainly used to correct the difference parameters and is set by prior knowledge.

[0029] ln() means taking the logarithmic function and performing logarithmic operations with e as the base, exp() means taking the exponential function and performing exponential operations with e as the base, g t represents the optimal population individual of the tth iteration, and the interval parameter sr is calculated as follows:

[0030]

[0031] Where: 2 is the variance of the entire population.

[0032] Furthermore, the grouping definition of the neighborhood bands in S5 is as follows:

[0033]

[0034] Where M is the number of bands, K is the number of groups, and P k The initial cluster centers in each group.

[0035] Furthermore, the definition of the sliding spectrum decomposition in S6 is as follows:

[0036]

[0037] Where: max is the sequence, γ is the hyperparameter that balances diversity and quality, j is the variable, r is the scale parameter for measuring sequence quality (r here has a different meaning from r in the previous text and needs to be modified to two different unknowns. The content in the manual also needs to be modified), T 1:trepresents the diversity of the j-th band sequence from the 1st to the tth iteration, v represents the singular value of the energy band, and t is the number of iterations;

[0038] The inflection point of the curve is used to determine the recommended number of bands. The definition of the inflection point of the curve is as follows:

[0039]

[0040] Where re represents the inflection point of the curve, i and K are independent variables, and w i represents the quality of the i-th band, arg max represents the parameter or index that takes the maximum value of the sequence, γ is a hyperparameter that balances diversity and quality, T i is the sequence diversity of the i-th band.

[0041] Furthermore, the classification method in S7 is a classification method combining KNN with random forest.

[0042] KNN and Random Forest are two common classification algorithms that can be used together to improve classification performance. The following are the steps for combining KNN and Random Forest:

[0043] Build a random forest: First, use the training set to build a random forest model. Random forest is an ensemble learning algorithm composed of multiple decision trees. Each decision tree is built based on a different random subset and feature subset, and the classification is performed through collective decision-making.

[0044] Make predictions for the test sample: For the test sample, input it into each decision tree in the random forest model and record the prediction results of each decision tree.

[0045] KNN classification: For each test sample, according to the KNN algorithm, the K nearest neighbor training samples are selected in the random forest (which can be the K samples with the closest feature distance to the sample), and voting is performed based on the labels of these K neighbors.

[0046] Prediction result: According to the results of KNN voting, the final classification label of the test sample is determined.

[0047] In this method, random forest is used to provide initial classification predictions, while the KNN algorithm is used to further vote and determine the final classification based on the random forest. This combination can fully utilize the ensemble learning advantages of random forest and the local neighborhood judgment ability of KNN to improve the accuracy and robustness of classification.

[0048] Compared with the prior art, the advantages of the present invention are:

[0049] 1. Band selection is performed on high-dimensional hyperspectral data to remove redundant bands while retaining high information content and low noise, reducing the amount of data required to be processed by the algorithm.

[0050] 2. Adopt heterogeneous search strategies to improve the exploration and development capabilities of the cuckoo search algorithm.

[0051] 3. Design a matched filter-based frequency band selection model to select a subset of frequency bands with a high signal-to-noise ratio. This subset of frequency bands contains the most discriminative and informative features in the image, which is crucial for classification tasks. It can better distinguish between different image categories, thereby improving classification accuracy. It is also robust to noise and image variations, suppressing noise interference and extracting more stable and reliable features.

[0052] 4. Design a mapping method based on a neighborhood grouping strategy to reduce the similarity between selected bands and increase the richness of selected features, ensuring that feature bands with greater differences are selected, which can better distinguish between different categories of images. Bands with less similarity can provide more unique information, thereby improving classification accuracy.

[0053] 5. Design an automatic band number recommendation strategy based on sliding spectrum decomposition. Sliding spectrum decomposition decomposes the image's spectral information to extract characteristic information from different bands. Furthermore, sliding spectrum decomposition can decompose the image's spectral information into distinct bands, each containing distinct characteristics. Automatically recommending the appropriate number of bands eliminates redundant information, retaining only the most useful features. This reduces the computational complexity of the classification algorithm, avoids unnecessary calculations and processing, and reduces computational time, improving efficiency and preventing misclassifications caused by redundant information. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic diagram of the overall process of the algorithm framework;

[0055] Figure 2 It is a schematic diagram of the mapping method based on neighborhood band grouping;

[0056] Figure 3 It is the algorithm flow chart. DETAILED DESCRIPTION

[0057] To further illustrate the technical means and effects of the present invention to achieve the intended purpose, a band selection hyperspectral classification method based on heterogeneous cuckoo search proposed by the present invention is described in detail below with reference to the accompanying drawings and preferred examples:

[0058] It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention. The algorithm of the present invention adopts heterogeneous search. The cuckoo algorithm consists of global exploration and local utilization. Global exploration is to achieve population update through Lévy flight, which is defined as follows:

[0059] is the i-th individual in the t-th iteration, is the i-th individual in the t+1th iteration, α=0.01. Levy(λ) represents the Levy flight algorithm formula, which is defined as follows:

[0060] Levy(λ)=μ / |v| 1 / β

[0061] where μ~N(0,σ 2 ),v~N(0,1), where N represents the normal distribution, μ and v represent random variables that obey the normal distribution, and β is the scale parameter.

[0062] Local utilization is achieved by performing population updates via a preference random walk, which is defined as follows:

[0063]

[0064] Where r represents the scale parameter, which is a random value between 0 and 1. and Represent two different individuals in the population at the tth iteration. The evolution is through the two individuals and However, the search capability is not enough to support the algorithm to achieve better performance in later iterations.

[0065] According to quantum mechanics, L is defined as the distance between two candidate solutions and the average position of all candidate solutions

[0066]

[0067] Where δ is a parameter that controls the difference between them. In order to enhance the perturbation ability of the cuckoo search algorithm in the search space, a heterogeneous search strategy is proposed based on formula (2) and quantum mechanism, which is defined as follows:

[0068]

[0069] Where, is the i-th individual in the t+1-th iteration, ζ=δln(1 / n), ξ=δexp(η), ln() represents the logarithmic function with e as the base, exp() represents the exponential function with e as the base, g t represents the best individual of generation t, and sr is defined as follows:

[0070]

[0071] σ 2 is the variance of the entire population. In addition, a heterogeneous cuckoo search algorithm is proposed to improve the exploration and utilization of the algorithm. The detailed process of the heterogeneous cuckoo search algorithm is shown below.

[0072]

[0073]

[0074] Inspired by the method of removing bad frequency bands based on matched filter weights, a band selection model based on a normalized matched filter detector is proposed to select a subset of frequency bands with a higher signal-to-noise ratio. The weights calculated by the matched filter detector represent the signal-to-noise ratio and image quality of each frequency band. The larger the weight, the higher the signal-to-noise ratio. Therefore, an unsupervised band selection model is designed, which is defined as follows:

[0075]

[0076] Where minf(x) represents the minimum function, K represents the number of selected bands, and X k Represents the selected band, W and H represent the width and height of the image respectively. represents the Xth k The weight of the i-th pixel in the frequency band.

[0077] In most cases, many methods only consider the amount of information and noise contained in the band, while ignoring the impact of inter-band correlation on hyperspectral image classification. However, how to minimize the similarity between selected bands is particularly important. Generally speaking, the similarity between adjacent bands is usually higher than that between non-adjacent bands. Therefore, through the design of novel mapping methods, the similarity of the selected band subsets is lower, such as Figure 2 In the neighborhood band grouping method, the hyperspectral image cube is first divided into K groups as evenly as possible, and the divided point set G is defined as follows:

[0078]

[0079] Where M is the number of bands, K is the number of groups, which is also the number of selected bands, and mod(·,·) represents the modulo operation. However, the initial division of energy bands according to formula (9) lacks a theoretical basis, resulting in unsatisfactory results. Therefore, the fine neighborhood grouping method is used to group G k Re-partition, where P k is considered as the initial cluster center in each group and is defined as follows:

[0080]

[0081] In the groups divided by the neighborhood grouping strategy, the frequency bands initially selected by the heterogeneous cuckoo search algorithm are distributed in different groups, which will result in three different situations: 0, 1, or ≥2 frequency bands selected in the group. For the selected frequency bands in each group, different strategies are used for mapping methods to obtain subsets of the selected frequency bands, such as Figure 2 As shown in the figure. If there are only 0 candidate frequency bands in the group, the frequency band with the largest matched filter weight is selected in the group. If there is 1 selected frequency band in the group, this frequency band is the final selected frequency band. If there are ≥ 2 candidate frequency bands in the group, the frequency band is selected using a roulette wheel selection strategy.

[0082] Currently, band selection methods focus more on the impact of different band numbers on classification accuracy, but there is less research on determining the number of selected bands. Therefore, we propose a recommendation strategy based on sliding decomposition to determine the number of bands. Sliding decomposition mainly provides high-quality and diverse content sequences in the content summary, which is defined as follows:

[0083]

[0084] Where max is the sequence, γ is the hyperparameter that balances diversity and quality, j is the variable, r and T 1:t Represent the sequence quality and diversity, respectively, and v represents the singular value of the energy band. In order to obtain the minimum number of selected frequency bands, the sliding decomposition measures the quality and diversity of the selected frequency bands, and the inflection point is the recommended number of selected frequency bands. This application obtains a smooth curve by fitting an exponential function to the discrete probability, and obtains the inflection point of the curve to determine the recommended number of bands, which is defined as follows:

[0085]

[0086] Where r represents the inflection point of the curve, i and K are independent variables, and w i represents the quality of the i-th frequency band, argmax represents the parameter or index that takes the maximum value of a function or sequence, γ is a hyperparameter that balances diversity and quality, T i For sequence diversity.

[0087] Experimental results and analysis: In the experiment, public datasets such as Botswana, Kennedy Space Center and Salinas were used to verify the effectiveness and applicability of the proposed algorithm. Experimental setup In order to verify the quality of the selected frequency bands, three classifiers, random forest, support vector machine and support vector machine based on edge-preserving filtering, were applied to classify the samples. Since all classifiers are supervised, we randomly selected 10% of each dataset as the training set. The two classifiers on different datasets use the same parameters. There are many performance indicators in hyperspectral image classification to evaluate the classification results. In this application, the overall accuracy (OA) is the percentage of correctly classified samples in all samples, the average accuracy (AA) is the average of the classification accuracy of each category, and the Kappa coefficient gives the percentage of correctly classified pixels. Three evaluation factors are used. The classification accuracy of the algorithm is compared with other advanced algorithms as shown in Table 1.

[0088] Table 1 Comparison of classification accuracy of the proposed classification algorithm and other state-of-the-art algorithms

[0089]

[0090]

[0091] The recommended number of bands for the three datasets is 13, 14, and 18, respectively. To fairly verify the performance of the algorithms used in this application, the above methods were run independently 30 times, where the parameters of the matched filter-based cuckoo search algorithm and the matched filter-based heterogeneous cuckoo search algorithm remained consistent throughout the experiment, such as population size N = 30 and discovery probability Pa = 0.25. If the fitness of the algorithm does not change in 5 consecutive iterations, the method is automatically terminated. The number of bands selected is 30. The bold font in all tables represents the best classification performance of all methods. As summarized in Table 1, the classification results of the frequency band subset characterized by a wider distribution are more satisfactory. For the recommended number of bands for the three datasets, the proposed matched filter-based heterogeneous cuckoo search algorithm outperforms the other compared methods in terms of classification results obtained using different classifiers. For the Botswana dataset, the numerical evaluation of the classification results of different classifiers of the proposed method and other methods on the selected band subsets are shown in Tables 2, 3 and 4, respectively, including the overall accuracy (OA), average accuracy (AA) and Kappa coefficient, as well as the classification accuracy of each category.

[0092] Table 2 Random forest classification accuracy of the proposed algorithm and other state-of-the-art algorithms in Botswana

[0093]

[0094]

[0095] Table 3 Classification accuracy of support vector machine for the algorithm proposed by Botswana and other state-of-the-art algorithms

[0096]

[0097] Table 4 Classification accuracy of the support vector machine based on edge-preserving filter for the algorithm proposed by Botswana and other state-of-the-art algorithms

[0098]

[0099]

[0100] As shown in Table 2, using random forest as the classifier, the proposed algorithm achieved the highest classification accuracy in 7 of the 14 categories, achieving the best performance across all metrics compared to the other methods. Besides the proposed method, the adaptive subspace partitioning strategy achieved the highest classification accuracy, with an overall accuracy of 84.92%, while the worst was the maximum variance principal component analysis, with an overall accuracy of 81.03%. Using the support vector machine as the classifier in Table 3, although the adaptive subspace partitioning strategy remained the best band selection method besides the proposed method, the number of classes achieving the highest classification accuracy decreased. Marginalized graph self-representation achieved the best classification accuracy in both 1 and 14 categories, with both achieving 100%, but its overall accuracy, average accuracy, and kappa coefficient were the worst among the compared methods. Although the cuckoo search algorithm with matched filters did not achieve the highest classification accuracy across all categories, its overall accuracy, average accuracy, and kappa coefficient were still higher than those of the neighborhood grouping normalized matched filter, robust dual-graph self-representation, and marginalized graph self-representation, demonstrating the rationality and superiority of the matched filter-based band selection model. When using a support vector machine with an edge-preserving filter as the classifier, all band selection methods except maximum variance principal component analysis achieved classification accuracies exceeding 98%, as shown in Table 4. The proposed algorithm achieved the highest classification accuracy in 9 of the 14 categories, with 100% accuracy achieved in 6 of them. The adaptive subspace partitioning strategy and the fast neighborhood grouping method for hyperspectral band selection achieved slightly better classification results than the proposed method, with overall accuracies of 98.84% and 98.85%, respectively. Comprehensive analysis of the above experimental results shows that the proposed algorithm outperforms other methods on different classifiers. Compared with the cuckoo search algorithm with matched filters, the effectiveness and superiority of the proposed heterogeneous search strategy are demonstrated. The performance of the proposed method under multi-scale cross-branch is studied. Experiments are conducted on the recommended number of selected bands, 30 selected bands and all bands of each dataset under the hyperspectral image classification method based on multi-scale cross-branch. Response and second-order channel attention are used respectively. 10% of the training samples are randomly selected in the experiment. The experimental results are shown in Table 5.

[0101] Table 5 Classification results under multi-scale cross-branch response and second-order channel attention

[0102]

[0103]

[0104] To highlight the superiority of the execution time of the proposed method, all band selection methods were executed in Matlab 2020a using a PC workstation (Intel Core i9-11900K CPU processor and 32GB RAM). The execution time of all band selection methods for selecting different numbers of bands in the three datasets is shown in Table 6.

[0105] Table 6 Execution time (seconds) of different methods for selecting a certain number of bands on three datasets

[0106]

[0107] Experimental results show that robust dual graph self-representation, region-level multi-graph fusion, and marginalized graph self-representation methods have very high execution times, while other methods have lower processing times. For the Salinas dataset, both robust dual graph self-representation and marginalized graph self-representation methods have execution times exceeding 200 seconds, which is prohibitively expensive. Compared to the cuckoo search algorithm with matched filters, the proposed method only takes approximately 0.05 seconds to execute. Clearly, improving classification performance within an acceptable timeframe is important and necessary. Compared to the neighborhood grouping normalized matched filter and the fast neighborhood grouping method with hyperspectral band selection, the proposed method has a lower execution time than the neighborhood grouping normalized matched filter and a higher execution time than the fast neighborhood grouping method with hyperspectral band selection, while achieving satisfactory classification performance.

[0108] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modification, equivalent change and modification made to the above embodiment by any technician familiar with the profession according to the technical essence of the present invention without departing from the scope of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A band-selective hyperspectral classification method based on heterogeneous cuckoo search, characterized by: The method comprises the following steps: S1. Initialize the weight parameters of the algorithm and determine the search space for the input hyperspectral image data; S2. Using the cuckoo search algorithm to perform global exploration and local utilization within the determined search space to obtain population individuals; S3. Perform heterogeneous search on the obtained population individuals to obtain individuals with weight labels; The definition of heterogeneous search is as follows: in: is the i-th individual in the t+1-th iteration, ε=δexp(j), ln() represents the logarithmic function with e as the base, exp() represents the exponential function with e as the base, g t represents the best individual of generation t, sr is as follows: Where: 2 is the variance of the entire population; S4. Filter out the bands with densely packed individuals with weighted labels to obtain a band subset with a high signal-to-noise ratio; S5. performing group mapping of neighboring bands on the subset of bands with high signal-to-noise ratio, and grouping the bands; S6. Recommending the number of bands for the grouped bands using sliding spectrum decomposition, and obtaining the bands for the final classification task from each group through recommendation; S7. Apply the classification algorithm to the recommended bands to obtain classification results, thereby completing the classification.

2. The band-selective hyperspectral classification method based on heterogeneous cuckoo search according to claim 1, characterized in that: The global exploration in S2 is completed by updating the population through Levy flight, which is defined as follows: in: is the i-th individual in the t-th iteration, is the i-th individual in the t+1th iteration, α is the weight scale parameter, which is 0.01, and is the artificial intervention correction of the proportion of the Levy flight algorithm. Levy(λ) represents the Levy flight algorithm formula, which is defined as follows: Levy(λ)=μ / |v| 1 / β Among them: μ~N(0,σ 2 ),v~N(0,1), Where: N represents the normal distribution, μ and v represent random variables that obey the normal distribution, β is the weight scale parameter of the random variable v, σ 2 is the variance.

3. The band selection hyperspectral classification method based on heterogeneous cuckoo search according to claim 1 is characterized by: The local utilization in S2 is completed by performing population update through preference random walk, which is defined as follows: Among them: rand represents the weight scale parameter of the population individual, which is a random real number between 0 and 1. represents the jth individual in the population at the tth iteration, represents the kth individual in the population at the tth iteration; In order to ensure that the algorithm achieves better performance in later iterations, a quantum mechanism is introduced, which is defined as follows: where δ is a variance parameter that controls the variance between individuals in the population.

4. The band selection hyperspectral classification method based on heterogeneous cuckoo search according to claim 1 is characterized by: The grouping definition of the neighborhood bands in S5 is as follows: Where M is the number of bands, K is the number of groups, and P k The initial cluster centers in each group.

5. The band selection hyperspectral classification method based on heterogeneous cuckoo search according to claim 1 is characterized by: The definition of the sliding spectrum decomposition in S6 is as follows: Among them: max is the sequence, γ is the hyperparameter that balances diversity and quality, j is the variable, r represents the scale parameter for measuring sequence quality, T 1:t represents the diversity of the j-th band sequence from the 1st to the tth iteration, v represents the singular value of the energy band, and t is the number of iterations; The inflection point of the curve is used to determine the recommended number of bands. The definition of the inflection point of the curve is as follows: Where re represents the inflection point of the curve, i and K are independent variables, and w i represents the quality of the i-th band, arg max represents the parameter or index that takes the maximum value of the sequence, γ is a hyperparameter that balances diversity and quality, T i is the sequence diversity of the i-th band.

6. The band selection hyperspectral classification method based on heterogeneous cuckoo search according to claim 1 is characterized by: The classification method in S7 is a KNN combined with random forest classification method; KNN and random forest are two common classification algorithms, which are used in combination to improve classification performance; the following are the steps of the classification method combining KNN and random forest: Constructing a random forest: First, a random forest model is constructed using the training set. Random forest is an ensemble learning algorithm consisting of multiple decision trees. Each decision tree is constructed based on a different random subset and feature subset, and classification is performed through collective decision-making. Predict the test sample: For the test sample, input it into each decision tree in the random forest model and record the prediction results of each decision tree; KNN classification: For each test sample, the K nearest neighbor training samples are selected in the random forest according to the KNN algorithm, and voting is performed based on the labels of these K neighbors; Prediction results: Determine the final classification label of the test sample based on the results of KNN voting; In this method, random forest is used to provide initial classification predictions, while the KNN algorithm is used to perform further voting and determine the final classification based on the random forest; this combination fully utilizes the ensemble learning advantages of random forest and the local neighborhood judgment ability of KNN to improve the accuracy and robustness of classification.

Images