A hyperspectral image classification method and system based on PCANet
By introducing non -linear functions and pre -processing steps in the high spectral image classification method, build a PCANET -based high -spectrum image classification model, solve the problem of low high spectral image classification accuracy in existing technology, and achieve more efficient calculations and higher computing and higher Classification accuracy.
Patent Information
- Application Number
- CN202210521660.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-13
- Publication Date
- 2025-01-24
- Estimated Expiration
- 2042-05-13
AI Technical Summary
When the existing hyperspectral image classification method deals with nonlinear light scattering, the classification accuracy is low and the Kappa index is too low to meet the needs of users.
Using PCANET -based high spectrometer image classification methods, by introducing non -linear functions and pre -processing to remove secondary spectral information, construct a high spectrum of the first layer of convolution nuclear, non -linear functions, second -layer convolution kernels and classifiers Image classification model.
It improves the accuracy of hyperspectral image classification, reduces the amount of calculation, and meets users' needs for hyperspectral image classification.
Smart Images
Figure CN114882360B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a hyperspectral image classification method. Background Art
[0002] Hyperspectral images are a collection of spectral image data that captures each pixel band in hundreds of continuous and narrow bands through remote sensing systems (e.g., satellite-borne or airborne sensors). Continuous band data provides detailed information that can be used to identify the target category of ground objects. Hyperspectral images can be used in many fields such as geological exploration, environmental monitoring, precision agriculture and vegetation detection. In the past decade, with the development of computer hardware and the improvement of computing efficiency, deep learning methods such as convolutional neural networks have been widely used in tasks such as vision and natural language processing, and have achieved relatively ideal results. In recent years, methods based on convolutional neural networks have also begun to be widely used in hyperspectral image classification. However, deep neural networks usually need to build a deeper and more complex network structure to extract abstract semantic features and other information, and often require a certain number of samples for training and adjusting model parameters. At the same time, when training the model, it is also necessary to consider regularization parameters and select numerical optimizers, and constantly adjust parameters according to data, which is also a cumbersome process.
[0003] At present, in order to solve the problem of slow speed of model training and testing, a hyperspectral image classification method based on FPGA deep edge filter is disclosed, which mainly solves the problems of slow classification speed, high power consumption and low overall classification accuracy of hyperspectral image in the prior art. The specific steps of the above method are as follows: 1) input the hyperspectral image to be classified; 2) preprocess the hyperspectral image to be classified; 3) perform dimensionality reduction operation on the preprocessed hyperspectral image; 4) perform deep edge filtering with deep edge filter; 5) generate training set and test set; 6) construct a hyperspectral image classification network; 7) train the hyperspectral image classification network; 8) classify the test set. Although the above hyperspectral image classification method solves the problems of slow classification speed, high power consumption and low overall classification accuracy, the above hyperspectral image classification method does not take into account the nonlinearity of hyperspectral, resulting in too low Kappa index and low classification accuracy, which cannot meet the needs of users. Summary of the invention
[0004] In order to overcome the defects of low hyperspectral image classification accuracy in the above-mentioned prior art, the present invention provides a hyperspectral image classification method and system based on PCANet.
[0005] In order to achieve the above invention purpose, the technical means adopted are:
[0006] A hyperspectral image classification method based on PCANet, the steps of which include:
[0007] Step 100: Acquire a hyperspectral image;
[0008] Step 200: preprocessing the hyperspectral image to obtain a training data set and a test data set;
[0009] Step 300: constructing a hyperspectral image classification model based on PCANet, inputting a training data set into the hyperspectral image classification model for training, and obtaining a trained hyperspectral image classification model;
[0010] The hyperspectral image classification model comprises a first layer of convolution kernels, a nonlinear function, a second layer of convolution kernels and a classifier connected in sequence; the training data set is input into the hyperspectral image classification model, and the first layer of convolution kernels is trained by extracting and analyzing image features of the training data set; the second layer of convolution kernels is trained after image features are extracted and analyzed on feature maps obtained by the first layer of convolution kernels and the nonlinear function in sequence; the classifier is trained after analyzing feature maps obtained by the first layer of convolution kernels, the nonlinear function and the second layer of convolution kernels in sequence;
[0011] Step 400: Put the test data set into the trained hyperspectral image classification model to obtain the hyperspectral image classification result.
[0012] The present invention also provides a hyperspectral image classification system based on PCANet, comprising:
[0013] A data acquisition module, used for acquiring hyperspectral images;
[0014] A preprocessing module is used to preprocess the hyperspectral image to obtain a training data set and a test data set;
[0015] A hyperspectral image classification module, wherein the hyperspectral image classification module includes a hyperspectral image classification model based on PCANet, and is used to perform hyperspectral image classification on an input hyperspectral image;
[0016] The training data set is input into the hyperspectral image classification model for training. After the trained hyperspectral image classification model is obtained, the test data set is input into the trained hyperspectral image classification model to obtain a classification result.
[0017] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: the hyperspectral image classification model based on PCANet improves the classification accuracy of hyperspectral images by introducing nonlinear functions to address the nonlinearity of hyperspectral imaging caused by light scattering; preprocessing the hyperspectral images to remove minor spectral information can greatly reduce the amount of calculation. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1This is a flow chart of the hyperspectral image classification method based on PCANet in Example 1;
[0019] Figure 2 This is a flow chart of the hyperspectral image classification method based on PCANet in Example 2;
[0020] Figure 3 This is a schematic diagram of a hyperspectral image classification system based on PCANet in Example 3. DETAILED DESCRIPTION
[0021] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;
[0022] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;
[0023] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0024] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0025] Example 1
[0026] This embodiment proposes a hyperspectral image classification method based on PCANet. Figure 1 As shown, the steps include:
[0027] Step 100: Acquire a hyperspectral image;
[0028] Step 200: preprocessing the hyperspectral image to obtain a training data set and a test data set;
[0029] Step 300: constructing a hyperspectral image classification model based on PCANet, inputting a training data set into the hyperspectral image classification model for training, and obtaining a trained hyperspectral image classification model;
[0030] The hyperspectral image classification model comprises a first layer of convolution kernels, a nonlinear function, a second layer of convolution kernels and a classifier connected in sequence; the training data set is input into the hyperspectral image classification model, and the first layer of convolution kernels is trained by extracting and analyzing image features of the training data set; the second layer of convolution kernels is trained after image features are extracted and analyzed on feature maps obtained by the first layer of convolution kernels and the nonlinear function in sequence; the classifier is trained after analyzing feature maps obtained by the first layer of convolution kernels, the nonlinear function and the second layer of convolution kernels in sequence;
[0031] Step 400: Put the test data set into the trained hyperspectral image classification model to obtain the hyperspectral image classification result.
[0032] In this embodiment, due to the non-linearity existing in hyperspectral imaging caused by light scattering, a non-linear function is introduced to improve the hyperspectral classification accuracy; preprocessing the hyperspectral data to remove secondary spectral information can greatly reduce the computational load.
[0033] Embodiment 2
[0034] This embodiment proposes a hyperspectral image classification method based on PCANet, as Figure 2 shown, and its steps include:
[0035] Step 100: Obtain a hyperspectral image;
[0036] Step 200: Preprocess the hyperspectral image to obtain a training data set and a test data set. The specific steps are as follows:
[0037] Step 210: Perform a dimensionality reduction operation on the hyperspectral image to obtain a hyperspectral main image; the dimensionality reduction operation in this embodiment can reduce unnecessary spectral information, thereby reducing the calculation and further simplifying the process of hyperspectral image classification. The specific dimensionality reduction operation implemented in this embodiment is as follows:
[0038] Step 211: Obtain the size w1×h1×C of the hyperspectral image, where w1 is the number of rows in the hyperspectral image matrix, h1 is the number of columns in the hyperspectral image matrix, and C is the number of spectral layers of the hyperspectral image;
[0039] Step 212: Use the principal components analysis (PCA) algorithm to perform dimensionality reduction on the spectral dimension of the hyperspectral image to obtain the eigenvalues of each layer of the spectrum in the hyperspectral image;
[0040] Step 213: Retain the k layers of spectra with the largest eigenvalues to form a hyperspectral main image, where k < C. In this embodiment, k is taken as 5, that is, retain the 5 layers of spectra with the largest eigenvalues. In actual operation, the number of retained layers can be determined according to experimental debugging.
[0041] Step 220: For each pixel point in the hyperspectral main image, form a pixel point image block with its adjacent pixel points;
[0042] Step 230: Mark the pixel point image blocks according to the ground object categories;
[0043] Step 240: Eliminate the unmarked pixel point image blocks to obtain a hyperspectral data set;
[0044] Step 250: Randomly divide the hyperspectral data set into a training data set and a test data set. In this embodiment, a certain number of pixel image blocks are first extracted from each type of land object in the hyperspectral data set, and then a certain proportion of pixel image blocks are randomly extracted from the remaining pixel image blocks. Finally, the pixel image blocks extracted twice are used as the training data set, and the remaining pixel image blocks are used as the test data set. This sampling method can ensure that pixel image blocks of each type of land object are extracted. In this embodiment, a total of 10% of the pixel image blocks are extracted as the training data set.
[0045] Step 300: construct a hyperspectral image classification model based on PCANet, input the training data set into the hyperspectral image classification model for training, and obtain a trained hyperspectral image classification model. The specific steps are as follows:
[0046] Step 310: extracting image features of the training data set through the window, and calculating the first layer convolution kernel of the hyperspectral image classification model according to the image features; the process of calculating the first layer convolution kernel in this embodiment is as follows:
[0047] Step 311: Obtain the size of the training data set w2×h2×k, and obtain the number of training data in the training data set N1, where w2 is the row in the training data set matrix, h2 is the column in the training data set matrix, and k is the number of spectral layers of the training data set;
[0048] Step 312: Set a window of size k1×k2 for the image layer of the cth channel in the ith training data in the training data set, extract N2 image features by sliding the window, and splice the extracted image features by column to obtain the window set of the cth channel in the ith training data. in, represents the N2th window in the cth channel of the ith training data, i=1,…,N1, c=1,…,k, N2=(w2-k1+1)×(h2-k2+1), k1 and k2 are set manually;
[0049] Step 313: Calculate the window average value of each channel in each training data;
[0050] Step 314: Subtract the window average value of the channel in which each window in the cth channel of the ith training data is located from the window average value of the channel in which it is located to obtain the window average value of the cth channel of the ith training data.
[0051] Step 315: Concatenate the mean values of each channel window of the i-th training data row by row to obtain the i-th training data set.
[0052] Step 316: Each training data set Splice by column to get the total set of training data
[0053] Step 317: Calculate the matrix XX T The eigenvalues and eigenvectors of , and the eigenvectors corresponding to the largest first L1 eigenvalues are retained to obtain the first layer of convolution kernel, which is expressed as follows:
[0054]
[0055] Among them, q l () means retaining the eigenvector corresponding to the largest first L1 eigenvalue and obtaining the convolution kernel The dimensions are L1×k1×k2×k.
[0056] Step 320: Convolve the training data set with the first layer convolution kernel and input it into the nonlinear function to obtain the nonlinear first layer feature map. The nonlinear function includes f n (), relu(), sigmoid(), the nonlinear function in this embodiment is f n (). In order to reduce the influence of light scattering on the accuracy of hyperspectral classification and further improve the classification accuracy, this embodiment firstly inputs the feature map obtained by convolving the training data set with the first layer of convolution kernel into the activation function to obtain the activated feature map, and finally introduces the shortcut connection. The specific operation is as follows:
[0057] The expression for convolution of the training data set with the first layer of convolution kernel is:
[0058]
[0059] Among them, I (patch) represents the training data set, Represents the feature map obtained after convolution;
[0060] Will Input to the activation function, its expression is as follows:
[0061]
[0062] in, Represents the activated feature map;
[0063] Finally, we introduce a shortcut connection. Adding it to the training data set gives the nonlinear first-layer feature map, which is expressed as follows:
[0064]
[0065] in, Represents the first layer feature map.
[0066] Step 330: Use the window to extract the image features of the first layer feature map, and calculate the second layer convolution kernel of the hyperspectral image classification model through the image features of the first layer feature map; the process of calculating the second layer convolution kernel in this embodiment is as follows:
[0067] Step 331: Adjust the size of the first layer feature map N1×w2×h2×L1 to (N1×L1)×w2×h2 to obtain a secondary training data set;
[0068] Step 332: Set a window of size k1×k2 for the zth training data in the training data set, extract w2×h2 image features of the training data by sliding the window, and splice the image features of the training data by column to obtain the window set of the zth training data. Among them, y z,1 represents the first window in the z-th training data, z=1,…,N1×L1;
[0069] Step 333: Calculate the window average of each training data;
[0070] Step 334: Subtract the window average value from each window of the z-th training data to obtain the window average value of the z-th training data.
[0071] Step 335: Concatenate the window means of each training data row by row to obtain the total set of training data.
[0072] Step 336: Calculate the matrix YY T The eigenvalues and eigenvectors of , and the eigenvector corresponding to the largest first L2 eigenvalue is retained to obtain the second layer of convolution kernel, which is expressed as follows:
[0073]
[0074] Among them, q l () means retaining the eigenvector corresponding to the largest first L2 eigenvalue and obtaining the second layer convolution kernel The dimensions are L2×k1×k2.
[0075] Step 340: Convolve the first layer feature map with the second layer convolution kernel to obtain the second layer feature map, which is expressed as:
[0076]
[0077] in, Represents the second layer feature map.
[0078] Step 350: Binarize and group the second layer feature map to obtain feature map blocks. The specific steps are as follows:
[0079] Step 351: Convert the L2 binary bit vector corresponding to the feature map obtained after the binarization operation into a decimal image to obtain a decimal feature map, which is expressed as:
[0080]
[0081] Among them, H refers to the binarization operation, that is, setting the values greater than 0 to 1 and the other values to 0; Represents the feature map after binarization.
[0082] Step 352: Divide the decimal feature map into blocks and groups to obtain feature map blocks, which are expressed as follows:
[0083]
[0084] in, It refers to the first feature map block of the l1th layer channel in the l2th decimal feature map. In this embodiment, the decimal feature map is divided into blocks of size B×B, that is, the stride in the above expression is set to B.
[0085] Step 360: Count the decimal histograms in each feature block, and connect all the histograms to form a histogram statistics vector Its expression is as follows:
[0086]
[0087] Step 370: input the histogram statistical vector into the classifier for training to obtain a trained hyperspectral image classification model.
[0088] Step 400: Put the test data set into the trained hyperspectral image classification model to obtain the hyperspectral image classification result. In this embodiment, the step of putting the test data set into the trained hyperspectral image classification model includes: after the test data set is input into the trained hyperspectral image classification model and extracted into the test feature graph, the test feature graph is binarized, grouped and histogram-counted to generate a histogram statistical vector; the histogram statistical vector is input into the classifier to obtain the prediction result.
[0089] In this embodiment, the technicians conducted tests on the Indian pines dataset, and the specific experimental results are as follows:
[0090]
[0091]
[0092] As shown in the table above, OA refers to the overall classification accuracy, which is mainly calculated by the number of correctly classified test samples / total number of test samples. AA refers to the average classification accuracy, which is mainly obtained by averaging the classification accuracy of different categories. KA refers to the Kappa coefficient, which is mainly used for consistency testing to measure the bias of the classification method. The higher the bias, the lower the score. Among them, the classification results of the hyperspectral image classification method based on PCANet in this embodiment are located in the ResPCANet column.
[0093] It can be seen from the above table that the hyperspectral image classification method based on PCANet in this embodiment has its own advantages compared with the classification results of the convolutional neural network HybridSN; the classification results of the hyperspectral image classification method based on PCANet in this embodiment are obviously better than those of the other classification methods.
[0094] Example 3
[0095] This embodiment proposes a hyperspectral image classification system based on PCANet, including:
[0096] A data acquisition module, used for acquiring hyperspectral images;
[0097] A preprocessing module is used to preprocess the hyperspectral image to obtain a training data set and a test data set;
[0098] A hyperspectral image classification module, wherein the hyperspectral image classification module includes a hyperspectral image classification model based on PCANet, and is used to perform hyperspectral image classification on an input hyperspectral image;
[0099] The training data set is input into the hyperspectral image classification model for training. After the trained hyperspectral image classification model is obtained, the test data set is input into the trained hyperspectral image classification model to obtain a classification result.
[0100] In this embodiment, a hyperspectral image is acquired through a data acquisition module, and then the hyperspectral image is transmitted to a preprocessing module. In the preprocessing module, the hyperspectral image undergoes dimensionality reduction and labeling to obtain a training data set and a test data set. Next, the training data set is input into a hyperspectral image classification module. After the training data set passes through the first layer of convolution kernels, nonlinear functions, and the second layer of convolution kernels, a hyperspectral image classification module that has completed training can be obtained. Finally, the test data set is input into the hyperspectral image classification module to obtain a classification result.
[0101] The same or similar reference numerals correspond to the same or similar components;
[0102] The terms used in the drawings to describe positional relationships are only used for illustrative purposes and should not be construed as limiting this patent;
[0103] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. A hyperspectral image classification method based on PCANet, characterized in that: The steps include: Step 100: Acquire a hyperspectral image; Step 200: preprocessing the hyperspectral image to obtain a training data set and a test data set; Step 300: constructing a hyperspectral image classification model based on PCANet, inputting a training data set into the hyperspectral image classification model for training, and obtaining a trained hyperspectral image classification model; The hyperspectral image classification model comprises a first layer of convolution kernels, a nonlinear function, a second layer of convolution kernels and a classifier connected in sequence; the training data set is input into the hyperspectral image classification model, and the first layer of convolution kernels is trained by extracting and analyzing image features of the training data set; the second layer of convolution kernels is trained after image features are extracted and analyzed on feature maps obtained by the first layer of convolution kernels and the nonlinear function in sequence; the classifier is trained after analyzing feature maps obtained by the first layer of convolution kernels, the nonlinear function and the second layer of convolution kernels in sequence; Step 400: Put the test data set into the trained hyperspectral image classification model to obtain the hyperspectral image classification result.
2. The hyperspectral image classification method based on PCANet according to claim 1, characterized in that: In step 200, the step of preprocessing the hyperspectral image includes: Step 210: performing a dimensionality reduction operation on the hyperspectral image to obtain a hyperspectral main image; Step 220: Each pixel point in the hyperspectral main image and its adjacent pixels form a pixel point image block; Step 230: marking the pixel image blocks according to the ground object category; Step 240: Eliminate unlabeled pixel image blocks to obtain a hyperspectral data set; Step 250: randomly divide the hyperspectral dataset into a training dataset and a test dataset.
3. The hyperspectral image classification method based on PCANet according to claim 2, characterized in that: In step 210, the step of performing a dimensionality reduction operation on the hyperspectral image includes: Step 211: using a principal component analysis algorithm to reduce the spectral dimension of the hyperspectral image to obtain characteristic values of each layer of the spectrum in the hyperspectral image; Step 212: retain the k-layer spectra with the largest eigenvalues to form a hyperspectral main image.
4. The hyperspectral image classification method based on PCANet according to claim 1, characterized in that: In step 300, the step of inputting the training data set into the hyperspectral image classification model for training includes: Step 310: extracting image features of the training data set through the window, and calculating the first layer convolution kernel of the hyperspectral image classification model according to the image features; Step 320: convolve the training data set with the first layer convolution kernel and input the convolution result into a nonlinear function to obtain a nonlinear first layer feature map; Step 330: extracting image features of the first layer feature map using a window, and calculating the second layer convolution kernel of the hyperspectral image classification model through the image features of the first layer feature map; Step 340: Convolve the first layer feature map with the second layer convolution kernel to obtain a second layer feature map; Step 350: Binarize and group the second layer feature map to obtain feature map blocks; Step 360: Count the decimal histograms in each feature block, and connect all the histograms to form a histogram statistics vector; Step 370: input the histogram statistical vector into the classifier for training to obtain a trained hyperspectral image classification model.
5. The hyperspectral image classification method based on PCANet according to claim 4 is characterized in that: In step 310, the steps of extracting image features of the training data set through a window and calculating the first layer of convolution kernel according to the image features include: Step 311: Obtain the size of the training data set w2×h2×k, and obtain the number of training data in the training data set N1, where w2 is the row in the training data set matrix, h2 is the column in the training data set matrix, and k is the number of spectral layers of the training data set; Step 312: Set a window of size k1×k2 for the image layer of the cth channel in the ith training data in the training data set, extract N2 image features by sliding the window, and splice the extracted image features by column to obtain the window set of the cth channel in the ith training data. in, represents the N2th window in the cth channel of the i-th training data, i = 1, ..., N1, c = 1, ..., k, N2 = (w2-k1+1) × (h2-k2+1); Step 313: Calculate the window average of each channel in each training data; Step 314: Subtract the window average value of the channel in which each window in the cth channel of the ith training data is located from the window average value of the channel in which it is located to obtain the window average value of the cth channel of the ith training data. Step 315: Concatenate the mean values of each channel window of the i-th training data row by row to obtain the i-th training data set. Step 316: Each training data set Splice by column to get the total set of training data Step 317: Calculate the matrix XX T The eigenvalues and eigenvectors of , and the eigenvectors corresponding to the largest first L1 eigenvalues are retained to obtain the first layer of convolution kernel, which is expressed as follows: Among them, q l () means retaining the eigenvector corresponding to the largest first L1 eigenvalue and obtaining the convolution kernel The dimensions are L1×k1×k2×k.
6. The hyperspectral image classification method based on PCANet according to claim 4, characterized in that: In step 320, the step of inputting a nonlinear function includes the operations of inputting an activation function and introducing a shortcut connection; Among them, the feature map obtained by the activation function is expressed as The training data set is denoted as I (patch) , the first-layer feature map obtained after the introduction of the quick connection is expressed as The expression of the shortcut connection is as follows:
7. The hyperspectral image classification method based on PCANet according to claim 6, characterized in that: In step 330, the steps of extracting the image features of the first layer feature map by using a window and obtaining the second layer convolution kernel by calculating the image features of the first layer feature map include: Step 331: Adjust the size of the first layer feature map N1×w2×h2×L1 to (N1×L1)×w2×h2 to obtain a secondary training data set; Step 332: Set a window of size k1×k2 for the zth training data in the training data set, extract w2×h2 image features of the training data by sliding the window, and splice the image features of the training data by column to obtain the window set of the zth training data. Among them, y z,1 represents the first window in the z-th training data, z=1,…,N1×L1; Step 333: Calculate the window average of each training data; Step 334: Subtract the window average value from each window of the z-th training data to obtain the window average value of the z-th training data. Step 335: Concatenate the window means of each training data row by row to obtain the total set of training data. Step 336: Calculate the matrix YY T The eigenvalues and eigenvectors of , and the eigenvector corresponding to the largest first L2 eigenvalue is retained to obtain the second layer of convolution kernel, which is expressed as follows: Among them, q l () means retaining the eigenvector corresponding to the largest first L2 eigenvalue, and obtaining the second layer convolution kernel W l2 The dimensions are L2×k1×k2.
8. The hyperspectral image classification method based on PCANet according to claim 7, characterized in that: In step 350, the step of grouping the second layer feature map into blocks includes: Step 351: convert the L2 binary bit vector corresponding to the feature map obtained after the binarization operation into a decimal image to obtain a decimal feature map; Step 352: Divide the decimal feature map into blocks and groups to obtain feature map blocks, which are expressed as follows: in, Refers to the first feature map block of the l1th layer channel in the l2th decimal feature map.
9. The hyperspectral image classification method based on PCANet according to any one of claims 1 to 8, characterized in that: In step 400, the step of placing the test data set into the trained hyperspectral image classification model includes: after the test data set is input into the trained hyperspectral image classification model and extracted into a test feature graph, the test feature graph is binarized, grouped and histogram-counted to generate a histogram statistical vector; and the histogram statistical vector is input into a classifier to obtain a prediction result.
10. A hyperspectral image classification system based on PCANet, applied to the hyperspectral image classification method based on PCANet according to any one of claims 1 to 9, characterized in that: include: A data acquisition module, used for acquiring hyperspectral images; A preprocessing module is used to preprocess the hyperspectral image to obtain a training data set and a test data set; A hyperspectral image classification module, wherein the hyperspectral image classification module includes a hyperspectral image classification model based on PCANet, and is used to perform hyperspectral image classification on an input hyperspectral image; The training data set is input into the hyperspectral image classification model for training. After the trained hyperspectral image classification model is obtained, the test data set is input into the trained hyperspectral image classification model to obtain a classification result.