CN110852227A - Hyperspectral image deep learning classification method, device, equipment and storage medium - Google Patents

Hyperspectral image deep learning classification method, device, equipment and storage medium Download PDF

Info

Publication number
CN110852227A
CN110852227A CN201911064083.XA CN201911064083A CN110852227A CN 110852227 A CN110852227 A CN 110852227A CN 201911064083 A CN201911064083 A CN 201911064083A CN 110852227 A CN110852227 A CN 110852227A
Authority
CN
China
Prior art keywords
classification
sample set
hyperspectral
image
deep learning
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201911064083.XA
Other languages
Chinese (zh)
Inventor
张霞
王楠
黄长平
岑奕
戚文超
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Institute of Remote Sensing and Digital Earth of CAS
Original Assignee
Institute of Remote Sensing and Digital Earth of CAS
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Institute of Remote Sensing and Digital Earth of CAS filed Critical Institute of Remote Sensing and Digital Earth of CAS
Priority to CN201911064083.XA priority Critical patent/CN110852227A/en
Publication of CN110852227A publication Critical patent/CN110852227A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB

Abstract

The invention relates to the technical field of hyperspectral image classification, and discloses a hyperspectral image deep learning classification method, device, equipment and storage medium, which are used for improving the accuracy and efficiency of hyperspectral image classification. The method comprises the following steps: acquiring a hyperspectral image to be classified; randomly cutting the hyperspectral images to be classified according to the size of a preset window and the marked sample set to obtain a sample set to be trained; expanding the data set through image transformation to obtain a corresponding deep learning sample set; extracting space spectrum features by adopting a convolution cyclic neural network and a three-dimensional convolution neural network; and classifying the hyperspectral images through a preset neural network classification model obtained through training to obtain a corresponding image classification result. By constructing the deep neural network model, deep abstract features of the hyperspectral image can be automatically extracted, the workload of manual extraction and feature optimization is effectively reduced, and end-to-end automatic identification and classification of the hyperspectral image are realized.

Description

Hyperspectral image deep learning classification method, device, equipment and storage medium
Technical Field
The invention relates to the technical field of hyperspectral image classification, in particular to a hyperspectral image deep learning classification method, device, equipment and storage medium.
Background
The hyperspectral image classification is used for endowing each pixel with a determined class label, and is a crucial technical means in the field of image information processing and analysis. The spectrum characteristics and the spatial characteristics that hundreds of continuous spectrum sections contain in the hyperspectral image can effectively promote classification precision and improve classification effect, and become indispensable important information in the hyperspectral image classification process. However, the dimension of the hyperspectral image is high, and the available labeled samples are few, so that the hyperspectral image classification still faces great challenges.
Therefore, how to efficiently and intelligently extract and utilize abundant spatial features and spectral features of a hyperspectral image based on limited labeled samples is a hotspot of research in the field of hyperspectral image processing and application.
The traditional high-spectral-resolution remote sensing image classification mainly comprises the steps of constructing an image classification feature library to obtain representative classification features, then optimizing the feature library, and training a classifier by utilizing the optimized features. However, the traditional hyperspectral image classification method is time-consuming and labor-consuming, the generalization capability of the classifier is weak, and the method cannot be rapidly and effectively popularized to different scenes for hyperspectral image classification, so that the traditional hyperspectral image classification method has the problems of low image classification accuracy and low efficiency.
Disclosure of Invention
The invention mainly aims to solve the problems of low image classification accuracy and low efficiency of the existing hyperspectral image classification method.
In order to achieve the above object, a first aspect of the present invention provides a hyperspectral image deep learning classification method, including:
acquiring a hyperspectral image to be classified;
randomly cutting the hyperspectral images to be classified according to the size of a preset window to obtain a corresponding data set;
performing data expansion on the data set in an image transformation mode to obtain a corresponding sample set to be classified;
performing classification label prediction on the hyperspectral images to be classified through a preset neural network classification model obtained through training, and filling the obtained classification labels into a matrix with the same size as the hyperspectral images to be classified to obtain image classification results corresponding to the hyperspectral images to be classified;
and evaluating the image classification result by adopting a preset classification evaluation index to obtain a hyperspectral image classification result and classification precision meeting the requirement of the corresponding evaluation index.
Optionally, in another embodiment of the first aspect of the present invention, before the randomly clipping the hyperspectral image to be classified according to the size of a preset window to obtain a corresponding data set, the method includes:
the method comprises the steps of obtaining a marked sample set with a preset data scale, marking hyperspectral images in the marked sample set according to classification labels, and dividing the hyperspectral marked sample set into a sample set A and a sample set B according to a preset sample proportion through hierarchical random sampling, wherein the sample set B is used as a test set;
according to the size of a preset window and the marked sample set A, randomly cutting the hyperspectral image to obtain a corresponding sample set to be trained;
and performing data expansion on the sample set to be trained in an image transformation mode to obtain a corresponding deep learning sample set.
Optionally, in another embodiment of the first aspect of the present invention, after performing data expansion on the sample set to be trained in an image transformation manner to obtain a corresponding deep learning sample set, the method includes:
dividing the deep learning sample set into a training set and a verification set by adopting random sampling; the training set is used for training and constructing the preset neural network classification model, the verification set is used for adjusting and optimizing training parameters of the preset neural network classification model, and the test set is used for testing the precision and the robustness of the preset neural network classification model.
Optionally, in another embodiment of the first aspect of the present invention, after performing data expansion on the to-be-trained sample set by using an image transformation manner to obtain a corresponding deep learning sample set, the method further includes:
inputting the deep learning sample set into a three-dimensional convolution residual error network model, and respectively constructing a spatial feature extraction network branch and a spectral feature extraction network branch by adopting residual error connection with different step lengths;
setting spatial dimension hyper-parameters of a three-dimensional convolution kernel in the spatial feature extraction network branch, extracting spatial features of the hyper-spectral images in the deep learning sample set in a parallel mode, and performing wave band superposition on the obtained spatial features;
setting spectral dimensional hyper-parameters of a three-dimensional convolution kernel in the spectral feature extraction network branch, extracting spectral features of the hyperspectral images in the deep learning sample set in a parallel mode, and performing wave band superposition on the obtained spectral features;
extracting the spectrum time sequence characteristics of the hyperspectral images in the deep learning sample set by adopting a convolution cyclic neural network model;
inputting the time sequence features, the superposition space features and the superposition spectrum features into an average pooling layer of the three-dimensional convolution residual error network model by using a cascading strategy for feature fusion to obtain a feature-fused hyperspectral image;
and automatically identifying and classifying the hyperspectral images with the fused features by utilizing a softmax function, and obtaining a constructed neural network classification model, wherein the construction of the neural network classification model is based on the three-dimensional convolution residual error network model and the convolution cyclic neural network model.
Optionally, in another embodiment of the first aspect of the present invention, after the automatically identifying and classifying the hyperspectral image of the feature fusion by using the softmax function and obtaining the constructed neural network classification model, the method further includes:
and carrying out algorithm iteration on the neural network classification model until the repetition times reach a preset iteration time or the classification precision reaches a preset precision requirement, and obtaining a corresponding preset neural network classification model.
Optionally, in another embodiment of the first aspect of the present invention, after the setting of the spectral dimensional hyper-parameter of the three-dimensional convolution kernel in the spectral feature extraction network branch, extracting the spectral features of the hyper-spectral images in the deep learning sample set in a parallel manner, and performing band stacking on the obtained spectral features, the method further includes:
and adding a Dropout layer in the three-dimensional convolution residual error network model, randomly discarding some neurons in the three-dimensional convolution residual error network model according to a preset probability, and returning the weight of a hidden layer or an output layer to zero.
Optionally, in another embodiment of the first aspect of the present invention, after the setting of the spectral dimensional hyper-parameter of the three-dimensional convolution kernel in the spectral feature extraction network branch, extracting the spectral features of the hyper-spectral images in the deep learning sample set in a parallel manner, and performing band stacking on the obtained spectral features, the method further includes:
and adopting an Adam optimizer in the three-dimensional convolution residual error network model, setting an initial learning rate in the Adam optimizer, and enabling the learning rate in the Adam optimizer to realize dynamic change so as to realize self-adaptive optimization of the space dimensional hyper-parameter and the spectrum dimensional hyper-parameter.
The invention provides a hyperspectral image deep learning classification device in a second aspect, which comprises:
the image to be classified acquisition module is used for acquiring a hyperspectral image to be classified;
the data set acquisition module is used for randomly cutting the hyperspectral images to be classified according to the size of a preset window to obtain a corresponding data set;
the to-be-classified sample set acquisition module is used for performing data expansion on the data set in an image transformation mode to obtain a corresponding to-be-classified sample set;
the image classification result acquisition module is used for predicting classification labels of the hyperspectral images in the sample set to be classified through a preset neural network classification model obtained through training, and filling the obtained classification labels into a matrix with the same size as the hyperspectral images to be classified to obtain image classification results corresponding to the hyperspectral images to be classified;
and the evaluation module is used for evaluating the image classification result by adopting a preset classification evaluation index to obtain a hyperspectral image classification result and classification precision meeting the requirement of the corresponding evaluation index.
Optionally, in another embodiment of the second aspect of the present invention, the hyperspectral image deep learning classification apparatus further includes:
the system comprises a marked sample set acquisition module, a high spectrum image analysis module and a high spectrum image analysis module, wherein the marked sample set acquisition module is used for acquiring a marked sample set with a preset data scale, the high spectrum image in the marked sample set is marked according to a classification label, and the high spectrum marked sample set is divided into a sample set A and a sample set B according to a preset sample proportion through hierarchical random sampling, wherein the sample set B is used as a test set;
the hyperspectral image acquisition module is used for randomly cutting the hyperspectral image according to the size of a preset window and the marked sample set A to obtain a corresponding sample set to be trained;
and the deep learning sample set acquisition module is used for performing data expansion on the sample set to be trained in an image transformation mode to obtain a corresponding deep learning sample set.
Optionally, in another embodiment of the second aspect of the present invention, the hyperspectral image deep learning classification apparatus further includes:
the deep learning sample set dividing module is used for dividing the deep learning sample set into a training set and a verification set by adopting random sampling; the training set is used for training and constructing the preset neural network classification model, the verification set is used for adjusting and optimizing training parameters of the preset neural network classification model, and the test set is used for testing the precision and the robustness of the preset neural network classification model.
Optionally, in another embodiment of the second aspect of the present invention, the hyperspectral image deep learning classification apparatus further includes:
the characteristic extraction branch building module is used for inputting the deep learning sample set into a three-dimensional convolution residual error network model and respectively building a spatial characteristic extraction network branch and a spectral characteristic extraction network branch by adopting residual error connection with different step lengths;
the spatial feature acquisition module is used for extracting spatial features of the hyperspectral images in the deep learning sample set in a parallel mode by setting spatial dimension hyper-parameters of a three-dimensional convolution kernel in the spatial feature extraction network branch and performing wave band superposition on the obtained spatial features;
the spectral feature acquisition module is used for extracting the spectral features of the hyperspectral images in the deep learning sample set in a parallel mode by setting spectral dimensional hyper-parameters of a three-dimensional convolution kernel in the spectral feature extraction network branch and carrying out wave band superposition on the obtained spectral features;
the spectrum time sequence feature acquisition module is used for extracting the spectrum time sequence features of the hyperspectral images in the deep learning sample set by adopting a convolution cyclic neural network model;
the characteristic fusion module is used for inputting the time sequence characteristics, the superposition space characteristics and the superposition spectrum characteristics into an average pooling layer of the three-dimensional convolution residual error network model by using a cascade strategy for characteristic fusion to obtain a characteristic-fused hyperspectral image;
and the neural network classification model building module is used for automatically identifying and classifying the hyperspectral image with the fused features by utilizing a softmax function and obtaining a built neural network classification model, and the building of the neural network classification model is based on the three-dimensional convolution residual error network model and the convolution cyclic neural network model.
Optionally, in another embodiment of the second aspect of the present invention, the hyperspectral image deep learning classification apparatus further includes:
and the preset neural network classification model obtaining module is used for carrying out algorithm iteration on the neural network classification model until the repetition times reach the preset iteration times or the classification precision reaches the preset precision requirement, so as to obtain the corresponding preset neural network classification model.
Optionally, in another embodiment of the second aspect of the present invention, the hyperspectral image deep learning classification apparatus further includes:
and the Dropout layer adding module is used for adding a Dropout layer in the three-dimensional convolution residual error network model, randomly discarding some neurons in the three-dimensional convolution residual error network model according to a preset probability, and zeroing the weight of a hidden layer or an output layer.
Optionally, in another embodiment of the second aspect of the present invention, the hyperspectral image deep learning classification apparatus further includes:
and the learning rate dynamic setting module is used for adopting an Adam optimizer in the three-dimensional convolution residual error network model, setting an initial learning rate in the Adam optimizer, and enabling the learning rate in the Adam optimizer to realize dynamic change so as to realize self-adaptive optimization of the space dimension hyperparameter and the spectrum dimension hyperparameter.
The third aspect of the present invention provides a hyperspectral image deep learning classification device, including: a memory having instructions stored therein and at least one processor, the memory and the at least one processor interconnected by a line; the at least one processor invokes the instructions in the memory to cause the hyper-spectral image deep learning classification apparatus to perform the method of the first aspect.
A fourth aspect of the present invention provides a computer-readable storage medium having stored therein instructions which, when run on a computer, cause the computer to perform the method of the first aspect described above.
According to the technical scheme provided by the invention, hyperspectral images to be classified are obtained; the method comprises the steps of obtaining a marked sample set with a preset data scale, marking hyperspectral images in the marked sample set according to classification labels, dividing the hyperspectral images into a sample set A and a sample set B according to a preset sample proportion through layering random sampling, wherein the sample set B is used as a test set, and randomly cutting the hyperspectral images to be classified according to the size of a preset window and the marked sample set A to obtain a sample set to be trained; performing data expansion on a sample set to be trained in an image transformation mode to obtain a corresponding deep learning sample set; automatically identifying and classifying hyperspectral images to be classified through a preset neural network classification model obtained through training to obtain corresponding image classification results; and evaluating the image classification result by adopting a preset classification evaluation index to obtain a hyperspectral image classification result and classification precision meeting the requirement of the corresponding evaluation index. According to the hyperspectral image classification method and device, the hyperspectral images to be classified are subjected to data preprocessing of random cutting and data expansion, the preset neural network classification model is input for automatic identification and classification, corresponding image classification results are obtained, the image classification results are evaluated through preset classification evaluation indexes, the hyperspectral image classification results and classification precision meeting the requirements of the corresponding evaluation indexes are obtained, and therefore the accuracy and the efficiency of hyperspectral image classification are improved.
Drawings
FIG. 1 is a schematic diagram of an embodiment of a hyperspectral image deep learning classification method in the embodiment of the invention;
FIG. 2 is a schematic diagram of another embodiment of a hyperspectral image deep learning classification method in the embodiment of the invention;
FIG. 3 is a schematic diagram of an embodiment of a hyperspectral image deep learning classification device in the embodiment of the invention;
FIG. 4 is a schematic diagram of another embodiment of a hyperspectral image deep learning classification device in the embodiment of the invention;
FIG. 5 is a schematic diagram of an embodiment of a hyperspectral image deep learning classification device in the embodiment of the invention.
Detailed Description
The embodiment of the invention provides a hyperspectral image deep learning classification method, device and equipment and a storage medium, which are used for improving the accuracy and efficiency of hyperspectral image classification.
In order to make the technical field of the invention better understand the scheme of the invention, the embodiment of the invention will be described in conjunction with the attached drawings in the embodiment of the invention.
The terms "first," "second," "third," "fourth," and the like in the description and in the claims, as well as in the drawings, if any, are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It will be appreciated that the data so used may be interchanged under appropriate circumstances such that the embodiments described herein may be practiced otherwise than as specifically illustrated or described herein. Furthermore, the terms "comprises," "comprising," or "having," and any variations thereof, are intended to cover non-exclusive inclusions, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.
Example one
For convenience of understanding, a specific flow of an embodiment of the present invention is described below, and referring to fig. 1, an embodiment of a hyperspectral image deep learning classification method in an embodiment of the present invention includes:
101. and acquiring the hyperspectral images to be classified.
Specifically, the server acquires hyperspectral images to be classified, for example, points are displayed on a Google map according to the GPS coordinates of field sample points, then the points are used as references, and the points are selected at corresponding positions of the hyperspectral images to be classified so as to obtain a known label sample set of original hyperspectral images to be classified. Wherein, the marked image pixel represents that the ground object type is known, and the unmarked image pixel represents that the ground object type is unknown.
102. And randomly cutting the hyperspectral images to be classified according to the size of a preset window to obtain a corresponding data set.
Specifically, assuming that the size of the hyperspectral image three-dimensional block to be classified is P × Q × B and the size of the preset window to be randomly cut is h × h, the server cuts the original hyperspectral image by using a certain pixel in the hyperspectral image to be classified as a center and adopting a random cutting mode to obtain a group of corresponding data sets, wherein the data sets comprise the hyperspectral images of which the sizes are h × h × B.
103. And performing data expansion on the data set in an image transformation mode to obtain a corresponding sample set to be classified.
Specifically, the server performs data expansion by using image transformation modes such as rotation, blurring and noise addition based on the cut hyperspectral image, and the expanded sample set to be classified can relieve the contradiction between high dimensionality of the original hyperspectral image and few training samples, and reduce errors caused by sample imbalance.
104. And performing classification label prediction on the hyperspectral images to be classified through a preset neural network classification model obtained through training, and filling the obtained classification labels into a matrix with the same size as the hyperspectral images to be classified to obtain image classification results corresponding to the hyperspectral images to be classified.
Specifically, the server performs classification label prediction on hyperspectral images in a sample set to be classified through a preset neural network classification model obtained through training based on the expanded sample set to be classified, fills the obtained classification labels into a matrix with the same size as the hyperspectral images to be classified to obtain image classification results corresponding to the hyperspectral images to be classified, and obtains a final hyperspectral image classification graph and classification precision, wherein the preset neural network classification model is an optimal classification model found through algorithm iteration during model training.
105. And evaluating the image classification result by adopting a preset classification evaluation index to obtain a hyperspectral image classification result and classification precision meeting the requirement of the corresponding evaluation index.
Specifically, the server evaluates the image classification result by adopting a preset classification evaluation index to obtain a hyperspectral image classification result and classification precision meeting the requirement of the corresponding evaluation index. The hyperspectral image classification result obtained by the deep learning can be evaluated by adopting the classification evaluation index, so that the image classification result and the classification precision required by the required evaluation index can be obtained. Wherein the classification evaluation index includes: (1) producer Accuracies (PA), which represent drawing accuracies, refer to the proportion of the number of pixels in the entire classification result chart, which are correctly classified into the surface type a, to the total number of pixels in the surface type, which is actually classified into the type a. (2) The total classification precision (OA) represents the proportion of the total number of pixels correctly classified in all surface feature types to the total number of pixels in the real category, i.e. the ratio of the total number of pixels distributed along the diagonal to the total number of pixels in the confusion matrix, and is calculated according to the following calculation formula:
wherein, PkiRepresents the number of samples for which class i is identified as k, N represents the total number of test samples, and m represents the number of classification classes.
(3) The Average of the percentage of correctly classified classes (AA) is calculated as follows:
Figure BDA0002258730590000092
wherein, PkiRepresents the number of samples for which class i is identified as k, N represents the total number of test samples, m represents the number of classification classes:
(4) the Kappa coefficient is obtained by multiplying the sum of the pixels of all the real earth surface categories by the diagonal of a confusion matrix, then subtracting the sum of the product of the total pixels of the real earth surface categories in a certain category and the total pixels to be classified in the category from the sum of all the categories, and then dividing the sum by the square difference of the total pixels to subtract the sum of the product of the total pixels of the real earth surface categories in a certain category and the total pixels to be classified in the category from the sum of all the categories, and the calculation is carried out according to the following calculation formula:
Figure BDA0002258730590000093
wherein K represents a Kappa coefficient, m and N represent the number of classification classes and the total number of samples, respectively, and ppiIs the total number of columns in the i-th class, pliThe total number of rows in which the ith class is located.
The overall classification precision and the Kappa coefficient reflect the classification precision of the whole classification chart, and when the Kappa coefficient is smaller than 0.5 according to the training experience value, the classification precision is judged to be poor, and when the Kappa coefficient is larger than 0.6, the classification precision is judged to be high in the concrete implementation of the invention.
In summary, the hyperspectral image to be classified is subjected to data preprocessing of random clipping and data expansion, then the preset neural network classification model is input for automatic identification and classification, a corresponding image classification result is obtained, the image classification result is evaluated through the preset classification evaluation index, and the hyperspectral image classification result and the classification precision meeting the requirement of the corresponding evaluation index are obtained. According to the hyperspectral image classification method, the hyperspectral images to be classified are sequentially cut according to the size of a preset window from left to right to obtain cubic image blocks with the same size as the size of the hyperspectral images to be classified, the cubic image blocks are input into the trained optimal classification model, image classification labels corresponding to the cubic image blocks are obtained through prediction, the image classification labels are filled into a matrix with the same size as the hyperspectral images to be classified, the final hyperspectral image classification images and classification precision are obtained, and therefore the accuracy and efficiency of hyperspectral image classification are improved.
Example two
On the basis of the above embodiment, referring to fig. 2, another embodiment of the hyperspectral image deep learning classification method in the embodiment of the present invention includes:
101. acquiring a hyperspectral image to be classified;
201. the method comprises the steps of obtaining a marked sample set with a preset data scale, marking hyperspectral images in the marked sample set according to classification labels, and dividing the hyperspectral marked sample set into a sample set A and a sample set B according to a preset sample proportion through hierarchical random sampling, wherein the sample set B is used as a test set;
specifically, the server obtains a labeled sample set of a predetermined data size, and the hyperspectral images in the labeled sample set are labeled according to the classification labels. For example, according to the GPS coordinates of the field sample points, points are spread on a Google map, then the points are used as references, and the training samples are selected at the corresponding positions of the hyperspectral images, wherein the marked image pixels represent that the types of the ground objects are known, and the unmarked image pixels represent that the types of the ground objects are unknown. For example, selecting samples of known classification categories as labeled samples to form labeled sample set STThen the sample set S is markedTCan be expressed as: (x)1,y1),(x2,y2)…,(xN,yN) (ii) a Wherein x isi∈Rm,yi∈{c1,c2,…ci…cn}，ciRepresenting the category, N being the number of marked samples, and N being the number of categories. By means of layered random sampling, dividing a hyperspectral marked sample set into a sample set A and a sample set B according to a preset sample proportion;
202. randomly cutting the hyperspectral images to be classified according to the size of a preset window and the marked sample set A to obtain a corresponding sample set to be trained;
specifically, assuming that the size of the hyperspectral image three-dimensional block in the marked sample set is P × Q × B, the size of the corresponding marked sample set is P × Q, and the size of the preset window to be randomly cut is h × h, the server cuts the marked sample set by using a certain pixel in the marked sample set as a center and adopting a random cutting mode to obtain a sample set to be trained, wherein the sample set to be trained comprises hyperspectral images of which the sizes are h × h × B, namely the hyperspectral images in all the marked sample sets are randomly cut to form a corresponding sample set to be trained.
203. And performing data expansion on the sample set to be trained in an image transformation mode to obtain a corresponding deep learning sample set.
Specifically, the server expands the training sample set by using image transformation modes such as rotation, blurring and noise addition based on the cut sample set to be trained, and the expanded deep learning sample set can relieve the contradiction between high dimensionality of the original hyperspectral image and few training samples, and reduce errors caused by sample imbalance.
Further, the server also divides the deep learning sample set into a training set and a verification set by adopting random sampling. The training set is used for training and constructing the preset neural network classification model, the verification set is used for adjusting and optimizing training parameters of the preset neural network classification model, and the test set is used for testing the precision and the robustness of the preset neural network classification model. Based on the extended deep learning sample set, the extended deep learning sample set is divided into a training set and a verification set by a random sampling method according to a preset division ratio so as to ensure that the distribution of different data sets is similar.
204. And inputting the deep learning sample set into a three-dimensional convolution residual error network model, and respectively constructing a spatial feature extraction network branch and a spectral feature extraction network branch by adopting residual error connection with different step lengths.
Specifically, the server inputs the deep learning sample set into a three-dimensional convolution residual error network model, and adopts residual error connections with different step lengths to respectively construct a spatial feature extraction network branch and a spectral feature extraction network branch. In the specific implementation of the present invention, the three-dimensional convolution formula in the three-dimensional convolution residual error network model is:
Figure BDA0002258730590000111
wherein, Pi×Qi×RiRepresenting the size of the convolution kernel, f (-) represents the activation function,
Figure BDA0002258730590000112
representing the value of a convolution kernel connected with the mth characteristic of the previous layer at (p, q, r); bijIs the bias of the jth feature of the ith layer, and v is the output of the jth layer at (x, y, z).
205. And setting a space dimension hyper-parameter of a three-dimensional convolution kernel in the spatial feature extraction network branch, extracting the spatial features of the hyperspectral images in the deep learning sample set in a parallel mode, and performing wave band superposition on the obtained spatial features.
206. And setting spectral dimensional hyper-parameters of a three-dimensional convolution kernel in the spectral feature extraction network branch, extracting the spectral features of the hyperspectral images in the deep learning sample set in a parallel mode, and performing band superposition on the obtained spectral features.
Specifically, as shown in steps 205 to 206, the server of the present invention sets the hyper-parameters of the three-dimensional convolution kernel in the feature extraction branch of the three-dimensional convolution residual network model to perform feature extraction, and in the deep learning model, the setting of the hyper-parameters is crucial to the training model, and different hyper-parameters affect the convergence rate and the classification accuracy of the deep learning model.
Further, when the method is implemented specifically, the server adds a Dropout layer in the three-dimensional convolution residual error network model, randomly discards some neurons in the three-dimensional convolution residual error network model according to a preset probability, and returns the weight of a hidden layer or an output layer to zero. After a spatial feature extraction branch and a spectral feature extraction branch of the three-dimensional convolution residual error network model are added, a Dropout layer is added, some neurons in the network are randomly discarded according to a certain probability, such as a probability of 0.5, and the weight of a hidden layer or an output layer is reset to zero, so that the effect of comprehensively classifying by adopting different network models is achieved while model parameters are reduced, and overfitting of the model is effectively avoided.
Further, when the method is implemented specifically, the server adopts an Adam optimizer in the three-dimensional convolution residual error network model, sets an initial learning rate in the Adam optimizer, and enables the learning rate in the Adam optimizer to realize dynamic change so as to realize self-adaptive optimization of the space dimension hyperparameter and the spectrum dimension hyperparameter. Namely, an Adam optimizer is used to set an initial learning rate, and then the learning rate dynamic change is used to realize the model parameter self-adaptive optimization, so that the model can automatically adjust the learning rate in different training stages, and the convergence rate of model training is improved.
Specifically, in the present invention, the learning rate is initialized to 0.00001, and then every 5 iterations, the learning rate is attenuated to 1/2 of the previous iteration, and so on, and the model is iterated for 20 times. In addition, due to limited training samples, the deep learning model is easy to generate an overfitting phenomenon, and the dropout layer skill can randomly discard some neurons in the network according to the preset probability of 0.5, so that the effect of comprehensively classifying by adopting different network models is achieved while model parameters are reduced, and overfitting of the model is effectively avoided.
Therefore, the dynamic learning rate and the dropout skill are applied to the deep learning model, so that the model convergence is effectively accelerated and the overfitting of the model is avoided.
207. And extracting the spectrum time sequence characteristics of the hyperspectral images in the deep learning sample set by adopting a convolution cyclic neural network model.
Specifically, the server extracts the spectrum time sequence characteristics of the hyperspectral images in the deep learning sample set by adopting a convolution cyclic neural network model. In specific implementation, the convolution cyclic neural network model is a long-time memory model based on a convolution form, the spectrum of the hyperspectral image is regarded as a time sequence, and the time sequence characteristics of the spectrum are extracted through a repeat point. The long-time and short-time memory model based on the convolution form can be expressed by any one of the following formulas:
Figure BDA0002258730590000121
Figure BDA0002258730590000122
wherein the content of the first and second substances,
Figure BDA0002258730590000124
is a cropped image block;
Figure BDA0002258730590000125
Figure BDA0002258730590000126
and
Figure BDA0002258730590000127
the gate comprises a forgetting gate, an input gate and an output gate;in the state of the hidden layer,
Figure BDA0002258730590000129
and
Figure BDA00022587305900001210
respectively, the expected cell state and the cell state; f is sigmoid function;and "+" is dot product and convolution operation; bf、bi、bcAnd boIs the bias of the respective layers; whfAnd WhiForget gate and input gate weight matrix are set respectively, and the rest is analogized.
208. And inputting the time sequence features, the superposition space features and the superposition spectrum features into an average pooling layer of the three-dimensional convolution residual error network model by using a cascading strategy for feature fusion to obtain a feature-fused hyperspectral image.
Specifically, the server further inputs the time sequence features, the superposition space features and the superposition spectrum features to an average pooling layer of the three-dimensional convolution residual error network model by using a cascading strategy for feature fusion, so as to obtain a feature-fused hyperspectral image. The method adopts a convolution cyclic neural network to extract the time sequence characteristics of the hyperspectral image, and combines the space characteristics and the spectrum characteristics extracted by a residual error network to realize the fusion of different levels of characteristics by using a cascading strategy.
209. And automatically identifying and classifying the hyperspectral images with the fused features by utilizing a softmax function, and obtaining a constructed neural network classification model, wherein the construction of the neural network classification model is based on the three-dimensional convolution residual error network model and the convolution cyclic neural network model.
Specifically, after deep abstract features of the hyperspectral images are extracted by the server and input to the average pooling layer for feature fusion, the automatic identification and classification of the hyperspectral images are carried out by utilizing a softmax function, model parameters are reduced, meanwhile, overfitting of the model is avoided, and a constructed neural network classification model is obtained.
In conclusion, the neural network models with different structures have different image classification performances, three-dimensional convolution residual error networks with different step lengths are connected to form a cascaded spatial feature and spectral feature extraction branch, the long-term and short-term memory model based on the convolution form is used for mainly mining the spectral time sequence features of the hyperspectral image, the deep spatial features and the spectral features which are rich in the hyperspectral image can be extracted by integrating different convolution operations, and finally the deep spatial features and the spectral features are input into a softmax function classifier to perform end-to-end deep learning hyperspectral image classification.
210. And carrying out algorithm iteration on the neural network classification model until the repetition times reach a preset iteration time or the classification precision reaches a preset precision requirement, and obtaining a corresponding preset neural network classification model.
Specifically, the server performs algorithm iteration on the built neural network classification model based on the three-dimensional convolution residual error network model and the convolution circulation neural network model until the repetition times reach a preset iteration time or the classification precision reaches a preset precision requirement, and a final classification model is obtained, namely the preset neural network classification model obtained through training.
104. Performing classification label prediction on the hyperspectral images in the sample set to be classified through a preset neural network classification model obtained through training, and filling the obtained classification labels into a matrix with the same size as the hyperspectral images to be classified to obtain image classification results corresponding to the hyperspectral images to be classified;
105. and evaluating the image classification result by adopting a preset classification evaluation index to obtain a hyperspectral image classification result and classification precision meeting the requirement of the corresponding evaluation index.
The specific implementation of the steps 101, 104, and 105 has been described in detail in the above embodiments, and is not described herein again.
Specifically, the invention further provides a specific implementation description for classifying the satellite hyperspectral images, which is specifically described as follows:
acquiring a hyperspectral data image shot by an Advanced Hyper-Spectral Imager (AHSI) of a GF-5 satellite visible short wave, wherein the Spectral resolution of a visible near red band is 5nm, and the visible near red band has 150 bands; the spectral resolution of the short wave band is 10nm, 180 wave bands exist, 291 wave bands remain after zero values and noise wave bands are removed, the amplitude is 60km, and the spatial resolution is 30 m.
Firstly, based on the acquired hyperspectral image, combining with the GPS coordinates of ground sampling points, uniformly selecting samples of known classes as marked samples, wherein the samples of known classes comprise 12 types of ground objects such as corns, rice, water bodies and the like.
Secondly, setting the ratio of the training sample to the verification sample to be 3:7 aiming at the marked sample, and then enhancing and expanding the sample library by adopting data such as rotation, noise adding and the like.
According to steps 208 to 210 in the second embodiment, a preset neural network classification model is constructed, and model hyper-parameter setting is carried out; setting the iteration number epoch to be 20, setting the batch sample number batch size to be 32, setting the initial learning rate lr to be 0.00001, then, attenuating the learning rate to 1/2 of the last iteration every 5 iterations, and so on; the Dropout layer randomly discards certain neurons in the network with a probability of 0.5.
And repeating the steps 208 to 210 of the second embodiment, and iteratively training the model until the classification precision meets the requirement or the iteration number reaches to epoch of 20.
According to the finally obtained classification result, the classification result of the hyperspectral image deep learning classification method provided by the invention is basically consistent with the artificial actual classification result. The result precision of the classification by using the hyperspectral image classification method provided by the invention is shown in the following table 1.
TABLE 1 Classification result precision Table
Figure BDA0002258730590000151
Wherein PA is producer precision, and C1-C12 correspond to the types of 12 ground objects such as corn, rice, water body and the like.
In the embodiment of the invention, under the condition that the training samples account for 30%, the image classification is carried out on the labeled samples obtained from the GF-5 satellite hyperspectral data images by using the hyperspectral image deep learning classification method provided by the embodiment, so that the overall classification precision of the classification result is 99.37%, the average classification precision is 99.54%, and the Kappa coefficient is 0.9927. Based on the classification result, the hyperspectral image deep learning classification method provided by the invention can avoid manual extraction and optimize the classification characteristics, can obtain a relatively ideal classification effect and relatively high classification precision, and also improves the image classification efficiency.
EXAMPLE III
The above description of the hyperspectral image deep learning classification method in the embodiment of the invention, and the following description of the hyperspectral image deep learning classification device in the embodiment of the invention, please refer to fig. 3, an embodiment of the hyperspectral image deep learning classification device in the embodiment of the invention includes:
and the image to be classified acquiring module 301 is used for acquiring the hyperspectral image to be classified.
And the data set acquisition module 302 is configured to randomly crop the hyperspectral image to be classified according to the size of a preset window to obtain a corresponding data set.
And a to-be-classified sample set obtaining module 303, configured to perform data expansion on the data set in an image transformation manner, so as to obtain a corresponding to-be-classified sample set.
The image classification result obtaining module 304 is configured to perform classification label prediction on the hyperspectral images in the sample set to be classified through a preset neural network classification model obtained through training, and fill the obtained classification labels into a matrix with the same size as that of the hyperspectral images to be classified to obtain an image classification result corresponding to the hyperspectral images to be classified.
The evaluation module 305 is configured to evaluate the image classification result by using a preset classification evaluation index, and obtain a hyperspectral image classification result and classification accuracy meeting the requirement of the corresponding evaluation index.
Optionally, as shown in fig. 4, the hyperspectral image deep learning classification apparatus further includes:
and the image to be classified acquiring module 301 is used for acquiring the hyperspectral image to be classified.
A marked sample set obtaining module 401, configured to obtain a marked sample set of a predetermined data scale, where a hyperspectral image in the marked sample set is marked according to a classification label, and the hyperspectral marked sample set is divided into a sample set a and a sample set B according to a preset sample proportion through hierarchical random sampling, where the sample set B is used as a test set;
and a to-be-trained sample set obtaining module 402, configured to randomly crop the to-be-classified hyperspectral images according to a preset window size and the labeled sample set a, so as to obtain a corresponding to-be-trained sample set.
A deep learning sample set obtaining module 403, configured to perform data expansion on the sample set to be trained in an image transformation manner, so as to obtain a corresponding deep learning sample set.
And a feature extraction branch construction module 404, configured to input the deep learning sample set into a three-dimensional convolution residual network model, and respectively construct a spatial feature extraction network branch and a spectral feature extraction network branch by using residual connections of different step lengths.
And the spatial feature acquisition module 405 is configured to set spatial dimension hyper-parameters of a three-dimensional convolution kernel in the spatial feature extraction network branch, extract spatial features of the hyper-spectral images in the deep learning sample set in a parallel manner, and perform band superposition on the obtained spatial features.
And the spectral feature acquisition module 406 is configured to set spectral dimensional hyper-parameters of a three-dimensional convolution kernel in the spectral feature extraction network branch, extract spectral features of the hyperspectral images in the deep learning sample set in a parallel manner, and perform band superposition on the obtained spectral features.
And the spectrum time sequence feature acquisition module 407 is configured to extract the spectrum time sequence features of the hyperspectral images in the deep learning sample set by using a convolution cyclic neural network model.
And the feature fusion module 408 is configured to input the time sequence features, the superposition space features, and the superposition spectrum features into an average pooling layer of the three-dimensional convolution residual error network model by using a cascading strategy to perform feature fusion, so as to obtain a feature-fused hyperspectral image.
And the neural network classification model construction module 409 is used for automatically identifying and classifying the hyperspectral images with the fused features by utilizing a softmax function, and obtaining a constructed neural network classification model, wherein the construction of the neural network classification model is based on the three-dimensional convolution residual error network model and the convolution cyclic neural network model.
The preset neural network classification model obtaining module 410 is configured to perform algorithm iteration on the neural network classification model until the repetition times reaches a preset iteration time or the classification accuracy reaches a preset accuracy requirement, so as to obtain a corresponding preset neural network classification model.
The image classification result obtaining module 304 is configured to perform classification label prediction on the hyperspectral images in the sample set to be classified through a preset neural network classification model obtained through training, and fill the obtained classification labels into a matrix with the same size as that of the hyperspectral images to be classified to obtain an image classification result corresponding to the hyperspectral images to be classified.
The evaluation module 305 is configured to evaluate the image classification result by using a preset classification evaluation index, and obtain a hyperspectral image classification result and classification accuracy meeting the requirement of the corresponding evaluation index.
Optionally, in another embodiment of the hyperspectral image deep learning classification apparatus according to the present invention, the hyperspectral image classification apparatus further includes:
and the deep learning sample set dividing module is used for dividing the deep learning sample set into a training set and a verification set by adopting random sampling. The training set is used for training and constructing the preset neural network classification model, the verification set is used for adjusting and optimizing training parameters of the preset neural network classification model, and the test set is used for testing the precision and the robustness of the preset neural network classification model.
Optionally, in another embodiment of the hyperspectral image deep learning classification device of the present invention, the hyperspectral image deep learning classification device further includes:
and the Dropout layer adding module is used for adding a Dropout layer in the three-dimensional convolution residual error network model, randomly discarding some neurons in the three-dimensional convolution residual error network model according to a preset probability, and zeroing the weight of a hidden layer or an output layer.
Optionally, in another embodiment of the hyperspectral image deep learning classification device of the present invention, the hyperspectral image deep learning classification device further includes:
and the learning rate dynamic setting module is used for adopting an Adam optimizer in the three-dimensional convolution residual error network model, setting an initial learning rate in the Adam optimizer, and enabling the learning rate in the Adam optimizer to realize dynamic change so as to realize self-adaptive optimization of the space dimension hyperparameter and the spectrum dimension hyperparameter.
The hyperspectral image deep learning classification device in the embodiment of the invention is described in detail in the aspect of the modular functional entity in the above fig. 3 and fig. 4, and the hyperspectral image deep learning classification device in the embodiment of the invention is described in detail in the aspect of hardware processing in the following.
Fig. 5 is a schematic structural diagram of a hyper-spectral image deep learning classification apparatus according to an embodiment of the present invention, where the hyper-spectral image classification apparatus 500 may generate relatively large differences due to different configurations or performances, and may include one or more processors (CPUs) 501 (e.g., one or more processors) and a memory 509, and one or more storage media 508 (e.g., one or more mass storage devices) storing an application 507 or data 506. Memory 509 and storage medium 508 may be, among other things, transient storage or persistent storage. The program stored on the storage medium 508 may include one or more modules (not shown), each of which may include a series of instruction operations in a boolean variable store computed on a graph. Still further, the processor 501 may be arranged to communicate with a storage medium 508, and execute a series of instruction operations in the storage medium 508 on the hyperspectral image classification apparatus 500.
The hyper-spectral image classification apparatus 500 may also include one or more power supplies 502, one or more wired or wireless network interfaces 503, one or more input-output interfaces 504, and/or one or more operating systems 505, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, and the like. Those skilled in the art will appreciate that the hyperspectral image deep learning classification apparatus structure shown in fig. 5 does not constitute a limitation of the hyperspectral image classification apparatus and may include more or fewer components than shown, or combine certain components, or a different arrangement of components.
It is clear to those skilled in the art that, for convenience and brevity of description, the specific working processes of the above-described systems, apparatuses and units may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.
In the embodiments provided in the present invention, it should be understood that the disclosed system, apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the units is only one logical division, and other divisions may be realized in practice, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.
The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.
In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.
The integrated unit, if implemented in the form of a software functional unit and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a read-only memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disk.
The above-mentioned embodiments are only used for illustrating the technical solutions of the present invention, and not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims (10)

1. A hyperspectral image deep learning classification method is characterized by comprising the following steps:
acquiring a hyperspectral image to be classified;
randomly cutting the hyperspectral images to be classified according to the size of a preset window to obtain a corresponding data set;
performing data expansion on the data set in an image transformation mode to obtain a corresponding sample set to be classified;
performing classification label prediction on the hyperspectral images in the sample set to be classified through a preset neural network classification model obtained through training, and filling the obtained classification labels into a matrix with the same size as the hyperspectral images to be classified to obtain image classification results corresponding to the hyperspectral images to be classified;
and evaluating the image classification result by adopting a preset classification evaluation index to obtain a hyperspectral image classification result and classification precision meeting the requirement of the corresponding evaluation index.
2. The hyperspectral image deep learning classification method according to claim 1, wherein before the hyperspectral images to be classified are randomly cut according to the size of a preset window and the marked sample set A to obtain a corresponding data set, the hyperspectral image deep learning classification method comprises the following steps:
the method comprises the steps of obtaining a marked sample set with a preset data scale, marking hyperspectral images in the marked sample set according to classification labels, and dividing the hyperspectral marked sample set into a sample set A and a sample set B according to a preset sample proportion through hierarchical random sampling, wherein the sample set B is used as a test set;
according to the size of a preset window and the marked sample set A, randomly cutting the hyperspectral image to obtain a corresponding sample set to be trained;
and performing data expansion on the sample set to be trained in an image transformation mode to obtain a corresponding deep learning sample set.
3. The hyperspectral image deep learning classification method according to claim 2, wherein the step of performing data expansion on the sample set to be trained in an image transformation mode to obtain a corresponding deep learning sample set comprises the following steps:
dividing the deep learning sample set into a training set and a verification set by adopting random sampling; the training set is used for training and constructing the preset neural network classification model, the verification set is used for adjusting and optimizing training parameters of the preset neural network classification model, and the test set is used for testing the precision and the robustness of the preset neural network classification model.
4. The hyperspectral image deep learning classification method according to claim 2 is characterized in that after the data expansion is performed on the sample set to be trained in an image transformation mode to obtain a corresponding deep learning sample set, the method further comprises the following steps:
inputting the deep learning sample set into a three-dimensional convolution residual error network model, and respectively constructing a spatial feature extraction network branch and a spectral feature extraction network branch by adopting residual error connection with different step lengths;
setting spatial dimension hyper-parameters of a three-dimensional convolution kernel in the spatial feature extraction network branch, extracting spatial features of the hyper-spectral images in the deep learning sample set in a parallel mode, and performing wave band superposition on the obtained spatial features;
setting spectral dimensional hyper-parameters of a three-dimensional convolution kernel in the spectral feature extraction network branch, extracting spectral features of the hyperspectral images in the deep learning sample set in a parallel mode, and performing wave band superposition on the obtained spectral features;
extracting the spectrum time sequence characteristics of the hyperspectral images in the deep learning sample set by adopting a convolution cyclic neural network model;
inputting the time sequence features, the superposition space features and the superposition spectrum features into an average pooling layer of the three-dimensional convolution residual error network model by using a cascading strategy for feature fusion to obtain a feature-fused hyperspectral image;
and automatically identifying and classifying the hyperspectral images with the fused features by utilizing a softmax function, and obtaining a constructed neural network classification model, wherein the construction of the neural network classification model is based on the three-dimensional convolution residual error network model and the convolution cyclic neural network model.
5. The hyperspectral image deep learning classification method according to claim 3, wherein after the hyperspectral image of the feature fusion is automatically identified and classified by utilizing a softmax function and a constructed neural network classification model is obtained, the method further comprises the following steps:
and carrying out algorithm iteration on the neural network classification model until the repetition times reach a preset iteration time or the classification precision reaches a preset precision requirement, and obtaining a corresponding preset neural network classification model.
6. The hyperspectral image deep learning classification method according to claim 3, wherein the method further comprises the steps of, after setting a spectral dimensional hyper-parameter of a three-dimensional convolution kernel in the spectral feature extraction network branch, extracting the spectral features of the hyperspectral images in the deep learning sample set in a parallel manner, and performing band superposition on the obtained spectral features:
and adding a Dropout layer in the three-dimensional convolution residual error network model, randomly discarding some neurons in the three-dimensional convolution residual error network model according to a preset probability, and returning the weight of a hidden layer or an output layer to zero.
7. The hyperspectral image deep learning classification method according to claim 3, wherein the method further comprises the steps of, after setting a spectral dimensional hyper-parameter of a three-dimensional convolution kernel in the spectral feature extraction network branch, extracting the spectral features of the hyperspectral images in the deep learning sample set in a parallel manner, and performing band superposition on the obtained spectral features:
and adopting an Adam optimizer in the three-dimensional convolution residual error network model, setting an initial learning rate in the Adam optimizer, and enabling the learning rate in the Adam optimizer to realize dynamic change so as to realize self-adaptive optimization of the space dimensional hyper-parameter and the spectrum dimensional hyper-parameter.
8. The utility model provides a hyperspectral image degree of depth study sorter which characterized in that includes:
the image to be classified acquisition module is used for acquiring a hyperspectral image to be classified;
the data set acquisition module is used for randomly cutting the hyperspectral images to be classified according to the size of a preset window to obtain a corresponding data set;
the to-be-classified sample set acquisition module is used for performing data expansion on the data set in an image transformation mode to obtain a corresponding to-be-classified sample set;
the image classification result acquisition module is used for predicting classification labels of the hyperspectral images in the sample set to be classified through a preset neural network classification model obtained through training, and filling the obtained classification labels into a matrix with the same size as the hyperspectral images to be classified to obtain image classification results corresponding to the hyperspectral images to be classified;
and the evaluation module is used for evaluating the image classification result by adopting a preset classification evaluation index to obtain a hyperspectral image classification result and classification precision meeting the requirement of the corresponding evaluation index.
9. The hyper-spectral image deep learning classification device is characterized by comprising: a memory having instructions stored therein and at least one processor, the memory and the at least one processor interconnected by a line;
the at least one processor invokes the instructions in the memory to cause the hyper-spectral image deep learning classification apparatus to perform the method of any of claims 1-7.
10. A computer-readable storage medium having stored thereon a computer program, characterized in that: the computer program realizing the steps of the method according to any one of claims 1-7 when executed by a processor.
CN201911064083.XA 2019-11-04 2019-11-04 Hyperspectral image deep learning classification method, device, equipment and storage medium Pending CN110852227A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911064083.XA CN110852227A (en) 2019-11-04 2019-11-04 Hyperspectral image deep learning classification method, device, equipment and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911064083.XA CN110852227A (en) 2019-11-04 2019-11-04 Hyperspectral image deep learning classification method, device, equipment and storage medium

Publications (1)

Publication Number Publication Date
CN110852227A true CN110852227A (en) 2020-02-28

Family

ID=69598258

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911064083.XA Pending CN110852227A (en) 2019-11-04 2019-11-04 Hyperspectral image deep learning classification method, device, equipment and storage medium

Country Status (1)

Country Link
CN (1) CN110852227A (en)

Cited By (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111353463A (en) * 2020-03-12 2020-06-30 北京工业大学 Hyperspectral image classification method based on random depth residual error network
CN111507409A (en) * 2020-04-17 2020-08-07 中国人民解放军战略支援部队信息工程大学 Hyperspectral image classification method and device based on depth multi-view learning
CN111582387A (en) * 2020-05-11 2020-08-25 吉林大学 Rock spectral feature fusion classification method and system
CN111611954A (en) * 2020-05-28 2020-09-01 云南电网有限责任公司电力科学研究院 Hyperspectral image classification method and device based on improved K-means algorithm
CN111667019A (en) * 2020-06-23 2020-09-15 哈尔滨工业大学 Hyperspectral image classification method based on deformable separation convolution
CN112037181A (en) * 2020-08-12 2020-12-04 深圳大学 2D SAXS atlas analysis model training method and device
CN112115850A (en) * 2020-09-16 2020-12-22 哈尔滨工程大学 Method for analyzing consistency of camouflage target and background spectrum
CN112364896A (en) * 2020-10-26 2021-02-12 平安科技（深圳）有限公司 Health information distribution determination method and device based on machine learning
CN112419324A (en) * 2020-11-24 2021-02-26 山西三友和智慧信息技术股份有限公司 Medical image data expansion method based on semi-supervised task driving
CN112560960A (en) * 2020-12-16 2021-03-26 北京影谱科技股份有限公司 Hyperspectral image classification method and device and computing equipment
CN112633401A (en) * 2020-12-29 2021-04-09 中国科学院长春光学精密机械与物理研究所 Hyperspectral remote sensing image classification method, device, equipment and storage medium
CN112651317A (en) * 2020-12-18 2021-04-13 中国电子科技集团公司信息科学研究院 Hyperspectral image classification method and system for sample relation learning
CN112733736A (en) * 2021-01-13 2021-04-30 西安电子科技大学 Class imbalance hyperspectral image classification method based on enhanced oversampling
CN112926789A (en) * 2021-03-17 2021-06-08 阳光电源股份有限公司 Satellite cloud picture prediction method, prediction device and readable storage medium
CN112966623A (en) * 2021-03-16 2021-06-15 长安大学 Multilayer width neural network and training method and application thereof
CN112966740A (en) * 2021-03-05 2021-06-15 西安邮电大学 Small sample hyperspectral image classification method based on core sample adaptive expansion
CN112990106A (en) * 2021-04-19 2021-06-18 中国人民解放军国防科技大学 Underwater object detection method and device, computer equipment and storage medium
CN113065403A (en) * 2021-03-05 2021-07-02 浙江大学 Hyperspectral imaging-based machine learning cell classification method and device
CN113160183A (en) * 2021-04-26 2021-07-23 山东深蓝智谱数字科技有限公司 Hyperspectral data processing method, device and medium
CN113158980A (en) * 2021-05-17 2021-07-23 四川农业大学 Tea leaf classification method based on hyperspectral image and deep learning
CN113392916A (en) * 2021-06-23 2021-09-14 华南农业大学 Method and system for detecting nutritional ingredients of bamboo shoots based on hyperspectral image and storage medium
CN115187870A (en) * 2022-09-13 2022-10-14 浙江蓝景科技有限公司杭州分公司 Marine plastic waste material identification method and system, electronic equipment and storage medium
CN115965953A (en) * 2023-01-04 2023-04-14 哈尔滨工业大学 Grain variety classification method based on hyperspectral imaging and deep learning
CN116188916A (en) * 2023-04-17 2023-05-30 杰创智能科技股份有限公司 Fine granularity image recognition method, device, equipment and storage medium
CN117095806B (en) * 2023-10-19 2023-12-15 邦士医疗科技股份有限公司 Quality management system of medical equipment in operating room

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108052911A (en) * 2017-12-20 2018-05-18 上海海洋大学 Multi-modal remote sensing image high-level characteristic integrated classification method based on deep learning
CN108491849A (en) * 2018-03-23 2018-09-04 上海理工大学 Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks
CN109754017A (en) * 2019-01-09 2019-05-14 西北工业大学 Based on separable three-dimensional residual error network and transfer learning hyperspectral image classification method
CN110263735A (en) * 2019-06-25 2019-09-20 北京林业大学 A method of tree species classification being carried out to artificial forest high-spectral data using Three dimensional convolution neural network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108052911A (en) * 2017-12-20 2018-05-18 上海海洋大学 Multi-modal remote sensing image high-level characteristic integrated classification method based on deep learning
CN108491849A (en) * 2018-03-23 2018-09-04 上海理工大学 Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks
CN109754017A (en) * 2019-01-09 2019-05-14 西北工业大学 Based on separable three-dimensional residual error network and transfer learning hyperspectral image classification method
CN110263735A (en) * 2019-06-25 2019-09-20 北京林业大学 A method of tree species classification being carried out to artificial forest high-spectral data using Three dimensional convolution neural network

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
WENCHAO QI ET AL.: "A Spectral-Spatial Cascaded 3D Convolutional Neural Network with a Convolutional Long Short-Term Memory Network for Hyperspectral Image Classification", 《REMOTE SENSING》 *

Cited By (37)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111353463A (en) * 2020-03-12 2020-06-30 北京工业大学 Hyperspectral image classification method based on random depth residual error network
CN111507409A (en) * 2020-04-17 2020-08-07 中国人民解放军战略支援部队信息工程大学 Hyperspectral image classification method and device based on depth multi-view learning
CN111582387A (en) * 2020-05-11 2020-08-25 吉林大学 Rock spectral feature fusion classification method and system
CN111611954A (en) * 2020-05-28 2020-09-01 云南电网有限责任公司电力科学研究院 Hyperspectral image classification method and device based on improved K-means algorithm
CN111611954B (en) * 2020-05-28 2023-11-24 云南电网有限责任公司电力科学研究院 Hyperspectral image classification method and device based on improved K-means algorithm
CN111667019A (en) * 2020-06-23 2020-09-15 哈尔滨工业大学 Hyperspectral image classification method based on deformable separation convolution
CN112037181A (en) * 2020-08-12 2020-12-04 深圳大学 2D SAXS atlas analysis model training method and device
CN112037181B (en) * 2020-08-12 2023-09-08 深圳大学 2D SAXS (three dimensional architecture) atlas analysis model training method and device
CN112115850A (en) * 2020-09-16 2020-12-22 哈尔滨工程大学 Method for analyzing consistency of camouflage target and background spectrum
CN112115850B (en) * 2020-09-16 2023-01-03 哈尔滨工程大学 Method for analyzing consistency of camouflage target and background spectrum
WO2021189982A1 (en) * 2020-10-26 2021-09-30 平安科技（深圳）有限公司 Health information distribution determination method and apparatus based on machine learning
CN112364896B (en) * 2020-10-26 2023-10-24 平安科技（深圳）有限公司 Method and device for determining health information distribution based on machine learning
CN112364896A (en) * 2020-10-26 2021-02-12 平安科技（深圳）有限公司 Health information distribution determination method and device based on machine learning
CN112419324A (en) * 2020-11-24 2021-02-26 山西三友和智慧信息技术股份有限公司 Medical image data expansion method based on semi-supervised task driving
CN112419324B (en) * 2020-11-24 2022-04-19 山西三友和智慧信息技术股份有限公司 Medical image data expansion method based on semi-supervised task driving
CN112560960A (en) * 2020-12-16 2021-03-26 北京影谱科技股份有限公司 Hyperspectral image classification method and device and computing equipment
CN112651317A (en) * 2020-12-18 2021-04-13 中国电子科技集团公司信息科学研究院 Hyperspectral image classification method and system for sample relation learning
CN112651317B (en) * 2020-12-18 2022-11-25 中国电子科技集团公司信息科学研究院 Hyperspectral image classification method and system for sample relation learning
CN112633401B (en) * 2020-12-29 2022-11-08 中国科学院长春光学精密机械与物理研究所 Hyperspectral remote sensing image classification method, device, equipment and storage medium
CN112633401A (en) * 2020-12-29 2021-04-09 中国科学院长春光学精密机械与物理研究所 Hyperspectral remote sensing image classification method, device, equipment and storage medium
CN112733736A (en) * 2021-01-13 2021-04-30 西安电子科技大学 Class imbalance hyperspectral image classification method based on enhanced oversampling
CN112966740A (en) * 2021-03-05 2021-06-15 西安邮电大学 Small sample hyperspectral image classification method based on core sample adaptive expansion
CN112966740B (en) * 2021-03-05 2024-03-29 西安邮电大学 Small sample hyperspectral image classification method based on core sample self-adaptive expansion
CN113065403A (en) * 2021-03-05 2021-07-02 浙江大学 Hyperspectral imaging-based machine learning cell classification method and device
CN112966623B (en) * 2021-03-16 2024-03-19 长安大学 Multi-layer wide neural network and training method and application thereof
CN112966623A (en) * 2021-03-16 2021-06-15 长安大学 Multilayer width neural network and training method and application thereof
CN112926789A (en) * 2021-03-17 2021-06-08 阳光电源股份有限公司 Satellite cloud picture prediction method, prediction device and readable storage medium
CN112990106A (en) * 2021-04-19 2021-06-18 中国人民解放军国防科技大学 Underwater object detection method and device, computer equipment and storage medium
CN113160183A (en) * 2021-04-26 2021-07-23 山东深蓝智谱数字科技有限公司 Hyperspectral data processing method, device and medium
CN113158980A (en) * 2021-05-17 2021-07-23 四川农业大学 Tea leaf classification method based on hyperspectral image and deep learning
CN113392916B (en) * 2021-06-23 2023-10-17 华南农业大学 Method, system and storage medium for detecting nutrition components of moso bamboo shoots based on hyperspectral image
CN113392916A (en) * 2021-06-23 2021-09-14 华南农业大学 Method and system for detecting nutritional ingredients of bamboo shoots based on hyperspectral image and storage medium
CN115187870A (en) * 2022-09-13 2022-10-14 浙江蓝景科技有限公司杭州分公司 Marine plastic waste material identification method and system, electronic equipment and storage medium
CN115965953B (en) * 2023-01-04 2023-08-22 哈尔滨工业大学 Grain Variety Classification Method Based on Hyperspectral Imaging and Deep Learning
CN115965953A (en) * 2023-01-04 2023-04-14 哈尔滨工业大学 Grain variety classification method based on hyperspectral imaging and deep learning
CN116188916A (en) * 2023-04-17 2023-05-30 杰创智能科技股份有限公司 Fine granularity image recognition method, device, equipment and storage medium
CN117095806B (en) * 2023-10-19 2023-12-15 邦士医疗科技股份有限公司 Quality management system of medical equipment in operating room

Similar Documents

Publication Publication Date Title
CN110852227A (en) Hyperspectral image deep learning classification method, device, equipment and storage medium
CN110084159B (en) Hyperspectral image classification method based on combined multistage spatial spectrum information CNN
CN112052755B (en) Semantic convolution hyperspectral image classification method based on multipath attention mechanism
CN111259898A (en) Crop segmentation method based on unmanned aerial vehicle aerial image
CN111680176A (en) Remote sensing image retrieval method and system based on attention and bidirectional feature fusion
CN110084311B (en) Hyperspectral image wave band selection method based on ternary weight convolution neural network
CN113095409B (en) Hyperspectral image classification method based on attention mechanism and weight sharing
CN112560966B (en) Polarized SAR image classification method, medium and equipment based on scattering map convolution network
Guirado et al. Deep-learning convolutional neural networks for scattered shrub detection with google earth imagery
CN105184314A (en) wrapper-type hyperspectral waveband selection method based on pixel clustering
CN111709318A (en) High-resolution remote sensing image classification method based on generation countermeasure network
CN113705580A (en) Hyperspectral image classification method based on deep migration learning
CN107680081B (en) Hyperspectral image unmixing method based on convolutional neural network
CN114937173A (en) Hyperspectral image rapid classification method based on dynamic graph convolution network
Jin et al. Deep learning for seasonal precipitation prediction over China
CN113435254A (en) Sentinel second image-based farmland deep learning extraction method
Zhan et al. A new hyperspectral band selection approach based on convolutional neural network
CN113486869B (en) Method, device and medium for lithology identification based on unsupervised feature extraction
Du et al. Training SegNet for cropland classification of high resolution remote sensing images
CN113392871A (en) Polarized SAR terrain classification method based on scattering mechanism multichannel expansion convolutional neural network
CN116958709A (en) Hyperspectral classification method of lightweight tensor convolution long-short-term memory network
CN109460788B (en) Hyperspectral image classification method based on low-rank-sparse information combination network
CN116664954A (en) Hyperspectral ground object classification method based on graph convolution and convolution fusion
Laban et al. Enhanced pixel based urban area classification of satellite images using convolutional neural network
CN113420812B (en) Polarized SAR image classification method based on evolutionary convolutional neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination