CN111931618A - Hyperspectral classification method based on separable residual three-dimensional dense convolution - Google Patents

Hyperspectral classification method based on separable residual three-dimensional dense convolution Download PDF

Info

Publication number
CN111931618A
CN111931618A CN202010744236.1A CN202010744236A CN111931618A CN 111931618 A CN111931618 A CN 111931618A CN 202010744236 A CN202010744236 A CN 202010744236A CN 111931618 A CN111931618 A CN 111931618A
Authority
CN
China
Prior art keywords
hyperspectral
residual
convolution
classification
dimensional
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
CN202010744236.1A
Other languages
Chinese (zh)
Inventor
邱云飞
吕舜尧
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.)
Liaoning Technical University
Original Assignee
Liaoning Technical University
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 Liaoning Technical University filed Critical Liaoning Technical University
Priority to CN202010744236.1A priority Critical patent/CN111931618A/en
Publication of CN111931618A publication Critical patent/CN111931618A/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
    • G06F18/24—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
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a hyperspectral classification method based on separable residual three-dimensional dense convolution, which comprises the following steps of: directly inputting a series of hyperspectral sample images subjected to data preprocessing into a network as input data; extracting the spatial spectrum characteristics of the hyperspectral image through residual three-dimensional separable dense convolution; and performing batch normalization, dropout and full connection operation, improving the separability of characteristics, reducing the number of model parameters, reducing the difficulty of model training, and predicting output data through a softmax classifier to obtain a classification result. According to the method, the three-dimensional residual separation dense convolution network is used for extracting the empty spectrum characteristics of the hyperspectral cube, the output of each unit is in short connection with the output of each unit of the next unit, information transfer is achieved, and finally the classification effect is achieved through the softmax classifier.

Description

Hyperspectral classification method based on separable residual three-dimensional dense convolution
Technical Field
The invention belongs to the technical field of hyperspectral image classification, and particularly relates to a hyperspectral classification method based on separable residual three-dimensional dense convolution.
Background
The Hyperspectral remote sensing image (Hyperspectral image) is a high-resolution image combining a spectrum technology and an image technology, and the Hyperspectral image plays a great role in various fields such as military affairs, agriculture, geographic detection, environmental monitoring and the like in recent years. The classification of hyperspectral images is one of the hot problems of current research. The hyperspectral image contains a plurality of spectral channels, and the high dimensionality is a big characteristic of the hyperspectral image. The spectrum mixer also has the characteristic of spectrum mixing, namely, one-dimensional spectrum information is added on the basis of two-dimensional space information, and the bandwidth is less than 10 nm. Each band in the image is a two-dimensional image. Each pixel in the spectral space is reflected as a continuous curve. Generally, the goal of hyperspectral remote sensing science is to acquire data using hundreds of spectral bands in order to provide detailed spectral and spatial information simultaneously. Therefore, high accuracy is particularly useful for HIS. The hyperspectral image is utilized to carry out ground object classification, target identification and target detection, which are research hotspots. Therefore, the development of the research of the hyperspectral remote sensing technology has necessary functions and significance. In recent years, scholars at home and abroad propose various classification invention methods of hyperspectral images. Many machine learning related image processing techniques are applied in hyperspectral classification. The classification method is roughly classified into a classification method based on spectral classification recognition and statistical recognition. The identification method based on the spectrum classification is to identify the ground feature classification by using the spectral data by adopting the method, and the statistical method is divided into unsupervised classification and supervised classification. Like maximum likelihood classification, SVM supports vector machines, but the above-mentioned methods either receive the influence of spectral fluctuations or receive the problem of classifier model parameters, making the classification inaccurate. Deep learning represented by a convolutional neural network (cnn) has made a breakthrough in image classification and pattern recognition in recent years, the convolutional neural network using deep learning can increase robustness, but problems of overfitting and gradient decrease easily occur with the increase of the number of network layers, and a network called residual error network (resnet) was proposed by the party of hokkming in 2015, which is a popularization of the convolutional neural network. The method utilizes a residual module which has jumping property, is convenient for gradient propagation, has stronger robustness and deeper system structure, and can reduce a plurality of parameters and overfitting. Zhong et al used the SSRN including the three-dimensional convolution layer to extract the characteristics of hyperspectral image cube in 2017, alleviated the gradient disappearance problem caused by the increase of network depth by introducing a residual structure, and the classification precision in hyperspectral image classification is greatly improved compared with the traditional method. The resnet is introduced into the hyperspectral classification, so that the problems of gradient disappearance and overlarge parameters can be solved well, and the method has a very good effect on the hyperspectral classification.
The two-dimensional convolutional neural network (2D-CNN) has outstanding performances in many classical fields, such as image classification, target detection, image evaluation and other image fields, and has good research results, the two-dimensional convolutional neural network can directly extract features of the images when the images are processed, an end-to-end processing mode is completed, but if the two-dimensional convolutional neural network is directly used for hyperspectral images, each channel needs to be convolved, each group of spectral bands of the hyperspectral images have many channels, a convolution kernel needs to be trained, a large number of parameters can be calculated, the calculation cost is greatly increased, the calculation efficiency is reduced, and overfitting can be generated.
Before three-dimensional convolution is proposed, researchers generally solve the problem by reducing the dimensionality of a spectrum through a data dimensionality reduction method, for example, PCA is used for dimensionality reduction, main component channels in a hyperspectral image are extracted, and then features are extracted through a two-dimensional convolution network.
Disclosure of Invention
Based on the defects of the prior art, the technical problem to be solved by the invention is to provide a hyperspectral classification method based on separable residual three-dimensional dense convolution, which is used for extracting features by inputting a three-dimensional original picture without dimension reduction, simultaneously extracting the features on a space channel and a spectrum channel to achieve space-spectrum combination, avoiding introducing a large number of parameters, preventing overfitting and improving the operation efficiency.
In order to solve the technical problem, the invention provides a hyperspectral classification method based on separable residual three-dimensional dense convolution, which comprises the following steps:
step 1: directly inputting a series of hyperspectral sample images subjected to data preprocessing into a network as input data;
step 2: extracting the spatial spectrum characteristics of the hyperspectral image through residual three-dimensional separable dense convolution;
and step 3: and performing batch normalization, dropout and full connection operation, improving the separability of characteristics, reducing the number of model parameters, reducing the difficulty of model training, and predicting output data through a softmax classifier to obtain a classification result.
Further, in step 1, a hyperspectral data cube with a neighborhood size of 7 × 7 × B is extracted with each target pixel as a center, and the network input is first preprocessed by a convolution kernel with 1 × 7 and a convolution layer with a step size of 1 × 2 and a maximum pooling layer with a kernel with 1 × 3 and a step size of 1 × 2.
From the above, in the classification of hyperspectral images, the existing method does not fully utilize all the hierarchical features extracted by a network model, and aiming at the characteristics of hyperspectral image dimension height and sample limitation, the invention designs a separable three-dimensional residual dense convolution network model, the network directly takes a simply preprocessed hyperspectral cube as input, the three-dimensional residual dense convolution network is used for extracting the empty spectrum features of the hyperspectral cube, the output of each unit is in short connection with the output of each unit of the next unit, information transmission is realized, and finally, the classification effect is achieved through a softmax classifier, and the experimental result shows that the classification precision is remarkably improved compared with the existing algorithm effect.
The foregoing description is only an overview of the technical solutions of the present invention, and in order to make the technical means of the present invention more clearly understood, the present invention may be implemented in accordance with the content of the description, and in order to make the above and other objects, features, and advantages of the present invention more clearly understood, the following detailed description is given in conjunction with the preferred embodiments, together with the accompanying drawings.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings of the embodiments will be briefly described below.
FIG. 1 is a flowchart of a hyperspectral classification method based on separable residual three-dimensional dense convolution according to the invention.
FIG. 2 is a graph of test results using overall classification accuracy (OA), average classification accuracy (AA), and Kappa coefficients as evaluation criteria on an Indian Pines dataset and a Pavia University dataset; wherein (a) and (b) are contrasted plots at different spatial sizes; (c) and (d) are contrasted plots at different training sample sizes;
FIG. 3 is a comparison of the various algorithms of Indian pings;
FIG. 4 is a comparison graph of different pavia unity algorithms.
Detailed Description
Other aspects, features and advantages of the present invention will become apparent from the following detailed description, taken in conjunction with the accompanying drawings, which form a part of this specification, and which illustrate, by way of example, the principles of the invention. In the referenced drawings, the same or similar components in different drawings are denoted by the same reference numerals.
In consideration of the conditions that the hyperspectral image has high dimensionality and the spectral feature and the spatial feature are not extracted sufficiently, the invention designs the three-dimensional residual separable dense convolutional neural network module for feature extraction, fully utilizes the layered features extracted by different units, combines the shallow feature and the deep feature to provide rich image detail information for image classification. Compared with the traditional methods such as 2DCNN, 3DCNN, resnet and the like, the results show that the method can effectively improve the classification precision and increase the stability.
The method utilizes the three-dimensional convolution neural network (3D-CNN) to directly extract the features of the hyperspectral image, has one convolution kernel dimension more than that of 2D-CNN, can directly act on the hyperspectral image compared with 2D-CNN, extracts the features by inputting a three-dimensional original image, does not need dimension reduction, extracts the features on a space and a spectrum channel simultaneously, achieves the combination of empty spectrums, can avoid introducing a large number of parameters, prevents overfitting and improves the operation efficiency. The 3D-CNN is more suitable for hyperspectral images with high dimensionality and multiple parameters.
The residual error network learning method is an important method for deep learning, can train networks of any layer, has many defects in the traditional network, and causes network degradation along with the deepening of the network, the accuracy of classification and identification quickly reaches saturation, and the phenomena that the network layer is deeper and the error rate is higher are generated. The residual error network can well solve the problems of gradient disappearance and gradient explosion. Let the input of the residual unit be x, the output of the network unit without short connection be F, the residual unit with added short connection have an output of h (x), and then h (x) F + x is present. Experiments have shown that in the residual unit, F is much easier to optimize than the original function mapping h (x).
Convolution combines the advantages of a residual error module and a separable convolution model, and aims to create a novel neural network which can have good performance in a deep network, can reduce network parameters and reduce the running time of the network, thereby improving the efficiency.
First is a channel-by-channel convolution, where one convolution kernel corresponds to one channel and only acts on that channel. The initial signature is then concatenated through N convolution kernels of 1 x 1, which can greatly reduce the number of parameters. For example, the present invention adds a short concatenation of residuals to the convolution model, assuming that the original channel is first reduced to 1/4 by 1 × 1 convolution. Then 3 x 3 convolution is performed and finally the output is expanded back to the original number by 1 x 1 convolution, so that the total number of executions is tested and only 1/9 of the original number of operations is needed. The structure can effectively reduce the calculation complexity and improve the operational capability of the network function.
The invention combines the advantages of the separated convolution and the dense network and designs a novel residual dense convolution structure. The residual error separation dense unit is composed of a plurality of convolution layers and an activation layer and plays a role in feature extraction. And each output tensor is directly input into the next convolutional layer to establish a short connection with the next convolutional layer. The method directly uses the data of the preprocessed hyperspectral center pixel and the field pixels thereof as the input of a residual error separation dense model, and the network model comprises 3 residual error separation dense modules.
The invention provides a hyperspectral image classification method based on residual three-dimensional separable dense convolution, which can simultaneously extract the spectral characteristics and the spatial characteristics of a hyperspectral image through a residual three-dimensional separable dense convolution module.
As shown in fig. 1, a series of hyperspectral sample images subjected to data preprocessing are directly input into a network as input data, and the specific method is to extract a hyperspectral data cube with the neighborhood size of 7 × 7 × B by taking each target pixel as a center. The net input is first preprocessed through a convolution kernel with 1 x 7 and a convolution layer with step size 1 x 2 and a max pooling layer with kernel 1 x 3 and step size 1 x 2. Then, the spatial spectrum features of the hyperspectral image are extracted through residual three-dimensional separable dense convolution, the problem of gradient disappearance can be effectively reduced through a residual network along with the increase of the network depth, and the structure can more effectively utilize the features and enhance the feature transfer among convolution layers. And then, carrying out batch normalization, dropout and Full Connection (FC) operation, improving the separability of features, reducing the number of model parameters, reducing the difficulty of model training, and predicting output data through a softmax classifier to obtain a classification result. The network contains 21 convolutional layers of which there are 2 residual separation dense convolutional modules. And (4) the ability of extracting nonlinear features accompanied by a relu function of the nonlinear activation layer after the convolutional layer.
To verify the effectiveness of the method of the invention, classification experiments were performed on the hyperspectral images by first inputting the network through a convolution kernel of 1 x 7 with a step size of 1 x 2 and a pooling layer with a kernel of 1 x 3 with a step size of 1 x 2 as input preprocessing. And then outputting the data through the established residual three-dimensional dense convolution model, passing through 2 residual three-dimensional dense convolution blocks in total, and inputting the result after the empty spectrum features are extracted into a softmax classifier to obtain the final result.
The experiment selects an Indian Pines data set commonly used for hyperspectrum, and a comparison experiment is carried out on the Pavia University data set.
The Indian Pines dataset was the earliest dataset for hyperspectral classification (see table 1) collected by an aviris sensor on the northwest Indian pine test site in indiana, usa in 6 months 1992, with a 145 x 145 pixel size and 224 spectral reflectance bands in the wavelength range of 0.4-2.5 μm. The total number of the wave bands is 200 after the water vapor absorption wave bands are removed, and the spatial resolution is 20 m.
TABLE 1
category notill mintill pasture trees windowed notill mintill clean woods total
training 200 200 200 200 200 200 200 200 200 200
testing 1248 650 283 530 280 770 2100 384 1066 7473
The Pavia University data is an Italian Pauia University image (see table 2) obtained by a ROSIS imaging spectrometer of the Germany space center, the pixel size is 610 multiplied by 340, the spectral imager continuously images 115 wave bands with the wavelength range of 0.43-0.86 mu m, the total wave band is 103 wave bands after the water vapor absorption wave band is removed, and the spatial resolution is 1.3 m.
TABLE 2
Figure BDA0002607796220000071
The network model mainly comprises a convolution layer, a pooling layer, a full-connection layer and the like, and the size of the convolution kernel is an important condition influencing the classification precision. The effect of different sizes of convolution kernel on the classification accuracy was tested on two data sets using convolution kernels of 3 x 3, 5 x 5, 7 x 7, 9 x 9, 11 x 11 size, respectively, and the results are shown in fig. 2 using overall classification accuracy (OA), average classification accuracy (AA) and Kappa coefficient as evaluation criteria.
In each data set, 20% of samples are used as training samples, 80% of samples are used as testing samples, 5 independent and non-repeated experiments are respectively carried out on 2 different data sets, finally, the classification accuracy is obtained by 5 average accuracies, and the classification results are shown in a table:
Figure DEST_PATH_IMAGE001
the method of the present invention is compared to typical classification methods, such as SVM, 3DCNN, ResNet, as illustrated by fig. 3 and 4. The improvement rates of the two groups of data respectively reach 98.89% and 98.88%, and the improvement rate is very obvious compared with the prior networks, residual error learning is realized on the network, the feature extraction capability in different layers is improved, and the stability and the classification accuracy of the network are improved.
In order to utilize the characteristics of each layer and improve the classification precision, the invention designs the separable three-dimensional network model with dense residual errors, combines all the characteristics, solves the problem of gradient disappearance, prevents overfitting, fully extracts the information of the image and enhances the transmission and utilization of the information. The method can be integrated with semi-supervised learning, so that the classification precision is improved.
While the foregoing is directed to the preferred embodiment of the present invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof, and the scope thereof is determined by the claims that follow.

Claims (2)

1. The hyperspectral classification method based on the separable residual three-dimensional dense convolution is characterized by comprising the following steps of:
step 1: directly inputting a series of hyperspectral sample images subjected to data preprocessing into a network as input data;
step 2: extracting the spatial spectrum characteristics of the hyperspectral image through residual three-dimensional separable dense convolution;
and step 3: and performing batch normalization, dropout and full connection operation, improving the separability of characteristics, reducing the number of model parameters, reducing the difficulty of model training, and predicting output data through a softmax classifier to obtain a classification result.
2. The method for hyperspectral classification based on separable residual three-dimensional dense convolution according to claim 1, wherein in step 1, a hyperspectral data cube with neighborhood size of 7 x B is extracted centering on each target pixel, and the network input is first preprocessed by a convolution layer with convolution kernel of 1 x 7 and step size of 1 x 2 and a maximum pooling layer with kernel of 1 x 3 and step size of 1 x 2.
CN202010744236.1A 2020-07-29 2020-07-29 Hyperspectral classification method based on separable residual three-dimensional dense convolution Pending CN111931618A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010744236.1A CN111931618A (en) 2020-07-29 2020-07-29 Hyperspectral classification method based on separable residual three-dimensional dense convolution

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010744236.1A CN111931618A (en) 2020-07-29 2020-07-29 Hyperspectral classification method based on separable residual three-dimensional dense convolution

Publications (1)

Publication Number Publication Date
CN111931618A true CN111931618A (en) 2020-11-13

Family

ID=73314779

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010744236.1A Pending CN111931618A (en) 2020-07-29 2020-07-29 Hyperspectral classification method based on separable residual three-dimensional dense convolution

Country Status (1)

Country Link
CN (1) CN111931618A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116051896A (en) * 2023-01-28 2023-05-02 西南交通大学 Hyperspectral image classification method of lightweight mixed tensor neural network

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111368896A (en) * 2020-02-28 2020-07-03 南京信息工程大学 Hyperspectral remote sensing image classification method based on dense residual three-dimensional convolutional neural network

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111368896A (en) * 2020-02-28 2020-07-03 南京信息工程大学 Hyperspectral remote sensing image classification method based on dense residual three-dimensional convolutional neural network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
丁杰等: "基于残差三维卷积神经网络的高光谱遥感图像分类", 《激光杂志》, vol. 40, no. 12, pages 45 - 52 *
蒋家旭: "基于自适应残差3D-CNN的高光谱图像跨域分类", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》, no. 01, pages 2 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116051896A (en) * 2023-01-28 2023-05-02 西南交通大学 Hyperspectral image classification method of lightweight mixed tensor neural network
CN116051896B (en) * 2023-01-28 2023-06-20 西南交通大学 Hyperspectral image classification method of lightweight mixed tensor neural network

Similar Documents

Publication Publication Date Title
CN110399909B (en) Hyperspectral image classification method based on label constraint elastic network graph model
CN112836773B (en) Hyperspectral image classification method based on global attention residual error network
CN110110596B (en) Hyperspectral image feature extraction, classification model construction and classification method
CN112200090B (en) Hyperspectral image classification method based on cross-grouping space-spectral feature enhancement network
CN111985543A (en) Construction method, classification method and system of hyperspectral image classification model
CN112580480B (en) Hyperspectral remote sensing image classification method and device
CN111652039B (en) Hyperspectral remote sensing ground object classification method based on residual error network and feature fusion module
CN111027509B (en) Hyperspectral image target detection method based on double-current convolution neural network
CN112836666A (en) Hyperspectral image classification and identification method
CN115222994A (en) Hyperspectral image classification method based on hybrid spectrum network and multi-head self-attention mechanism
CN110837808A (en) Hyperspectral image classification method based on improved capsule network model
CN113902622B (en) Spectrum super-resolution method based on depth priori joint attention
CN110569884A (en) Hyperspectral remote sensing image classification method based on deep learning and morphology
CN110147725A (en) A kind of high spectrum image feature extracting method for protecting projection based on orthogonal index office
CN116580195A (en) Remote sensing image semantic segmentation method and system based on ConvNeXt convolution
CN115331104A (en) Crop planting information extraction method based on convolutional neural network
CN112949725A (en) Wheat seed classification method based on multi-scale feature extraction
CN115019178A (en) Hyperspectral image classification method based on large kernel convolution attention
CN114898157A (en) Global learning device and method for hyperspectral image classification
CN111931618A (en) Hyperspectral classification method based on separable residual three-dimensional dense convolution
CN113505856A (en) Hyperspectral image unsupervised self-adaptive classification method
Song et al. Multi-source remote sensing image classification based on two-channel densely connected convolutional networks.
CN117115675A (en) Cross-time-phase light-weight spatial spectrum feature fusion hyperspectral change detection method, system, equipment and medium
Laban et al. Enhanced pixel based urban area classification of satellite images using convolutional neural network
CN114373080B (en) Hyperspectral classification method of lightweight hybrid convolution model based on global reasoning

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