深度學(xué)習(xí)算法在高光譜影像分類中的應(yīng)用研究
發(fā)布時(shí)間：2018-11-16 20:35
【摘要】：高光譜遙感近年來發(fā)展迅速在許多領(lǐng)域得到了較好的應(yīng)用,逐漸成為對(duì)地物目標(biāo)進(jìn)行定量研究的一種重要方法。由于其光譜分辨率高和圖譜合一的特點(diǎn),令其在地表物質(zhì)的精細(xì)識(shí)別及分類等方面具有不可比擬的優(yōu)勢(shì),對(duì)高光譜影像進(jìn)行分類成為高光譜遙感應(yīng)用研究中的重要環(huán)節(jié)。高光譜影像數(shù)據(jù)高維數(shù)的特點(diǎn)也給影像分類帶來了困難,即對(duì)于高維特征數(shù)據(jù),在進(jìn)行監(jiān)督分類時(shí)需要大量帶標(biāo)記訓(xùn)練樣本來對(duì)分類模型進(jìn)行訓(xùn)練才能保證分類器的精度,也就是Hughes現(xiàn)象。通常采用降維算法來降低高光譜影像數(shù)據(jù)的維數(shù)被廣泛應(yīng)用到高光譜影像分類中,通過降維的手段可以較好的解決Hughes難題。但常用的降維方法,往往局限在提取像元的淺層特征,不能得到像元的深層特征,這一定程度上限制了分類器的表現(xiàn)。魯棒的深層特征中往往包含像元的抽象結(jié)構(gòu)信息,這更有利于分類精度的提高。本文將深度學(xué)習(xí)應(yīng)用到高光譜影像分類中,嘗試提取更有利于分類的像元深度特征,同時(shí)分析三種常用深度學(xué)習(xí)算法的優(yōu)勢(shì)和特點(diǎn),著重研究將非監(jiān)督學(xué)習(xí)的堆棧式自編碼器和深度信念網(wǎng)運(yùn)用于高光譜影像像元的特征提取中來解決Hughes難題。研究流程為:首先證明高光譜影像中像元的非線性特性和分類時(shí)Hughes現(xiàn)象的發(fā)生以論證深度學(xué)習(xí)在高光譜影像分類中的適用性;其次通過與傳統(tǒng)降維方法進(jìn)行類比分析,得出深度學(xué)習(xí)理論中的自編碼器和受限波茲曼機(jī)兩種特征提取算法性能優(yōu)于傳統(tǒng)算法;然后采用模型參數(shù)調(diào)優(yōu),可視化分析,連接不同分類器下的精度評(píng)價(jià)等實(shí)驗(yàn)方法綜合分析,得到了堆棧式自編碼器和深度信念網(wǎng)下的最優(yōu)分類模型;最終對(duì)分類性能更優(yōu)的堆棧式自編碼器進(jìn)行進(jìn)一步的優(yōu)化,即在自編碼器中加入稀疏表示的限制條件以及引入GPU并行運(yùn)算,來進(jìn)一步提升分類精度和分類速度。本文將深度學(xué)習(xí)理論引入到高光譜影像分類中,通過非監(jiān)督的學(xué)習(xí)方法可利用大量的無標(biāo)簽數(shù)據(jù),并且能夠提取像元的深度特征。實(shí)驗(yàn)證明深度學(xué)習(xí)算法優(yōu)于傳統(tǒng)的特征提取算法,并得出基于深度學(xué)習(xí)的最優(yōu)分類模型,即堆棧式稀疏自編碼器,在兩種實(shí)驗(yàn)數(shù)據(jù)下分類精度可達(dá)到93.41%和94.92%,針對(duì)模型訓(xùn)練時(shí)間長(zhǎng)的問題,采用并行運(yùn)算的方式可使模型的訓(xùn)練速度提升7倍多。
[Abstract]:Hyperspectral remote sensing has been developed rapidly in many fields in recent years, and has gradually become an important method for quantitative study of ground objects. Because of its high spectral resolution and the unity of spectrum, it has an incomparable advantage in the fine recognition and classification of surface materials, so the classification of hyperspectral images has become an important link in the application of hyperspectral remote sensing. The characteristics of high dimension of hyperspectral image data also bring difficulties to image classification, that is, for high-dimensional feature data, a large number of labeled training samples are needed to train the classification model in order to ensure the accuracy of the classifier. This is the Hughes phenomenon. Usually reducing dimension algorithm to reduce the dimension of hyperspectral image data is widely used in hyperspectral image classification, by reducing the dimension of the method can solve the problem of Hughes. However, the commonly used dimensionality reduction methods are often limited in extracting shallow features of pixels, and can not obtain the deep features of pixels, which limits the performance of classifiers to some extent. Robust deep features often contain abstract structure information of pixels, which is more conducive to the improvement of classification accuracy. This paper applies depth learning to hyperspectral image classification, tries to extract pixel depth features that are more favorable to classification, and analyzes the advantages and characteristics of three commonly used depth learning algorithms. This paper focuses on the application of unsupervised learning stack self-encoder and depth belief net to feature extraction of hyperspectral image pixels to solve the Hughes problem. The research flow is as follows: firstly, the nonlinear characteristics of pixels in hyperspectral images and the occurrence of Hughes phenomenon in classification are proved to demonstrate the applicability of depth learning in hyperspectral image classification. Secondly, through the analogy analysis with the traditional dimensionality reduction method, it is concluded that the performance of the self-encoder and the constrained Boltzmann algorithm in the depth learning theory is better than the traditional algorithm. Then the optimal classification model under stack self-encoder and depth belief net is obtained by comprehensive analysis of model parameter tuning visual analysis and precision evaluation under different classifiers. Finally, the stack self-encoder with better classification performance is further optimized, that is, adding the constraint condition of sparse representation to the self-coder and introducing GPU parallel operation to further improve the classification accuracy and classification speed. In this paper, depth learning theory is introduced into hyperspectral image classification. Through unsupervised learning method, a large number of untagged data can be used, and depth features of pixels can be extracted. Experimental results show that the depth learning algorithm is superior to the traditional feature extraction algorithm, and the optimal classification model based on depth learning, i.e. stack sparse self-encoder, can achieve the classification accuracy of 93.41% and 94.92% under two kinds of experimental data. In view of the long training time of the model, the training speed of the model can be increased more than 7 times by parallel operation.
【學(xué)位授予單位】：成都理工大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2017
【分類號(hào)】：P237
[Abstract]:Hyperspectral remote sensing has been developed rapidly in many fields in recent years, and has gradually become an important method for quantitative study of ground objects. Because of its high spectral resolution and the unity of spectrum, it has an incomparable advantage in the fine recognition and classification of surface materials, so the classification of hyperspectral images has become an important link in the application of hyperspectral remote sensing. The characteristics of high dimension of hyperspectral image data also bring difficulties to image classification, that is, for high-dimensional feature data, a large number of labeled training samples are needed to train the classification model in order to ensure the accuracy of the classifier. This is the Hughes phenomenon. Usually reducing dimension algorithm to reduce the dimension of hyperspectral image data is widely used in hyperspectral image classification, by reducing the dimension of the method can solve the problem of Hughes. However, the commonly used dimensionality reduction methods are often limited in extracting shallow features of pixels, and can not obtain the deep features of pixels, which limits the performance of classifiers to some extent. Robust deep features often contain abstract structure information of pixels, which is more conducive to the improvement of classification accuracy. This paper applies depth learning to hyperspectral image classification, tries to extract pixel depth features that are more favorable to classification, and analyzes the advantages and characteristics of three commonly used depth learning algorithms. This paper focuses on the application of unsupervised learning stack self-encoder and depth belief net to feature extraction of hyperspectral image pixels to solve the Hughes problem. The research flow is as follows: firstly, the nonlinear characteristics of pixels in hyperspectral images and the occurrence of Hughes phenomenon in classification are proved to demonstrate the applicability of depth learning in hyperspectral image classification. Secondly, through the analogy analysis with the traditional dimensionality reduction method, it is concluded that the performance of the self-encoder and the constrained Boltzmann algorithm in the depth learning theory is better than the traditional algorithm. Then the optimal classification model under stack self-encoder and depth belief net is obtained by comprehensive analysis of model parameter tuning visual analysis and precision evaluation under different classifiers. Finally, the stack self-encoder with better classification performance is further optimized, that is, adding the constraint condition of sparse representation to the self-coder and introducing GPU parallel operation to further improve the classification accuracy and classification speed. In this paper, depth learning theory is introduced into hyperspectral image classification. Through unsupervised learning method, a large number of untagged data can be used, and depth features of pixels can be extracted. Experimental results show that the depth learning algorithm is superior to the traditional feature extraction algorithm, and the optimal classification model based on depth learning, i.e. stack sparse self-encoder, can achieve the classification accuracy of 93.41% and 94.92% under two kinds of experimental data. In view of the long training time of the model, the training speed of the model can be increased more than 7 times by parallel operation.
【學(xué)位授予單位】：成都理工大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2017
【分類號(hào)】：P237
【參考文獻(xiàn)】
相關(guān)期刊論文 前10條
1 方敏;王君;王紅艷;李天涯;;應(yīng)用監(jiān)督近鄰重構(gòu)分析的高光譜遙感數(shù)據(jù)特征提取[J];紅外與激光工程;2016年10期
2 黃鴻;鄭新磊;羅甫林;;融合加權(quán)均值濾波與流形重構(gòu)保持嵌入的高光譜影像分類[J];光子學(xué)報(bào);2016年10期
3 李新國(guó);黃曉晴;;一種基于DBN的高光譜遙感圖像分類方法[J];電子測(cè)量技術(shù);2016年07期
4 戴曉愛;郭守恒;任m，
本文編號(hào)：2336576
本文鏈接：http://sikaile.net/kejilunwen/dizhicehuilunwen/2336576.html
最近更新
教材專著
