遥感高光谱成像能够获得丰富的地物光谱信息,这使得在传统的宽波段遥感中不可分辨的物质,在高光谱遥感中可以被分辨出来。高光谱图像具有“图谱合一”的特点,充分的利用高光谱图像中的光谱信息和空间信息是获得精确分类结果的前提。深度学习模型中的自编码神经网络能够实现高维数据的非线性降维,而卷积神经网络(Convolutional Neural Network, CNN)则能够自动的从图像中提取空间特征,基于此,本文提出了一种基于深度学习的Autoencoder-CNN高光谱图像分类方法。首先利用自编码神经网络对高光谱数据进行光谱维的降维,然后将卷积神经网络作为分类器,将待分类像元及其邻域像元一同作为卷积神经网络的输入,实现高光谱图像的空谱联合分类。 Remote sensing hyperspectral imaging can obtain abundant spectral information, which provides the possibility for the analysis of high precision terrain. The hyperspectral image has the characteristics of “map in one”, and the full use of spectral information and spatial information in hy- perspectral image is the premise of obtaining accurate classification results. Deep learning stack machine model in automatic encoding (Stack Auto-Encoder SAE) can effectively extract data in nonlinear information, and convolutional neural network (Convolutional Neural Network, CNN) can automatically extract features from the image. Based on this, this paper presents a classification method of hyperspectral images based on deep learning. Firstly, the spectral dimension of the hyperspectral data is reduced using automatic encoding machine, then convolutional neural network is used as the classifier, and the pixel and its neighborhood pixels are classified together as the input of the classifier, so as to realize the hyperspectral image classification with spectral space.
本实验使用两组高光谱数据集检验本文算法的分类性能,分别是IndianPines数据集和University of Pavia数据集。前者为AVIRIS采集的农业区高光谱图像,图像大小为145像素 × 145像素,共包含220个波段,去掉其中的20个水吸收严重的波段,得到包含200个波段的高光谱数据。后者为ROSIS采集
图2. 用于高光谱图像分类的卷积神经网络结构
的城区高光谱图像,图像大小为610像素 × 340像素,共包含115个波段,去除其中水吸收严重的12个波段,将其余103个波段作为高光谱数据。Indian Pines数据以及University of Pavia数据的假彩色图及标记模板如图3(a)、图3(b)和图4(a)、图4(b)所示。
为了进一步证明本文的算法确实有效,使用同样的方法对University of Pavia数据集进行分类。PCA降维方式同样保留前40个主成分,Autoencoder的结构为103-40-40-40,即使用深度为4的自动编码网络进行降维处理,降维后的光谱维度为40。Pavia大学数据集的类别数量以及训练和测试样本数量如表2所示,图4为分类的结果,图6为每一类地物分类的精度。从Pavia大学数据的分类结果可以得到与
图5. Indian Pines数据分类精度
图6. Pavia University数据分类精度
Class labels and train-test distribution of samples for the Indian Pines dataset
No
Name
Train
Test
1
Asphalt
10
36
2
Corn-notill
428
1000
3
Corn-minitill
200
630
4
Corn
47
197
5
Grass-pasture
150
333
6
Grass-tress
210
520
7
Grass-pasture-mowed
7
19
8
Hay-windrowed
119
360
9
Oats
6
14
10
Soybean-notill
300
672
11
Soybean-minitill
520
1935
12
Soybean-clean
100
493
13
Wheat
50
155
14
Woods
200
1065
15
Build-Grass-Tree-Drives
186
200
16
Stone-Steel-Towers
30
63
表1. Indian Pines高光谱影像的类别和样本数
Class labels and train-test distribution of samples for the University of Pavi
袁 林,胡少兴,张爱武,柴沙陀,王 兴. 基于深度学习的高光谱图像分类方法 A Classification Method for Hyperspectral Imagery Based on Deep Learning[J]. 人工智能与机器人研究, 2017, 06(01): 31-39. http://dx.doi.org/10.12677/AIRR.2017.61005
参考文献 (References)ReferencesLandgrebe, D. (2002) Hyperspectral Image Data Analysis. Signal Processing Magazine, 19, 17-28.
https://doi.org/10.1109/79.974718Wange, A., Lu, J., Cai, J., Wang, G. and Charm, T.J. (2015) Unsupervised Joint Feature Learning and Encoding for RGB-D Scene Labeling. IEEE Transactions on Image Process, 24, 4459-4473.
https://doi.org/10.1109/TIP.2015.2465133Huges, G. (1968) On the Mean Accuracy of Statistical Pattern Recognizers. IEEE Transactions on Information Theory, 14, 55-63. https://doi.org/10.1109/TIT.1968.1054102Kuo, B.-C. and Landgrebe, D.A. (2004) Nonparametric Weighted Feature Extraction for Classification for Classification. IEEE Transactions on Geoscience and Remote Sensing, 42, 1096-1105.
https://doi.org/10.1109/TGRS.2004.825578Landgrebe, D.A. (2003) Signal Theory Methods in Multispectral Remote Sensing. Wiley, Hoboken.Liu, Y., Gu, Y. and Zhang, Y. (2006) Hyperspectral Feature Extraction Using Selective PCA Based on Genetic Algorithm with Subgroups. 1st International Conference on Innovative Computing, Information and Control, Beijing, 30 August-1 September 2006, 652-656.Zhao, Y., Zhang, L., Li, P. and Huang, B. (2007) Classification of High Spatial Resolution Imagery Using Improved Gaussian Markov Random-Field-Based Texture Features. IEEE Transactions on Geoscience and Remote Sensing, 45, 1458-1468. https://doi.org/10.1109/TGRS.2007.892602Xu, H., Li, P. and Shen, Y. (2008) The Application of Invariant Moments to High Resolution Remote Sensing Image Classification. Remote Sensing for Land & Resources, 20, 9-13.Luo, B. and Zhang, L. (2014) Robust Autodual Morphological Profiles for the Classification of High-Resolution Satellite Images. IEEE Transactions on Geoscience and Remote Sensing, 52, 1451-1462.
https://doi.org/10.1109/TGRS.2013.2251468Bengio, Y., Lamblin, P., Popovici, D., et al. (2007) Greedy Layer-Wise Training of Deep Networks. Advances in Neural Information Processing Systems, 19, 153.Fukushima, K. (1988) Neocognitron: A Hierarchical Neural Network Capable of Visual Pattern Recognition. Neural Networks, 1, 119-130. https://doi.org/10.1016/0893-6080(88)90014-7LeCun, Y., Bottou, L., Bengio, Y. and Haffner, P. (1998) Gradient-Based Learning Applied to Document Recognition. Proceedings of the IEEE, 86, 2278-2323. https://doi.org/10.1109/5.726791Taigman, Y., Yang, M., Ranzato, M. and Wolf, L. (2014) DeepFace: Closing the Gap to Human-Level Performance in Face Verification. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Columbus, 23-28 June 2014, 1701-1708.Sainath, T.N., Mohamed, A.-R., Kingsbury, B. and Ramabhadran, B. (2013) Deep Convolutional Neural Networks for LVCSR. Proceedings of the 38th IEEE International Conference on Acoustics, Speech, and Signal Processing, Vancouver, 26-31 May 2013, 8614-8618. https://doi.org/10.1109/icassp.2013.6639347Chen, Y., Lin, Z., Zhao, X., Wang, G. and Gu, Y. (2014) Deep Learning-Based Classification of Hyperspectral Data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7, 2094-2107.
https://doi.org/10.1109/JSTARS.2014.2329330Lawrence, S., Giles, C.L., Tsoi, A.C. and Back, A.D. (1997) Face Recognition: A Convolutional Neural-Network Approach. IEEE Transactions on Neural Networks, 8, 98-113. https://doi.org/10.1109/72.554195