面向机载高光谱数据的3D-CNN亚热带森林树种分类

doi:10.11707/j.1001-7488.20201110

Abstract

Abstract:

Objective: This study was implemented to explore the potential of deep convolutional neural networks in airborne hyperspectral data classification,so as to improve the classification accuracy of forest tree species in subtropical regions. Method: The aeronautical hyperspectral data of Nanning Gaofeng forest farm in Guangxi Zhuang Autonomous Region obtained by the LiCHy system of Chinese Academy of Forestry was used. The CNN model used in this paper aimed to deal with hyperspectral image analysis problems in an end-to-end manner. It could take raw data as input,without dimension reduction or feature screening,eliminating the need for traditional classification methods to manually feature selection in different degrees. The 3D convolutional layers in the network could extract spectral features and spatial features simultaneously,learn the local signal changes in the spatial and spectral dimensions of the feature cube,and classify them with important recognition features to improve the discriminating ability of hyperspectral images. For the problem of high dimensionality of airborne hyperspectral data and relatively few training samples,the CNN model was optimized to avoid over-fitting. Result: Compared with the traditional feature selection and object-oriented segmentation method,CNN could obtain a higher classification accuracy,the overall accuracy reached 98.38%,Kappa coefficient was 0.98. Compared with support vector machine combined with random forest (RF) feature selection classification,the overall accuracy was improved by 8.82%,and the Kappa coefficient was increased by 0.11. In the case of small sample training (75% reduction in training samples size),the overall accuracy still reached 95.89%,and the Kappa coefficient was 0.94. Conclusion: The three-dimensional convolutional neural network could fully utilize the rich information in the image processing of the feature extraction and classification of airborne hyperspectral imagery,which could achieve high-precision discrimination of subtropical forest tree species; in addition,reasonable network structure and training strategy (adding the Dropout layer) could greatly improve the network training speed and still get good results in small sample training,and could achieve efficient and accurate classification of forest species.

Key words: hyperspectral remote sensing, convolutional neural network, tree species classification, small sample

CLC Number:

S757

Lin Zhao,Xiaoli Zhang,Yanshuang Wu,Bin Zhang. Subtropical Forest Tree Species Classification Based on 3D-CNN for Airborne Hyperspectral Data[J]. Scientia Silvae Sinicae, 2020, 56(11): 97-107.

Figures/Tables 12

Fig.1

Table 1

Fig.2

Fig.3

Table 2

Table 3

Table 4

Table 5

Fig.4

Fig.5

Fig.6

Fig.7

References 0

	李竺强, 朱瑞飞, 高放, 等. 三维卷积神经网络模型联合条件随机场优化的高光谱遥感影像分类. 光学学报, 2018, 38 (8): 404- 413.
	Li Z Q , Zhu R F , Gao F , et al. Hyperspectral remote sensing image classification based on three-dimensional convolution neural network combined with conditional random field optimization. Acta Optica Sinica, 2018, 38 (8): 404- 413.
	荚文, 庞勇, 岳彩荣, 等. 机载AISA EagleⅡ高光谱数据处理——以额济纳旗试验区为例. 遥感技术与应用, 2016, 31 (3): 504- 510.
	Jia W , Pang Y , Yue C R , et al. The procesing of airborne AISA EagleⅡ data in Ejina Banner study area. Remote Sensing Technology and Application, 2016, 31 (3): 504- 510.
	毛学刚, 陈文曲, 魏晶昱, 等. 分割尺度对面向对象树种分类的影响及评价. 林业科学, 2017, 53 (12): 73- 83. doi: 10.11707/j.1001-7488.20171208
	Mao X G , Chen W Q , Wei J Y , et al. Effect and evaluation of segmentation scale on object-based forest species classification. Scientia Silvae Sinicae, 2017, 53 (12): 73- 83. doi: 10.11707/j.1001-7488.20171208
	张莹, 张晓丽, 李宏志, 等. 基于改进转换分离度特征选择规则的土地覆盖分类比较. 林业科学, 2018, 54 (8): 88- 98.
	Zhang Y , Zhang X L , Li H Z , et al. A comparison of landcover classification based on the improved transformed divergence analysis. Scientia Silvae Sinicae, 2018, 54 (8): 88- 98.
	Bannari A , Morin D , Bonn F , et al. A review of vegetation indices. Remote Sensing Reviews, 1995, 13 (1/2): 95- 120.
	Chen Y , Jiang H , Li C , et al. Deep feature extraction and classification of hyperspectral images based on convolutional neural networks. IEEE Transactions on Geoscience and Remote Sensing, 2016, 54 (10): 6232- 6251. doi: 10.1109/TGRS.2016.2584107
	Girshick R , Donahue J , Darrell T , et al. Rich feature hierarchies for accurate object detection and semantic segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2014, 580- 587.
	Gitelson A A , Merzlyak M N . Optical properties and nondestructive estimation of anthocyanin content in plant leaves. Photochem Photobiol, 2001, 74 (1): 38- 45.
	Haboudane D , Miller J R , Tremblay N , et al. Integrated narrow-band vegetation indices for prediction of crop chlorophyll content for application to precision agriculture. Remote Sensing of Environment, 2002, 81 (2/3): 416- 426.
	Hernández-Clemente R , Navarro-Cerrillo R M , Suárez L , et al. Assessing structural effects on PRI for stress detection in conifer forests. Remote Sensing of Environment, 2011, 115 (9): 2360- 2375. doi: 10.1016/j.rse.2011.04.036
	Hinton G E , Osindero S , Teh Y W . A fast learning algorithm for deep belief nets. Neural Computation, 2006a, 18 (7): 1527- 1554. doi: 10.1162/neco.2006.18.7.1527
	Hinton G E , Salakhutdinov R R . Reducing the dimensionality of data with neural networks. Science, 2006b, 313 (5786): 504- 507. doi: 10.1126/science.1127647
	Hinton G E, Srivastava N, Krizhevsky A, et al. 2012. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv: 1207.0580.
	Ji S , Xu W , Yang M , et al. 3D convolutional neural networks for human action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013, 35 (1): 221- 231. doi: 10.1109/TPAMI.2012.59
	Kobayashi S , Sanga-Ngoie K . A comparative study of radiometric correction methods for optical remote sensing imagery: the IRC vs. other image-based C-correction methods. International Journal of Remote Sensing, 2009, 30 (2): 285- 314.
	Krizhevsky A , Sutskever I , Hinton G E . Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 2012, 25 (2): 1097- 1105.
	Le Roux N , Bengio Y . Deep belief networks are compact universal approximators. Neural Computation, 2010, 22 (8): 2192- 2207. doi: 10.1162/neco.2010.08-09-1081
	Li Y , Zhang H , Shen Q . Spectral-spatial classification of hyperspectral imagery with 3D convolutional neural network. Remote Sensing, 2017, 9 (1): 67.
	Lichtenthaler H K , Lang M , Sowinska M , et al. Detection of vegetation stress via a new high resolution fluorescence imaging system. Journal of Plant Physiology, 1996, 148 (5): 599- 612.
	Makantasis K , Karantzalos K , Doulamis A , et al. Deep supervised learning for hyperspectral data classification through convolutional neural networks. Geoscience and Remote Sensing Symposium (IGARSS), IEEE International, 2015, 4959- 4962.
	Merzlyak M N , Gitelson A A , Chivkunova O B , et al. Non-destructive optical detection of pigment changes during leaf senescence and fruit ripening. Physiologia Plantarum, 1999, 106 (1): 135- 141.
	Simonyan K, SZisserman A. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint: 1409.1556.
	Sims D A , Gamon J A . Relationships between leaf pigment content and spectral reflectance across a wide range of species, leaf structures and developmental stages. Remote Sensing of Environment, 2002, 81 (2/3): 337- 354.
	Szegedy C , Liu W , Jia Y , et al. Going deeper with convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, 1- 9.
	Tran D , Bourdev L , Fergus R T , et al. Learning spatiotemporal features with 3d convolutional networks. Proceedings of the IEEE International Conference on Computer Vision, 2015, 4489- 4497.
	Vincent P , Larochelle H , Lajoie I , et al. Stacked denoising autoencoders: learning useful representations in a deep network with a local denoising criterion. Journal of Machine Learning Research, 2010, 11 (12): 3371- 3408.
	Xue J , Su B . Significant remote sensing vegetation indices: a review of developments and applications. Journal of Sensors, 2017, 2017, 1- 17.
	Yue J , Zhao W , Mao S , et al. Spectral-spatial classification of hyperspectral images using deep convolutional neural networks. Remote Sensing Letters, 2015, 6 (6): 468- 477.

参数Parameter	数值Value
光谱范围Spectral range	400~1 000 nm
光谱分辨率Spectral resolution	3.3 nm
视场角Field of view	37.7°
瞬间视场角Instantaneous field of view	0.646 m·rad
焦距Focal length	18.1 mm
波段数Number of band	125
空间像元数Number of spatial pixel	1 024
光谱采样间隔Sampling interval in spectral	4.6 nm
量化值Bit depth	12 bits

网络层 Layer	卷积核大小 Filter size	卷积核数量 Number of filter	步长 Stride	输出大小 Output size	特征立方体数量 Feature volumes	参数数量 Number of parameters
输入层Input	—	—		11×11×125	1	0
3D卷积层1 3D_CONV Layer1	3×3×7	1×N	1, 1, 1	9×9×119	4	256
3D卷积层2 3D_CONV Layer2	3×3×7	N×2N	1, 1, 1	7×7×113	8	2 024
3D卷积层3 3D_CONV Layer3	3×3×7	2N×N×4N	1, 1, 1	5×5×117	16	8 080
3D卷积层4 3D_CONV Layer4	3×3×7	8N×N	1, 1, 1	3×3×111	32	32 288
全连接层 Fully_Connect_Layer	—	—	—	1×1×1	128	3 723 392
Softmax层 Softmax	—	—	—	1×1×1	12	1 548

特征Feature	数量Number	方法Method
光谱变换 Spectral transform	5	ICA(独立成分分析)变换前5个主成分ICA (independent component analysis)top 5 principal components
光谱指数 Spectral indices	9	NDVI(Bannari et al., 1995)、PSRI(Merzlyak et al., 1999)、MRESRI(Sims et al., 2002)、 MRENDVI(Sims et al., 2002)、GNDVI(Lichtenthaler et al., 1996)、 PRI(Hernández-Clemente et al., 2011)、SIPI(Haboudane et al., 2002)、 ARI1(Gitelson et al., 2001)、VOG1(Xue et al., 2017)
纹理特征 Texture feature	24	高光谱图像的第482、550和650波段通过GLCM(灰度共生矩阵)获取纹理特征, 纹理窗口大小为17×17，得到24个纹理特征变量The 482^th, 550^th and 650^th bands of hyperspectral images obtain texture features by GLCM(gray level co-occurrence matrix). The texture window size is 17×17, and 24 texture feature variables are obtained
高度特征Height feature	1	利用LiCHy系统同步获取的LiDAR数据得到数字冠层高度模型作为高度特征变量，以确定各树种的高度信息Using the LiDAR data obtained synchronously by LiCHy system, the digital canopy height model is obtained as the height characteristic variable to determine the height information of each tree species

序号 No.	地物类别Categories	样本Samples
序号 No.	地物类别Categories	训练样本Trian	验证样本Validation	测试样本Test
1	杉木Cunninghamia lanceolata	8 074	8 074	64 594
2	马尾松Pinus massoniana	534	534	4 272
3	湿地松Pinus elliottii	417	417	3 340
4	巨尾桉Eucalyptus grandis × E.urophylla	1 493	1 493	11 945
5	尾叶桉Eucalyptus urophylla	4 027	4 027	32 220
6	红椎Castanopsis hystrix	2 546	2 546	20 369
7	米老排Mytilaria laosensis	300	300	2 400
8	油茶Camellia oleifera	832	832	6 659
9	其他阔叶树Other broadleaf forest	1 074	1 074	8 593
10	道路Road	1 047	1 047	8 378
11	采伐迹地Cutting site	1 017	1 017	8 136
12	建筑用地Building land	7	7	54
	总计Total	21 369	21 369	170 960

项目Item	ICA+光谱指数+纹理特征 ICA+ spectral indices+ texture feature	随机森林特征筛选 Random forest feature screening	2D-CNN (W=9)	2D-CNN (W=27)	3D-CNN (W=5, N=4)	3D-CNN (W=9, N=4)	3D-CNN (W=11, N=4)	3D-CNN (W=11, N=8)	3D-CNN (W=13, N=8)
总体精度Overall accuracy(%)	86.66	89.56	75.16	91.62	93.90	96.82	98.38	98.44	98.37
平均精度Average accuracy(%)	74.73	80.03	59.96	81.53	93.64	97.39	98.80	98.74	98.92
Kappa系数(K×100)	83.03	86.87	67.49	89.29	92.27	95.99	97.96	98.03	97.94
杉木 Cunninghamia lanceolata	90.63	93.55	83.41	90.66	93.93	96.93	98.37	98.52	97.88
马尾松 Pinus massoniana	34.76	32.38	0	62.37	93.32	98.40	99.19	99.24	99.05
湿地松 Pinus elliottii	75.21	78.37	36.16	77.69	89.16	97.25	99.06	98.74	99.79
巨尾桉 Eucalyptus grandis × E.urophylla	97.83	94.02	68.12	96.43	97.10	98.35	99.28	99.61	99.44
尾叶桉 Eucalyptus urophylla	98.16	91.47	61.82	93.05	95.37	98.83	99.67	99.69	99.78
红椎 Castanopsis hystrix	85.59	70.71	57.23	86.86	85.25	89.61	94.32	94.88	95.36
米老排 Mytilaria laosensis	47.38	10.95	69.20	88.29	89.81	97.75	98.95	99.34	99.00
油茶 Camellia oleifera	98.72	98.60	79.14	91.49	97.62	99.44	99.72	99.91	99.75
其他阔叶树 Other broadleaf forest	75.27	75.20	79.52	95.35	94.03	95.99	97.46	95.43	97.44
道路Road	90.99	88.05	98.53	98.02	99.65	99.29	99.65	99.61	99.65
采伐迹地Cutting site	99.87	99.88	86.37	98.16	98.76	99.10	99.96	99.94	99.94
建筑用地Buliding land	65.91	63.64	0	0	89.66	97.78	100.00	100.00	100.00
需要训练参数的数量 Number of training parameters	—	—	251 244	202 092	507 172	458 020	3 767 588	7 618 172	10 386 724
预测需要的时间 Time required for predicting	—	—	35	39	29	44	62	87	92
训练所需要的时间 Time required for training	—	—	14.5	13.0	N/A	35.5	61.5	90.0	101.5

Subtropical Forest Tree Species Classification Based on 3D-CNN for Airborne Hyperspectral Data

RichHTML

PDF (PC)

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 12

References 0

Related Articles 3

Recommended Articles

Metrics

Comments

[1]	Guangqun Zhang,Yingjie Li,Hangjun Wang,Houkui Zhou. Forest Image Classification Based on Fine-Tuning CaffeNet [J]. Scientia Silvae Sinicae, 2020, 56(10): 121-128.
[2]	Zhiwei Lin,Qilu Ding,Jinfu Liu. Bird Species Identification Based on Deep Convolutional Network with Fusing Global and Local Features [J]. Scientia Silvae Sinicae, 2020, 56(1): 133-144.
[3]	Liu Jinli, Chen Zhao, Gao Jinping, Gao Xianlian, Sun Zhongqiu. Research on the Method of Determining the Optimal Segmentation Scale for Tree Species Classification of High-Resolution Image [J]. Scientia Silvae Sinicae, 2019, 55(11): 95-104.