Please wait a minute...
img

官方微信

遥感技术与应用  2019, Vol. 34 Issue (4): 704-711    DOI: 10.11873/j.issn.1004-0323.2019.4.0704
CNN 专栏     
基于DenseNet的无人机光学图像树种分类研究
林志玮1,2,3(),丁启禄1,黄嘉航2,涂伟豪1,胡典1,刘金福1,4
1. 福建农林大学计算机与信息学院, 福建 福州 350002
2. 福建农林大学林学院, 福建 福州 350002
3. 福建农林大学林学博士后流动站, 福建 福州 350002
4. 福建省高校生态与资源统计重点实验室, 福建 福州 350002
Study on Tree Species Classification of UAV Optical Image based on DenseNet
Zhiwei Lin1,2,3(),Qilu Ding1,Jiahang Huang2,Weihao Tu1,Dian Hu1,Jinfu Liu1,4
1. College of Computer and Information Science, Fujian Agriculture and Forestry University, Fuzhou 350002, China
2. College of Forestry, Fujian Agriculture and Forestry University, Fuzhou 350002, China
3. Forestry Post-doctoral station of Fujian Agriculture and Forestry University Key Laboratory for Ecology and Resource Statistics of Fujian Province, Fuzhou 350002, China
4. Key Laboratory for Ecology and Resource Statistics of Fujian Province, Fuzhou 350002, China
 全文: PDF(8632 KB)   HTML
摘要:

利用无人机航拍获得光学影像数据,结合深度学习理论,建立树种识别模型,以期为大规模树种识别提供一种新的方式。首先以福建安溪县为例,采用无人机获取20 m及40 m高度的航拍影像。其次,以树种为对象,对航拍影像进行分割,获得12种树种影像。最后,结合深度学习理论,采用DenseNet卷积神经网络建立树种识别模型,探讨不同航拍高度以及不同网络深度对树种识别的影响。结果表明:不同航拍高度的树种识别模型,其分类精度均达80%以上,最高精度为87.54%。从航拍影像解析度分析,随着航拍影像解析度的下降,模型识别精度呈现下降趋势,以20 m航拍影像数据建构的树种识别模型,其分类精度高于40 m模型;从模型网络深度分析,随着模型网络层数的增加,模型分类精度出现下降现象,DenseNet121模型分类精度高于DenseNet169模型分类精度。综上所述,基于无人机航拍影像,结合深度卷积神经网络,提出了新的树种识别方式,并以安溪县森林树种识别为例证明了该分类框架的有效性。

关键词: 无人机深度学习树种识别光学影像    
Abstract:

To provide a new idea for large-scale tree species identification, the UAV is used to obtain optical images, and is associated with the theory of deep learning to establish tree species recognition models. First, the Anxi County in Fujian Province is taken as an example, UAV was photographed at different heights of 20 m and 40 m to obtain aerial images of trees. Second, using the tree species as the object, aerial images were segmented to obtain 12 species of tree images. Finally, combined with the deep learning theory, DenseNet is used to establish the tree species recognition model, and the effects of different aerial heights and different depths of network on tree species recognition are discussed. The classification accuracy of tree species identification models with different aerial heights reached more than 80%, and the highest precision was 87.54%. From the analysis of the resolution of aerial image, with the decline of the resolution of aerial image, the accuracy of model presented a downward trend. The tree species recognition model constructed with 20m aerial image data had a higher classification accuracy than the 40m model. From the depth analysis of the network, with the increase of the number of network layers of the model, the classification accuracy of the model decreased. The accuracy of the DenseNet121 model was higher than that of the DenseNet169 model. Based on UAV aerial images and combined with deep convolutional neural network, a new tree species identification method was proposed. The identification of forest tree species in Anxi County was used as an example to prove the validity of the classification framework.

Key words: UAV    Deep learning    Tree identification    Optical image
收稿日期: 2018-04-28 出版日期: 2019-10-16
ZTFLH:  TP79  
基金资助: 海峡博士后交流资助计划;中国博士后科学基金面上项目(2018M632565);福建省自然科学基金项目(2016J01718)
作者简介: 林志玮(1981-),男,台湾台北人,硕士生导师,讲师,主要从事计算机科学、图像处理、图形识别以及机器学习等方面的研究。E-mail:cwlin@fafu.edu.cn
服务  
把本文推荐给朋友
加入引用管理器
E-mail Alert
RSS
作者相关文章  
林志玮
丁启禄
黄嘉航
涂伟豪
胡典
刘金福

引用本文:

林志玮,丁启禄,黄嘉航,涂伟豪,胡典,刘金福. 基于DenseNet的无人机光学图像树种分类研究[J]. 遥感技术与应用, 2019, 34(4): 704-711.

Zhiwei Lin,Qilu Ding,Jiahang Huang,Weihao Tu,Dian Hu,Jinfu Liu. Study on Tree Species Classification of UAV Optical Image based on DenseNet. Remote Sensing Technology and Application, 2019, 34(4): 704-711.

链接本文:

http://www.rsta.ac.cn/CN/10.11873/j.issn.1004-0323.2019.4.0704        http://www.rsta.ac.cn/CN/Y2019/V34/I4/704

航拍高度 20 m 40 m
图片数量 617 585
平均解析度 960×859 867×737
表1  不同航拍高度图像信息
图1  树种影像分割图
图2  Denseblock与transitionlayer
图3  DenseNet框架图
神经网络层 输出特征图尺寸 DenseNet121 DenseNet169
卷积层 112×112 7×7 卷积 7×7 卷积
池化层 56×56 3×3 最大池化 3×3 最大池化
Dense Block (1) 56×56 1 × 1 3 × 3 × 6 1 × 1 3 × 3 × 6
Transition Layer(1) 56×56 1×1卷积 1×1卷积
28×28 2×2 平均池化 2×2 平均池化
Dense Block (2) 28×28 1 × 1 3 × 3 × 12 1 × 1 3 × 3 × 12
Transition Layer(2) 28×28 1×1卷积 1×1卷积
14×14 2×2 平均池化 2×2 平均池化
Dense Block (3) 14×14 1 × 1 3 × 3 × 24 1 × 1 3 × 3 × 32
Transition Layer(3) 14×14 1×1卷积 1×1卷积
7×7 2×2 平均池化 2×2 平均池化
Dense Block (4) 7×7 1 × 1 3 × 3 × 16 1 × 1 3 × 3 × 32
分类层 1×1 7×7全局平均池化 7×7全局平均池化
softmax分类 softmax分类
表2  DenseNet121和DenseNet169模型参数
图4  DenseNet121和DenseNet169模型混淆矩阵图
航拍高度 20 m 40 m
DenseNet121 87.54% 84.38%
DenseNet169 86.23% 83.33%
表3  不同航拍高度DenseNet121和DenseNet169分类结果
1 Liu Xusheng , Zhao Xiaoli . Research Advances and Countermeasures of Remote Sensing Classification of Forest Vegetation[J]. Forest Resources Management, 2004(1):61-64.
1 刘旭升,张晓丽 .森林植被遥感分类研究进展与对策[J].林业资源管理,2004(1):61-64.
2 Zhang Chenkun , Han Min . Spectral-spatial Joint Classification of Hyperspectral Image with Edge-preserving Filtering[J]. Acta Automatica Sinica, 2018, 44(2): 280-288.
2 张成坤, 韩敏 . 基于边缘保持滤波的高光谱影像光谱-空间联合分类[J]. 自动化学报, 2018, 44(2): 280-288.
3 Liu Xiuying , Lin Hui , Xiong Jianli , et al .Band Selection from Hyperspectral Data of Forestry Species [J]. Remote Sensing Information,2005(4):41-44.
3 刘秀英,林辉,熊建利,等 .森林树种高光谱波段的选择[J].遥感信息,2005(4):41-44.
4 Wang Zhihui , Ding Lixia . Tree Species Discrimination based on Leaf-level Hyperspectral Characteristic Analysis[J]. Spectroscopy and Spectral Analysis,2010,30(7):1825-1829.
4 王志辉,丁丽霞 .基于叶片高光谱特性分析的树种识别[J].光谱学与光谱分析,2010,30(7):1825-1829.
5 Chen Zhuo , Ma Hongchao .Automatic Extracting and Modeling Approach of City Cloverleaf from Airborne LiDAR Data[J]. Acta Geodaetica et Cartographica Sinica, 2012, 41(2): 252-258.
5 陈卓,马洪超 . 基于机载LiDAR数据的大型立交桥自动提取与建模方法[J]. 测绘学报, 2012, 41(2): 252-258.
6 Liu Lihuan , Pang Yong , Fan Yiwen , et al .Fused Airborne Li DAR and Hyperspectral Data for Tree Species Identification in a Natural Temperate Forest[J]. Journal of Remote Sensing,2013,17(3):679-695.
6 刘丽娟,庞勇,范文义,等 .机载LiDAR和高光谱融合实现温带天然林树种识别[J].遥感学报,2013,17(3):679-695.
7 Kwak D A , Lee W K , Lee J H , et al . Detection of Individual Trees and Estimation of Tree Height Using LiDAR Data[J]. Journal of Forest Research, 2007, 12(6):425-434.
8 Chen Jianming , Chen Zhibo , Yang Meng , et al .Research on Tree Species Identification Algorithm based on Combination of Leaf Traditional Characteristics and Distance Matrix as well as Corner Matrix[J]. Journal of Beijing Forestry University,2017,39(2):108-116.
8 陈明健,陈志泊,杨猛,等 .叶片传统特征和距离矩阵与角点矩阵相结合的树种识别算法[J].北京林业大学学报,2017,39(2):108-116.
9 Iwata T , Saitoh T .Tree Recognition based on Leaf Images[C]∥IEEE, Sice Conference. 2014:2489-2494.
10 Krizhevsky A , Sutskever I , Hinton G E . ImageNet Classification with Deep Convolutional Neural Networks[C]∥International Conference on Neural Information Processing Systems. Curran Associates Inc. 2012:1097-1105.
11 Simonyan K , Zisserman A . Very Deep Convolutional Networks for Large-scale Image Recognition[J]. arXiv preprint arXiv: 1409.1556, 2014.
12 Szegedy C , Liu W , Jia Y , et al . Going Deeper with Convolutions[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2015: 1-9.
13 He K , Zhang X , Ren S , et al . Deep Residual Learning for Image Recognition[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 770-778.
14 Huang G , Liu Z , Maaten V D , et al . Densely Connected Convolutional Networks[C]⫽ arXiv Preprint arXiv: 1608.06993, 2016.
15 O'Shea K , Nash R . An Introduction to Convolutional Neural Networks[J]. arXiv preprint arXiv: 1511.08458, 2015.
16 Zeiler M D , Fergus R . Visualizing and Understanding Convolutional Networks[C]∥European Conference on Computer Vision. Springer, Cham, 2014: 818-833.
17 Luo W , Li Y , Urtasun R , et al . Understanding the Effective Receptive Field in Deep Convolutional Neural Networks[C]∥Advances in Neural Information Processing Systems. 2016: 4898-4906.
18 Geng M , Wang Y , Xiang T , et al . Deep Transfer Learning for Person Re-identification[J]. arXiv Preprint arXiv: 1611.05244, 2016.
19 Shao L , Zhu F , Li X . Transfer Learning for Visual Categorization: A Survey[J]. IEEE Transactions on Neural Networks and Learning Systems, 2015, 26(5): 1019-1034.
20 Xie M , Jean N , Burke M , et al . Transfer Learning from Deep Features for Remote Sensing and Poverty Mapping[J]. arXiv Preprint arXiv: 1510.00098, 2015.
[1] 刘天福,陈学泓,董琪,曹鑫,陈晋. 深度学习在GlobeLand30-2010产品分类精度优化中应用研究[J]. 遥感技术与应用, 2019, 34(4): 685-693.
[2] 田慧慧, 冯 莉, 赵璊璊, 郭 松, 董继伟. 无人机热红外城市地表温度精细特征研究 [J]. 遥感技术与应用, 2019, 34(3): 553-563.
[3] 徐梦竹, 徐佳, 邓鸿儒, 袁春琦. 基于全极化SAR影像的海岛地物分类[J]. 遥感技术与应用, 2019, 34(3): 647-654.
[4] 李伟, 唐伶俐, 吴昊昊, 腾格尔, 周梅. 轻小型无人机载激光雷达系统研制及电力巡线应用[J]. 遥感技术与应用, 2019, 34(2): 269-274.
[5] 李淑敏, 冯权泷, 梁其椿, 张学庆. 基于深度学习的国产高分遥感影像飞机目标自动检测[J]. 遥感技术与应用, 2018, 33(6): 1095-1102.
[6] 赵云,谢东海,邓磊,闫亚男,李博旭. 利用多角度影像计算BRDF的方法与系统实现[J]. 遥感技术与应用, 2018, 33(4): 741-749.
[7] 何艺,周小成,黄洪宇,许雪琴. 基于无人机遥感的亚热带森林林分株数提取[J]. 遥感技术与应用, 2018, 33(1): 168-176.
[8] 何海清,庞燕,陈晓勇. 面向遥感影像场景的深度卷积神经网络递归识别模型[J]. 遥感技术与应用, 2017, 32(6): 1078-1082.
[9] 周在明,杨燕明,陈本清. 基于无人机影像的滩涂入侵种互花米草植被信息提取与覆盖度研究[J]. 遥感技术与应用, 2017, 32(4): 714-720.
[10] 褚洪亮,肖青,柏军华,程娟. 基于无人机遥感的叶面积指数反演[J]. 遥感技术与应用, 2017, 32(1): 140-148.
[11] 张正健,李爱农,边金虎,赵伟,南希,雷光斌,谭剑波,夏浩铭,汪阳春,杜小林,林家元. 基于无人机的山地遥感观测平台及可靠性分析—以若尔盖试验为例[J]. 遥感技术与应用, 2016, 31(3): 417-429.
[12] 李爱农,边金虎,张正健,赵伟,南希,孙志宇,唐明坤,俄尕. 若尔盖高原区域碳收支参量多尺度遥感综合观测试验:科学目标与试验设计[J]. 遥感技术与应用, 2016, 31(3): 405-416.
[13] 张正健,李爱农,边金虎,赵伟,南希,靳华安,谭剑波. 基于无人机影像可见光植被指数的若尔盖草地地上生物量估算研究[J]. 遥感技术与应用, 2016, 31(1): 51-62.
[14] 李宜展,朱秀芳,张锦水,潘耀忠,李慕义. 与抽样相结合的县级作物遥感面积估算应用实例[J]. 遥感技术与应用, 2015, 30(5): 891-898.
[15] 宋耀鑫,张丹丹,唐伶俐,李传荣,马灵玲. 基于ASIFT算法的低重叠度无人机影像拼接方法[J]. 遥感技术与应用, 2015, 30(4): 725-730.