一种改进的高空间分辨率遥感影像森林类型深度学习精细分类方法:双支FCN-8s

doi:10.11707/j.1001-7488.20200306

摘要/Abstract

摘要：

目的: 对全卷积神经网络模型进行双支化改进，探索高空间分辨率遥感影像森林类型深度学习分类新方法，为提高森林资源遥感调查精度提供技术支撑。方法: 双支FCN-8s包含2个FCN-8s子模型，一个子模型基于R、G、B三波段特征，采用微调方式构建；另一个子模型基于五特征构建。将2个子模型8、16、32倍下的采样结果进行融合并分类，得到每个像元的类别。以旺业甸林场为研究区，采用GF-2卫星遥感影像提取标准化植被指数（NDVI），构建基于R+G+B三波段特征、R+G+B+NIR四波段特征和R+G+B+NIR+NDVI五特征的数据集，对双支FCN-8s优化方法的有效性进行定量评价。结果: 1）双支FCN-8s方法的总体分类精度为85.89%，Kappa系数为0.84；相比传统FCN-8s，双支FCN-8s方法可提高大部分森林类型的分类精度，尤其对油松、红松、白桦等类别改善效果明显。2）相对于传统基于特征优选的SVM模型而言，双支FCN-8s方法的总体分类精度由75%上升至85.89%，精度提升大于10%，各类别的分类效果均有改善。3）使用微调策略以及加入NDVI特征后，模型可有效改善油松、山杨及白桦等树种的分类效果。结论: 双支FCN-8s高空间分辨率遥感影像森林类型深度学习精细分类方法可有效提升森林类型的细分程度和分类精度。

关键词: 森林类型, 深度学习, 全卷积神经网络, GF-2, 双支FCN-8s

Abstract:

Objective: The study was conducted to perform the two-branch improvement on fully convolutional network models to explore a new method for the deep learning classification of remote sensing images of forest types with a high spatial resolution, with an aim to provide technical support for the improvement of the accuracy of remote sensing investigation of forest resources. Method: The two-branch FCN-8s contained two FCN-8s sub-models. One sub-model was constructed by fine-tuning based on RGB three-band features and the other sub-model was constructed based on five-band features. The sampling results of these two sub-models at 8, 16, and 32 magnifications were fused and classified to obtain the category of each pixel. Taking Wangyedian forest farm as the research area, the normalized difference vegetation index (NDVI) was extracted based on GF-2 satellite remote sensing images. In addition, the data set based on R+G+B three-band features, R+G+B + NIR four-band features and R+G+B + NIR + NDVI five-band features was constructed to evaluate the effectiveness of the two-branch FCN-8s optimization method quantitatively. Result: 1) The overall classification accuracy of the two-branch FCN-8s method is 85.89% and the Kappa coefficient is 0.84. Compared with the traditional FCN-8s, the dual-branch FCN-8s method can improve the classification accuracy of most forest types, especially for Chinese pine, Korean pine and white birch, and the improvement effects are significant. 2) Compared with the SVM model based on traditional features, the overall classification accuracy of the dual-branch FCN-8s method increaseds from 75% to 85.89%. The accuracy is improved by over 10%. The classification effects of each category are also improved. 3) With the fine-tuning strategy and NDVI features, the model can effectively improve the classification effects of Chinese pine, aspen and white birch. Conclusion: The deep learning precision classification method for forest types with dual-branch FCN-8s high-resolution remote sensing images can effectively improve the subdivision degree and the classification accuracy of forest types.

Key words: forest type, deep learning, FCN-8s, GF-2, two-branch FCN-8s

中图分类号:

S757

郭颖,李增元,陈尔学,张旭,赵磊,陈艳,王雅慧. 一种改进的高空间分辨率遥感影像森林类型深度学习精细分类方法:双支FCN-8s[J]. 林业科学, 2020, 56(3): 48-60.

Ying Guo,Zengyuan Li,Erxue Chen,Xu Zhang,Lei Zhao,Yan Chen,Yahui Wang. A Deep Learning Method for Forest Fine Classification Based on High Resolution Remote Sensing Images: Two-Branch FCN-8s[J]. Scientia Silvae Sinicae, 2020, 56(3): 48-60.

图/表 13

图1

图2

图3

表1

图4

图5

图6

表2

图7

表3

图8

表4

图9

参考文献 0

	陈广胜, 陈守玉, 景维鹏, 等. 基于遗传算法与全卷积网络的遥感影像像素级分类方法. 小型微型计算机系统, 2018. 39 (7): 1590- 1595. doi: 10.3969/j.issn.1000-1220.2018.07.041
	Chen G S , Chen S Y , Jing W P , et al. Remote sensing image pixel-level classification with genetic algorithm and fully convolutional network. Journal of Chinese Computer Systems, 2018. 39 (7): 1590- 1595. doi: 10.3969/j.issn.1000-1220.2018.07.041
	丁世飞, 齐丙娟, 谭红艳. 支持向量机理论与算法研究综述. 电子科技大学学报, 2011. 40 (1): 2- 10.
	Ding S F , Qi B J , Tan H Y . An overview on theory and algorithm of support vector machines. Journal of University of Electronic Science and Technology of China, 2011. 40 (1): 2- 10.
	方旭, 王光辉, 杨化超, 等. 结合均值漂移分割与全卷积神经网络的高分辨遥感影像分类. 激光与光电子学进展, 2018. 55 (2): 446- 454.
	Fang X , Wang G H , Yang H C , et al. High resolution remote sensing image classification combining with mean-shift segmentation and fully convolution neural network. Laser & Optoelectronics Progress, 2018. 55 (2): 446- 454.
	高永年, 张万昌. 遥感影像地形校正研究进展及其比较实验. 地理研究, 2008. 27 (02): 467- 478. doi: 10.3321/j.issn:1000-0585.2008.02.024
	Gao Y N , Zhang W C . Comparison test and research progress of topographic correction on remotely sensed data. Geographical Research, 2008. 27 (02): 467- 478. doi: 10.3321/j.issn:1000-0585.2008.02.024
	何小飞, 邹峥嵘, 陶超, 等. 联合显著性和多层卷积神经网络的高分影像场景分类. 测绘学报, 2016. 45 (9): 1073- 1080.
	He X F , Zou Z R , Tao C , et al. Combined saliency with multi-convolutional neural network for high resolution remote sensing scene classification. Acta Geodaetica et Cartographica Sinica, 2016. 45 (9): 1073- 1080.
	黄建文, 鞠洪波. 山区林地多季相TM影像综合分类研究. 福建林学院学报, 2004. 24 (2): 136- 139. doi: 10.3969/j.issn.1001-389X.2004.02.010
	Huang J W , Ju H B . Study on integrated vegetation classification by using multi-temporal TM data in mountainous area of northeast China. Journal of Fujian College of Forestry, 2004. 24 (2): 136- 139. doi: 10.3969/j.issn.1001-389X.2004.02.010
	雷光斌, 李爱农, 边金虎, 等. 基于阈值法的山区森林常绿、落叶特征遥感自动识别方法——以贡嘎山地区为例. 生态学报, 2014. 34 (24): 7210- 7221.
	Lei G B , Li A N , Bian J H , et al. An practical method for automatically identifying the evergreen and deciduous characteristic of forests at mountainous areas:a case study in Mt.Gongga Region. Acta Ecologica Sinica, 2014. 34 (24): 7210- 7221.
	刘大伟, 韩玲, 韩晓勇. 基于深度学习的高分辨率遥感影像分类研究. 光学学报, 2016. 36 (4): 306- 314.
	Liu D W , Han L , Han X Y . High spatial resolution remote sensing image classification based on deep learning. Acta Optica Sinica, 2016. 36 (4): 306- 314.
	任冲, 鞠洪波, 张怀清, 等. 多源数据林地类型的精细分类方法. 林业科学, 2016. 52 (6): 54- 65.
	Ren C , Ju H B , Zhang H Q , et al. Multi-source data for forest land type precise classification. Scientia Silvae Sinicae, 2016. 52 (6): 54- 65.
	徐慧敏. 2018.基于深度学习U-Net模型的高分辨率遥感影像分类方法研究.成都:西南交通大学硕士学位论文.
	Xu H M. 2018. Method research of high resolution remote sensing imagery classification based on U-Net model of deep learning. Chengdu: MS thesis of Southwest Jiaotong University.[in Chinese]
	尹宝才, 王文通, 王立春. 深度学习研究综述. 北京工业大学学报, 2015. 41 (1): 48- 59.
	Yin B C , Wang W T , Wang L C . Review of deep learning. Journal of Beijing University of Technology, 2015. 41 (1): 48- 59.
	张永宏, 夏广浩, 阚希, 等. 基于全卷积神经网络的多源高分辨率遥感道路提取. 计算机应用, 2018. 38 (7): 246- 251.
	Zhang Y H , Xia G H , Kan X , et al. Road extraction from multi-source high resolution remote sensing image based on fully convolutional neural network. Journal of Computer Applications, 2018. 38 (7): 246- 251.
	Agata H , Aneta L . Mapping forest type and tree species on a regional scale using multi-temporal Sentinel-2 data. Remote Sensing, 2019. 11 (8): 929.
	Bengio Y . Learning deep architectures for AI. Foundations and Trends^® R in Machine Learning, 2009. 2 (1): 1- 127.
	Feng Y , Lu D , Chen Q , et al. Examining effective use of data sources and modeling algorithms for improving biomass estimation in a moist tropical forest of the Brazilian Amazon. International Journal of Digital Earth, 2017. 10, 996- 1016.
	Foody G . Assessing the accuracy of remotely sensed data:principles and practices. The Photogrammetric Record, 2010. 25 (130): 204- 205.
	Fu G , Liu C , Zhou R , et al. Classification for high resolution remote sensing imagery using a fully convolutional network. Remote Sensing, 2017. 9 (5): 498.
	Gaetano R , Ienco D , Ose K , et al. A two-branch CNN architecture for land cover classification of PAN and MS imagery. Remote Sensing, 2018. 10 (11): 1746. doi: 10.3390/rs10111746
	Haralick R M , Shanmugam K , Dinstein I . Textural features for image classification. Studies in Media and Communication, 1973. SMC-3 (6): 610- 621.
	He K , Zhang X , Ren S , et al. Identity mappings in deep residual networks. European Conference on Computer Vision, 2016. 2016 (4): 630- 645.
	Kingma D, Ba J. 2014. Adam: a method for stochastic optimization. arXiv preprint arXiv: 1412.6980.
	Längkvist M , Kiselev A , Alirezaie M , et al. Classification and segmentation of satellite ortho-imagery using convolutional neural networks. Remote Sensing, 2016. 8 (4): 329.
	Liu X , Jiao L , Zhao J , et al. Deep multiple instance learning-based spatial-spectral classification for PAN and MS imagery. IEEE Transactions on Geoscience and Remote Sensing, 2018. 56 (1): 461- 473.
	Long J , Shelhamer E , Darrell T . Fully Convolutional networks for semantic segmentation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2014. 39 (4): 640- 651.
	Marmanis D , Schindler K , Wegner J D , et al. Classification with an edge:improving semantic image segmentation with boundary detection. ISPRS Journal of Photogrammetry and Remote Sensing, 2018. 135, 158- 172. doi: 10.1016/j.isprsjprs.2017.11.009
	Nogueira K , Penatti , Otávio A B , et al. Towards better exploiting convolutional neural networks for remote sensing scene classification. Pattern Recognition, 2016. 61, 539- 556.
	Rezaee M , Mahdianpari M , Zhang Y , et al. Deep convolutional neural network for complex wetland classification using optical remote sensing imagery. IEEE Journal of Selected Topics in Applied Earch Observations and Remote Sensing, 2018. 11 (9): 3030- 3039. doi: 10.1109/JSTARS.2018.2846178
	Sherrah J. 2016. Fully convolutional networks for dense semantic labelling of high resolution aerial imagery. arXiv preprint arXiv: 1606.02585.
	Sun W , Wang R . Fully convolutional networks for semantic segmentation of very high resolution remotely sensed images combined with DSM. IEEE Geoscience and Remote Sensing Letters, 2018. 99, 1- 5.
	Tan Y H , Xiong S Z , Li Y S . Automatic extraction of built-up areas from panchromatic and multispectral remote sensing images using double-stream deep convolutional neural networks. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2018. 11 (11): 3988- 4004.
	Vapnik V , Lerner A . Pattern recognition using generalized portrait method. Automation & Remote Control, 1963. 24, 774- 780.
	Wang Y , Lu D . Mapping torreyagrandis spatial distribution using high spatial resolution satellite imagery with the expert rules-based approach. Remote Sensing, 2017. 9 (6): 564.
	Wang X , Wang M , Wang S , et al. Extraction of vegetation information from visible unmanned aerial vehicle images. Transactions of the Chinese Society of Agricultural Engineering, 2015. 31 (5): 152- 159.
	Xie Z , Chen Y , Lu D , et al. Classification of land cover, forest, and tree species classes with Ziyuan-3 multispectral and stereo data. Remote Sensing, 2019. 11 (2): 164.

一级 Class one	二级 Class two	三级 Class three
林地Forest land	有林地Woodland	油松Pinus tabulaeformis
		华北落叶松Larix principis-rupprechtii
		红松Pinus koraiensis
		白桦Betula platyphylla
		山杨Populus davidiana
		蒙古栎Quercus mongolica
非林地Non-forest land	灌木林地Shrub land	/
	其他林地Other forest land	/
	耕地Cultivated land	/
	草地Grassland	/
	建设用地Construction land	/
	其他非林地Other non-forest land	/

地类 Land cover type	双支FCN-8s Two-branch FCN-8s		FCN-8s		SVM
地类 Land cover type	生产者精度 Producer accuracy(%)	用户精度 User accuracy(%)	生产者精度 Producer accuracy(%)	用户精度 User accuracy(%)	生产者精度 Producer accuracy(%)	用户精度 User accuracy(%)
油松Pinus tabulaeformis	94.59	87.50	90.54	82.72	79.73	81.94
华北落叶松Larix principis-rupprechtii	89.86	93.94	94.20	87.84	88.41	70.93
红松Pinus koraiensis	100.00	93.75	100.00	93.75	93.33	100.00
白桦Betula platyphylla	50.00	82.14	26.09	80.00	56.52	68.42
蒙古栎Quercus mongolica	96.43	90.00	100.00	82.35	96.43	79.41
山杨Populus davidiana	40.00	33.33	40.00	26.09	66.67	43.48
其他林地Other forest land	50.00	38.10	56.25	42.86	31.25	62.50
耕地Cultivated land	94.44	87.18	97.22	87.50	97.22	56.45
建设用地Construction land	100.00	100.00	100.00	100.00	97.50	97.50
灌木林地Shrub land	91.30	80.77	86.96	95.24	21.74	100.00
草地Grassland	91.67	100.00	75.00	100.00	50.00	100.00
其他非林地Other non-forest land	100.00	90.91	93.33	93.33	53.33	100.00
总体精度Overall accuracy(%)	85.89		82.67		75.00
Kappa系数Kappa coefficient	0.841 4		0.805 0		0.716 8

地类 Land cover type	双支FCN-8s Two-branch FCN-8s		双支FCN-8s不使用微调策略 Two-branch FCN-8s without finetune strategy		双支FCN-8s不加入NDVI特征 Two-branch FCN-8s without NDVI
地类 Land cover type	生产者精度 Producer accuracy(%)	用户精度 User accuracy(%)	生产者精度 Producer accuracy(%)	用户精度 User accuracy(%)	生产者精度 Producer accuracy(%)	用户精度 User accuracy(%)
油松Pinus tabulaeformis	94.59	87.50	82.43	88.41	86.49	84.21
华北落叶松Larix principis-rupprechtii	89.86	93.94	89.86	82.67	89.86	87.32
红松Pinus koraiensis	100.00	93.75	100.00	100.00	100.00	100.00
白桦Betula platyphylla	50.00	82.14	36.96	73.91	56.52	70.27
蒙古栎Quercus mongolica	96.43	90.00	92.86	74.29	96.43	90.00
山杨Populus davidiana	40.00	33.33	46.67	36.84	26.67	30.77
其他林地Other forest land	50.00	38.10	31.25	25.00	43.75	38.89
耕地Cultivated land	94.44	87.18	91.67	78.57	97.22	70.00
建设用地Construction land	100.00	100.00	95.00	100.00	82.50	100.00
灌木林地Shrub land	91.30	80.77	82.61	76.00	91.30	95.45
草地Grassland	91.67	100.00	75.00	81.82	66.67	100.00
其他非林地Other non-forest land	100.00	90.91	96.67	90.63	96.67	93.55
总体精度Overall accuracy(%)	85.89		79.46		81.93
Kappa系数Kappa coefficient	0.841 4		0.769 3		0.796 3

[1]	刘金丽, 陈钊, 高金萍, 高显连, 孙忠秋. 高分影像树种分类的最优分割尺度确定方法[J]. 林业科学, 2019, 55(11): 95-104.
[2]	田静, 邢艳秋, 姚松涛, 曾旭婧, 焦义涛. 基于元胞自动机和BP神经网络算法的Landsat-TM遥感影像森林类型分类比较[J]. 林业科学, 2017, 53(2): 26-34.
[3]	彭舜磊;王得祥. 秦岭主要森林类型近自然度评价[J]. 林业科学, 2011, 47(1): 135-142.
[4]	董希斌姜帆. 帽儿山不同森林类型生物多样性恢复效果分析[J]. 林业科学, 2008, 44(12): 77-82.
[5]	许克福吴泽民管露露. 马鞍山市不同城市森林类型树种组成及多样性[J]. 林业科学, 2008, 44(10): 142-147.
[6]	王原吴泽民张磊赵霞管露露. 马鞍山城市森林景观镶嵌与其城郊分布梯度格局研究[J]. 林业科学, 2007, 43(3): 51-58.
[7]	陈尔学李增元谭炳香梁毓照张则路. 高光谱数据森林类型统计模式识别方法比较评价[J]. 林业科学, 2007, 43(1): 84-89.
[8]	汤孟平唐守正洪玲霞雷相东张会儒. 森林类型多样性最大覆盖模型与一种可行算法[J]. 林业科学, 2003, 39(3): 70-75.
[9]	金森. 黑龙江省林火规律研究 Ⅲ.大尺度水平林火与森林类型之间的关系研究[J]. 林业科学, 2002, 38(4): 171-175.