随机森林方法支持的复杂地形区土地利用/土地覆被分类研究

doi:10.12082/dpxxkx.2019.180346

摘要/Abstract

摘要：

随机森林方法目前已经成为遥感分类机器学习中一种有效方法,探索基于中等分辨率的Landsat卫星数据与随机森林方法相结合对复杂地形区长时间序列数据的获取及土地利用/土地覆被变化及模拟研究是非常有意义的。本文基于Landsat8OLI卫星多光谱数据,采用随机森林分类方法对青海省湟水流域复杂地形区土地利用类型进行了分类研究。针对复杂地形区域的情况,将研究区进行地理分区,根据每个分区的特点,选择相应的地形特征参数,并通过提取Landsat 8数据的光谱信息与纹理信息构建最优特征集,探索随机森林方法在复杂地形区土地利用分类的适用性。结果表明：使用Landsat8OLI数据进行随机森林分类,能较好地得到湟水流域复杂地形区域的土地利用类型结果;光谱、地形及纹理信息的结合在不同分区的表现结果不同。在脑山区光谱与地形信息结合能使随机森林分类效果最佳,总体精度达到91.33%,Kappa系数为0.886;而在浅山区与川水区综合考虑光谱、地形、纹理信息进行随机森林分类效果最佳,浅山区与川水区总体精度分别达到92.09%和87.85%,Kappa系数分别为0.902和0.859;利用随机森林算法进行优化选择纹理特征组合可以在保证分类精度的同时能够快速地提取土地利用类型信息,为复杂地形区土地利用类型的区分提供了实际可行的方法。

关键词: 随机森林算法, 复杂地形区, 土地利用/土地覆盖分类, 特征选择, 湟水流域

Abstract:

Random forest classification has become an effective method in remote sensing classification of machine learning. It is of great significance to combine the Landsat satellite data and random forest method to obtain long time series data in the complex terrain areas and to explore its land use/land cover change. Based on the multi-spectral data of landsat8 OLI satellite, this paper adopted the random forest classification method to classify the land use types of Huangshui basin complex topography areas in Qinghai province. According to the characteristics of complex terrain areas, the study area was divided into different geographical regions. The topographic parameters were then selected, and the optimal feature collection was constructed by extracting spectral and texture information of Landsat8 data. The objective of this papers was to explore the applicability of random forest methods in land use classification on the complex topographic regions. The results showed that RFC classification with the landsat8 OLI data can be well used to obtain the land use types in the Huangshui basin. The combination of spectral, topographic, and texture information performed differently in different areas. In the middle and high mountain areas, the combination of spectral and topographic information can obtain the best results in the random forest classification with the overall accuracy of 91.33% and Kappa coefficient of 0.886. In the shallow mountain areas and valley plain, however, the random forest classification can obtain the best results by combining spectral, topographic, and texture information with the overall accuracy of 92.09% and 87.85% and Kappa coefficient of 0.902 and 0.859, respectively. Using the random forest algorithm to optimize the selection of texture feature combination can extract the land use type information quickly and ensure its accuracy. Random forest classification combined multi-source information can be used effectively to classify land use types, which can provide some enlightenment and reference values for the renewal of land use status and the development of social economy in the study area.

Key words: random forest algorithm, complex terrain area, land use/land cover classification, feature selection, the Huangshui basin

马慧娟, 高小红, 谷晓天. 随机森林方法支持的复杂地形区土地利用/土地覆被分类研究[J]. 地球信息科学学报, 2019, 21(3): 359-371.DOI:10.12082/dpxxkx.2019.180346

Huijuan MA, Xiaohong GAO, Xiaotian GU. Random Forest Classification of Landsat 8 Imagery for the Complex Terrain Area based on the Combination of Spectral, Topographic and Texture Information[J]. Journal of Geo-information Science, 2019, 21(3): 359-371.DOI:10.12082/dpxxkx.2019.180346

图/表 16

图1

表1

图2

图3

表2

表3

表4

表5

图4

表6

图5

表7

图6

表8

表9

图7

参考文献 24

[1]	杨超. 基于混合像元分解决策树的土地利用/覆被分类研究[D].云南:云南师范大学,2017.
	[ Yang C.Research on land use/land cover classification based on decision tree of mixed pixel decomposition[D]. Yunnan: Yunnan normal university, 2017.
[2]	乔平林,张继贤,卢秀山,等.基于神经网络的区域生态环境分类方法研究[J].地理与地理信息科学,2004,20(2):97-99. doi: 10.3969/j.issn.1672-0504.2004.02.023
	[ Qiao P L, Zhang J X, lu X S, et al. Regional ecological environment classification methods based on neural networks[J]. Geographic and Geographic Information Science, 2004,20(2):97-99. ] doi: 10.3969/j.issn.1672-0504.2004.02.023
[3]	徐凯健,田庆久,杨闫君,等.遥感土地覆被分类的空间尺度响应研究[J].地球信息科学学报,2018,20(2):246-253.
	[ Xu K J, tian Q J, Yang Y J, et al. Research on spatial scale response of remote sensing land cover classification[J]. Journal of Geo-information Science, 2018,20(2):246-253. ]
[4]	杨雪峰,王雪梅.基于决策树的多角度遥感影像分类[J].地球信息科学学报, 2016,18(3):416-422. doi: 10.3724/SP.J.1047.2016.00416
	[ Yang X F, wang X M. Classification of multi-angle remote sensing images based on decision tree[J]. Journal of Geoinformation Science, 2016,18(3):416-422. ] doi: 10.3724/SP.J.1047.2016.00416
[5]	Pal M. Random forests for land cover classification[C]//Geoscience and Remote Sensing Symposium,2003.IGARSS'03. Proceedings of 2003 IEEE International,IEEE, 2004(6):3510-3512.
[6]	Chan C W, Desiré P.Evaluation of Random Forest and Adaboost tree-based ensemble classification and spectral band selection for ecotope mapping using airborne hyperspectral imagery[J]. Remote Sensing of Environment, 2008,112(6):2999-3011. doi: 10.1016/j.rse.2008.02.011
[7]	Waske B, Braun M.Classifier ensembles for land cover mapping using multitemporal SAR imagery[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2009,64(5):450-457. doi: 10.1016/j.isprsjprs.2009.01.003
[8]	田绍鸿,张显峰.采用随机森林法的天绘数据干旱区城市土地覆盖分类[J].国土资源遥感,2016,28(1):43-49. doi: 10.6046/gtzyyg.2016.01.07
	[ Tian S F, Zhang X F.Classification of urban land cover in arid region using sky-painted data of random forest method[J]. Remote Sensing for Land and Resources, 2016,28(1):43-49. ] doi: 10.6046/gtzyyg.2016.01.07
[9]	马玥,姜琦刚,孟治国,等.基于随机森林算法的农耕区土地利用分类研究[J].农业机械学报,2016,47(1):297-303. doi: 10.6041/j.issn.1000-1298.2016.01.040
	[ Ma Y, Jiang Q G, Meng Z G, et al.Land use classification in agricultural land based on random forest algorithm[J]. Journal of Agricultural Machinery, 2016,47(1):297-303. ] doi: 10.6041/j.issn.1000-1298.2016.01.040
[10]	韩燕,王玲,罗冲.随机森林算法在干旱区土地利用遥感分类中的应用研究[J].石河大学学报(自科版),2017,35(1):95-101. doi: 10.13880/j.cnki.65-1174/n.2017.01.016
	[ Han Y, Wang L, Luo C.Application of random forest algorithm in remote sensing classification of land use in arid areas[J]. Journal of Shihe University (Natural Science), 2017,35(1):95-101. ] doi: 10.13880/j.cnki.65-1174/n.2017.01.016
[11]	谷晓天. 基于机器学习的湟水流域土地利用/土地覆被分类研究[D].青海:青海师范大学,2018.
	[ Gu X T.Research on land use/land cover classification based on machine learning in huangshui river basin[D]. Qinghai. Qinghai normal university, 2018. ]
[12]	曾永年,靳文凭,何丽丽,等.青海高原东部农业区土地利用遥感分类制图[J].农业工程学报,2012,28(16):225-231. doi: 10.3969/j.issn.1002-6819.2012.16.035
	[ Zeng Y N, Jin W P, He L L, et al.Remote sensing classification mapping of land use in the eastern qinghai plateau agricultural region[J]. Journal of Agricultural Engineering, 2012,28(16):225-231. ] doi: 10.3969/j.issn.1002-6819.2012.16.035
[13]	熊有平. 湟水流域川水区、浅山区、脑山区和石山林区划分及特点[J].水利科技与经济,2012,18(2):14-15. doi: 10.3969/j.issn.1006-7175.2012.02.007
	[ Xiong Y P.Division and characteristics of valley plains, loess hills, middle and high mountains and stone mountain forests in huangshui river basin[J]. Water Science and Technology and Economy, 2012,18(2):14-15. ] doi: 10.3969/j.issn.1006-7175.2012.02.007
[14]	李金山. 复杂地形区土地利用/土地覆被分类研究——以湟水流域为例[D].青海:青海师范大学,2013:1-82.
	[ Li J S.A study on land use/land cover changes in complex topographical areas: A case study of the Qinshui river basin[D]. Qinghai: Qinghai Normal University, 2013:1-82. ]
[15]	贾伟,高小红,杨灵玉,等.面向对象方法的复杂地形区地表覆盖信息提取[J].兰州大学学报(自然科学版),2018,54(04):486-493.
	[ Jia W, Gao X H,Yang L Y, et al.Surface cover information extraction of complex topographic areas by object-oriented method[J]. Journal of Lanzhou University (Natural Science Edition), 2008,54(04):486-493. ]
[16]	李欣海. 随机森林模型在分类与回归分析中的应用[J].应用昆虫学报,2013,50(4):1190-1197. doi: 10.7679/j.issn.2095-1353.2013.163
	[ Li xinhai. Application of stochastic forest model in classification and regression analysis[J]. Chinese Journal of Applied Entomology, 2013,50(4):1190-1197. ] doi: 10.7679/j.issn.2095-1353.2013.163
[17]	董师师,黄哲学.随机森林理论浅析[J].集成技术,2013,2(1):1-7.
	[ Dong S S, Huang Z X.Analysis of random forest theory[J]. Integrated Ttechnology, 2013,2(1):1-7. ]
[18]	Menze B H, Kelm B M, Splitthoff D N, et al.On oblique random forests[C]// Machine Learning and Knowledge Discovery in Databases, Berlin: Springer, 2011:453-469.
[19]	Singh V P, Srivastava S, Srivastava R.Effective mammogram classification based on center symmetric-LBP features in wavelet domain using random forests[J]. Technology and Health Care, 2016,25(4):709-727. doi: 10.3233/THC-170851 pmid: 28582938
[20]	王栋,岳彩荣,田传召,等.基于随机森林的大姚县TM遥感影像分类研究[J].林业调查规划,2014,39(2):1-5. doi: 10.3969/j.issn.1671-3168.2014.02.001
	[ Wang D, Yue C R, Tian C Z, et al.Remote sensing image classification of dayao county based on random forest[J]. Journal of Forestry Investigation and Planning, 2014,39(2):1-5. ] doi: 10.3969/j.issn.1671-3168.2014.02.001
[21]	张大川,刘小平,姚尧,等.基于随机森林CA的东莞市多类土地利用变化模拟[J].地理与地理信息科学,2016,32(5):29-36. doi: 10.3969/j.issn.1672-0504.2016.05.005
	[ Zhang D C, Liu X P, Yao Y, et al.Simulation of multiple land use changes in dongguan city based on stochastic forest CA[J]. Geoscience and Geographic Information Science, 2016,32(5):29-36. ] doi: 10.3969/j.issn.1672-0504.2016.05.005
[22]	白琳,徐永明,何苗,等.基于随机森林算法的近地表气温遥感反演研究[J].地球信息科学学报,2017,19(3):390-397. doi: 10.3724/SP.J.1047.2017.00390
	[ Bai L, Xu Y M, He M, et al.Research on remote sensing inversion of near-surface temperature based on stochastic forest algorithm[J]. Journal of Geoinformation Science, 2017,19(3):390-397. ] doi: 10.3724/SP.J.1047.2017.00390
[23]	雷震. 随机森林及其在遥感影像处理中应用研究[D].上海:上海交通大学,2012:1-130.
	[ Lei Z.Application of random forest and its application in remote sensing image processing[D]. Shanghai: Shanghai Jiao Tong University, 2012:1-130. ]
[24]	Hernandez I E R, Shi W. A Random Forests classification method for urban land-use mapping integrating spatial metrics and texture analysis[J]. International Journal of Remote Sensing, 2018,39(4):1175-1198. doi: 10.1080/01431161.2017.1395968

影像编号	接收卫星	传感器类型	影像获取时间	分辨率/m	波段数
LC81330342016211LGN00	Landsat8	Operational Land Image （OLI）	2016-07-29	30	2-7波段
LC81310352016213LGN00			2016-07-31
LC81320342015233LGN00			2015-08-21
LC81320352014198LGN00			2014-07-17

特征参数	特征参数特点及意义
Landsat8OLI 多光谱2~7 波段	2-蓝波段	对水体穿透强,可获得更多水下信息;能够区分土壤和植被、分析土地利用结构变化。
	3-绿波段	主要观测植被在绿波段中的反射峰值,这一波段位于叶绿素的两个吸收带之间,利用这一波段增强鉴别植被的能力
	4-红波段	该波段为叶绿素的主要吸收波段,能增强植被覆盖与无植被覆盖之间的反差,亦能增强同类植被的反差,反映不同植物叶绿素吸收,植物健康状况,用于区分植物种类与植物覆盖率
	5-近红外波段	对植被类别差异最敏感,可以区别植被类型;由于处于水体强吸收区,因此呈现的水体轮廓清晰,便于与其他地物的区分
	7-短波红外2	反映植物和土壤水分含量敏感,可以区别雪和云
	8-短波红外2	可用于区分主要岩石类型;处于水的强吸收带,在影像上该波段的水体呈黑色;对植物水分敏感
指数信息	NDVI	归一化差值植被指数,也称为生物量指标变化,可使植被从水和土中分离出来,是植被生长及植被覆盖度最佳指示因子
	NDBI	归一化建筑物指数,利用了不透水面的中红外波段反射率高于近红外反射率的规律,该指数有助于城乡、工矿与居住建设用地的提取
	MNDWI	改进归一化差值水体指数,用遥感影像特定的波段进行改进的归一化差值处理,以突显影像中的水体信息
地形信息	DEM	数字高程模型,可以提取坡度与坡向,影响研究区土地利用/土地覆被类型分布格局
纹理信息	纹理特征	纹理信息以灰度共生矩阵为主,是一种通过研究灰度的空间相关特性来描述纹理的常用方法;可以反映图像灰度分布均匀程度和纹理粗细程度

地理分区/m	特征信息	特征参数
脑山区(>3200 )	光谱信息地形信息	B,G,R,NIR,SWIR1,SWIR2,NDVI 高程,坡度,坡向
浅山区(2600~3200 )	光谱信息纹理信息地形信息	B,G,R,NIR,SWIR1,SWIR2,NDVI,NDBI,MNDWI 均值,方差,同质性,对比度,非相似性,熵,二阶矩,相关性高程,坡度,坡向
川水区(<2600 )	光谱信息纹理信息地形信息	B,G,R,NIR,SWIR1,SWIR2,NDVI,NDBI,MNDWI,PCA1,PCA2 均值,方差,同质性,对比度,非相似性,熵,二阶矩,相关性,高程,坡度,坡向

土地利用类型	6MS		6MS+NDVI		6MS+DEM		6MS+DEM+NDVI
土地利用类型	制图精度	用户精度	制图精度	用户精度	制图精度	用户精度	制图精度	用户精度
林地	84.62	97.06	84.62	97.06	85.71	93.33	85.71	95.45
高草地	94.07	86.72	95.76	86.92	92.65	88.73	94.12	88.89
中草地	85.71	87.50	85.71	87.50	96.30	83.87	96.30	83.87
水域	92.73	94.44	90.91	94.34	92.31	94.70	92.31	97.70
未利用土地	93.48	91.49	93.48	91.49	88.46	92.00	88.46	92.00
总体精度	90.46		90.75		90.82		91.33
Kappa系数	0.876		0.879		0.880		0.886

土地利用类型	6MS+SL+AS		6MS+SL+AS+NDVI		6MS+SL+AS +OIF+TXT
土地利用类型	制图精度	用户精度	制图精度	用户精度	制图精度	用户精度
水浇地	90.91	95.00	90.91	95.24	90.91	95.24
旱地	94.74	88.34	94.74	91.14	94.74	90.57
林地	91.78	97.10	92.47	96.43	92.47	96.43
高草地	88.46	92.00	88.46	90.20	94.23	96.08
中草地	94.83	85.94	97.10	84.06	96.55	90.32
水域	88.89	94.12	88.89	94.12	91.67	80.49
城乡工矿居民用地	83.93	85.45	80.36	90.00	78.57	89.80
总体精度	91.54		91.91		92.09
Kappa系数	0.895		0.900		0.902