植物生态学报 ›› 2017, Vol. 41 ›› Issue (12): 1273-1288.DOI: 10.17521/cjpe.2017.0231

所属专题: 遥感生态学

• 研究论文 • 上一篇    下一篇

植冠下土壤类型差异对遥感估算冬小麦叶面积指数的影响

高林1,2,*, 王晓菲1,2,3, 顾行发4, 田庆久1,2,3,**(), 焦俊男1,2, 王培燕1,2, 李丹4   

  1. 1南京大学国际地球系统科学研究所, 南京 210023
    2南京大学江苏省地理信息技术重点实验室, 南京 210023
    3南京大学中国南海研究协同创新中心, 南京 210023
    4中国科学院遥感与数字地球研究所, 北京 100101
  • 出版日期:2017-12-10 发布日期:2018-02-23
  • 通讯作者: 田庆久
  • 基金资助:
    国家自然科学基金(41771370)和国家重点研发计划(2017YFD0600903)

Exploring the influence of soil types underneath the canopy in winter wheat leaf area index remote estimating

GAO Lin1,2,*, WANG Xiao-Fei1,2,3, GU Xing-Fa4, TIAN Qing-Jiu1,2,3,**(), JIAO Jun-Nan1,2, WANG Pei-Yan1,2, LI Dan4   

  1. 1International Institute for Earth System Science, Nanjing University, Nanjing 210023, China

    2Jiangsu Provincial Key Laboratory of Geographic Information Science and Technology, Nanjing University, Nanjing 210023, China

    3Collaborative Innovation Center of South China Sea Studies, Nanjing University, Nanjing 210023, China
    and
    4Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing 100101, China
  • Online:2017-12-10 Published:2018-02-23
  • Contact: TIAN Qing-Jiu

摘要:

遥感是从田块到区域乃至全球范围无损探测叶面积指数(LAI)的有效方法。土壤背景是LAI遥感研究的重要制约因素之一, 而土壤类型是组成土壤背景的主要部分, 对植被冠层-土壤的光学性质有重要影响, 但目前植冠下土壤类型背景对遥感LAI估算的影响尚不明确。该文通过分析归一化差异植被指数、修正型土壤调节植被指数、修正的叶绿素吸收比率指数、红边拐点、红边振幅、红边面积、红边对数指数和归一化差异光谱指数在不同土壤类型下对LAI的敏感性, 挖掘最不敏感的光谱参数; 通过比较两种回归模型(偏最小二乘回归和随机森林回归)在单一土壤类型和多种土壤类型区对LAI的预测精度, 探究将单一土壤类型下发展的LAI估算模型应用到复杂土壤类型地区时可能出现的问题。结果表明: (1)虽然8种光谱指数对LAI的敏感性因土壤类型不同而差异明显, 但红边拐点受植冠下土壤类型影响最小; “lambda-by-lambda”波段优选算法不仅可以提供对LAI最敏感的光谱区间, 而且可在一定程度上为抵抗植冠下土壤类型差异影响的光谱指数构建提供可行思路; (2)回归模型的LAI预测精度因是否考虑土壤类型而不同, 但在小区域尤其是田块尺度研究时, 对变量的解释能力是选择模型的第一考虑, 而偏最小二乘回归在此方面优于随机森林回归; 在未知地表先验知识的前提下, 随机森林回归对大区域LAI估算比偏最小二乘回归适合, 但地表先验知识的获取对LAI遥感估算仍然十分必要。

关键词: 遥感, 土壤类型, 冬小麦叶面积指数, 光谱指数, 偏最小二乘回归, 随机森林回归

Abstract: Aims Remote sensing is an effective and nondestructive way to retrieve leaf area index (LAI) from plot, regional and global range. Soil background is one of the confounding factors limiting remotely estimating LAI. And soil type contains a large proportion of soil background information, which can influence the optical properties of vegetation canopy and soil. However, our knowledge on the effects stemmed from soil types underneath the canopy on LAI remote estimating have been in shortage. Thus, this study aims to explore the influences of soil types underneath the canopy on winter wheat LAI remote estimating. Methods We analyzed the sensitivity variation of eight spectral indices, named normalized difference vegetation index (NDVI), modified soil-adjusted vegetation index (MSAVI), modified chlorophyll absorption ratio index 2 (MCARI2), red edge inflection point (REIP), red edge amplitude (Dr), red edge area (SDr), red edge symmetry (RES), normalized difference spectral index (NDSI), to LAI in different soil types, and then we identified some spectral intervals or parameters that were insensitive to soil type variations underneath the canopy. We also compared the accuracy of two commonly used regression models, partial least squares regression (PLSR) and random forest regression (RFR), in estimating LAI for different soil types. We also explored the problems arising from applying the regression model developed in single soil type area to complex soil types area in retrieving LAI. Important findings This paper demonstrates the effects of soil types underneath the canopy on LAI retrieving. 1) The sensitivity of spectral indices to LAI is significantly different due to the soil type variation, but REIP has the least effects from soil type variation among the eight spectral indices. Meanwhile, the band selection algorithm of lambda-by-lambda not only chooses the most sensitive spectral interval for LAI, but also provides a feasible way to construct the spectral index that exhibits strong resistances to the effects of soil types underneath the canopy. 2) The accuracy of LAI estimation by regression models differs under soil type considered or not. So we suggest that in small scale researches, especially in a field scale, the ability of regression models in explaining variables is the priority consideration, while the PLSR is superior to RFR in this respect. Under the premise of unknown priori knowledge of land surfaces, the RFR is more suitable for retrieving LAI than PLSR, but land surface priori knowledge is still necessary. These findings provide the theoretical basis and methods for developing remotely sensing estimating LAI models adapted to various land surfaces. Further analysis is needed in applying the findings in more crop types, cultivars and growth stages.

Key words: remote sensing, soil type, winter wheat leaf area index, spectral indices, partial least squares regression, random forest regression