地球信息科学学报 ›› 2016, Vol. 18 ›› Issue (8): 1123-1132.doi: 10.3724/SP.J.1047.2016.01123

• 遥感科学与应用技术 • 上一篇    下一篇

变换光谱数据对土壤氮素PLSR模型的影响研究

乔星星(), 冯美臣*(), 杨武德, 孙慧, 郭小丽, 史超超   

  1. 山西农业大学旱作农业工程研究所,太谷 030801
  • 收稿日期:2015-12-07 修回日期:2016-02-17 出版日期:2016-08-10 发布日期:2016-08-10
  • 通讯作者: 冯美臣 E-mail:qxx1702@163.com;fmc101@163.com
  • 作者简介:

    作者简介:乔星星(1989-),女,山西长子人,硕士生,研究方向为作物生态和信息技术。E-mail:qxx1702@163.com

  • 基金资助:
    国家自然科学基金项目(31371572、31201168);山西省科学技术发展计划项目(201603D221037-3);山西省归国人员重点资助项目(2014-重点4);山西省科技攻关项目(20110311038);山西省青年基金项目(2012021023-5)

Effect of Spectral Transformation Processes on the PLSR Models of Soil Nitrogen

QIAO Xingxing(), FENG Meichen*(), YANG Wude, SUN Hui, GUO Xiaoli, SHI Chaochao   

  1. Institute of Dry Farming Engineering, Shanxi Agricultural University, Taigu 030801, China
  • Received:2015-12-07 Revised:2016-02-17 Online:2016-08-10 Published:2016-08-10
  • Contact: FENG Meichen E-mail:qxx1702@163.com;fmc101@163.com

摘要:

光谱数据变换对消除背景、噪音影响以及提取光谱特征有重要的作用,是光谱数据分析过程中的必要步骤。为了研究光谱变换处理对土壤氮素PLSR模型的影响精度,并选择最佳光谱变换处理方法,本文对原始光谱数据进行了15种典型光谱变换,通过比较不同变换光谱与土壤氮素的相关性,实现土壤氮素的PLSR精确诊断,并综合评定最佳光谱数据变换方法。结果表明,涉及微分处理后的光谱变换,尤其是先进行开方(T8、T11)、对数(T6、T12)等变换后再进行微分处理,可提高其与土壤氮素的相关性。在引入较少因子变量个数的条件下,该方法使因变量解释量达到了98%。综合考虑模型的校正、验证效果及模型复杂度(模型最佳因子变量个数),可得出光谱平方根的一阶微分变换处理(T8)为最佳的土壤光谱变换算法。该条件下的土壤氮素的校正模型表现为R2=0.985、RMSEC=0.000132、Fn=6,验证模型的表现为R2=0.9853、RMSEV=0.000162,结果表明基于T8的光谱数据变换可实现本试验条件下土壤氮素的光谱估算。另外,可以考虑将原始光谱的一阶微分(T9)、对数和对数倒数的一阶微分(T6、T7)以及平方根和对数的二阶微分(T11、T12)作为光谱数据变换方法。本文研究结果可为土壤氮素估算和光谱数据预处理提供技术参考。

关键词: 光谱变换, 土壤氮素, 诊断, 精度, 偏最小二乘法

Abstract:

Spectral transformation is an essential pre-treatment technique as it can eliminate the effects of background and noise, and it plays a vital role in extracting spectral features′ information and constructing the optimal model. In order to explore the effects of different spectral transformation methods on the accuracy of the PLSR model in monitoring soil nitrogen and determining the optimal spectral transformation, the raw spectrum was transformed with respect to fifteen transformation algorithms and the correlations between each pair of transformed spectrum and soil nitrogen were analyzed. Furthermore, the performances of the PLSR models in monitoring the soil nitrogen based on different transformed spectra were evaluated. The results showed that, for cases involving the first or second-order differential reprocessing transformations, the correlation coefficient between the soil nitrogen and the relevant transformed spectrum increased more significantly than with the raw spectrum, especially when applying the transformation algorithms of square root (T8 and T11) and logarithm (T6 and T12) firstly. Also, fewer optimal factors for these pre-treatments were needed and selected to achieve the threshold of 98% in explaining the dependent variable. Moreover, the first-order differential reprocessing of the square root of raw spectrum (T8) had a higher accuracy (R2=0.985022, RMSEC=0.000132; R2=0.9853, RMSEV=0.000162, Fn=6) for the calibrated model and the validated model respectively, after the comprehensive evaluation of the predicting performance and the complexity of different models. Finally, the first-order differential reprocessing of the square root of raw spectrum (T8) was determined as the recommended transformation method to evaluate the soil nitrogen. In addition, the first-order and second-order differential of the logarithm of raw spectrum (T6 and T12), the first-order differential of the logarithmic reciprocal of raw spectrum (T7), the first-order differential of raw spectrum (T9), as well as the second-order differential of the square root of raw spectrum (T11) could also be considered and chosen as alternatives. The study would provide some theoretical techniques and references to the evaluation of soil nitrogen and spectrum processing.

Key words: spectral transformation, soil nitrogen, diagnosis, accuracy, PLSR