结合光谱变换和Kennard-Stone算法的水稻土全氮光谱估算模型校正集构建策略研究

发布时间：2018-10-29 15:23

【摘要】：土壤组分光谱估算过程中校正样本集的构建会影响模型的预测精度。当前结合反射光谱和Kennard-Stone(KS)算法的校正样本集构建策略忽视了土壤反射光谱是土壤属性的综合反映,构建的样本集通常无法很好地代表目标土壤组分的变异。光谱变换方法可以突出目标组分的光谱特征,为此,本文以湖北省江汉平原滨湖地区水稻土为研究对象,结合包括一阶微分(FD)、Savitzky-Golay(SG)、Haar小波变换、标准正态变量变换(SNV)和多元散射校正(MSC)在内的光谱变换方法和KS算法进行校正样本集建构,通过对比不同样本集构建策略对使用偏最小二乘回归(PLSR)建立的土壤全氮含量光谱估算模型预测精度的影响,研究光谱变换是否有助于提高基于KS算法构建的校正样本集的代表性。结果表明:不同光谱变换会影响校正样本集的构建。反射光谱经过SG或Haar小波变换后,再使用KS算法构建校正样本集与直接基于反射光谱使用KS算法构建的校正样本集相同,建立的估算模型精度不变,相对分析误差(RPD)分别为1.41和1.27。结合FD,SNV或MSC变换和KS算法构建的校正集与基于反射光谱使用KS算法构建的校正集不同,建立的估算模型RPD分别从0.95,1.48和1.42提高到1.13、1.78和2.20。研究表明SNV和MSC等光谱变换方法可以提高基于KS算法构建的校正样本集的代表性,并可有效提高模型预测精度。
[Abstract]:In the process of estimating soil component spectrum, the construction of calibration sample set will affect the prediction accuracy of the model. The current strategy of constructing the corrected sample set by combining the reflectance spectrum with the Kennard-Stone (KS) algorithm ignores that the soil reflectance spectrum is a comprehensive reflection of the soil attributes, and the constructed sample set can not well represent the variation of the target soil component. The spectral transformation method can highlight the spectral characteristics of the target components. In this paper, the paddy soil in Jianghan Plain of Hubei Province is taken as the research object, which includes the first-order differential (FD), Savitzky-Golay (SG), Haar wavelet transform. The standard normal variable transform (SNV) and multivariate scattering correction (MSC) spectral transformation method and KS algorithm are used to construct the calibration sample set. The effects of different sample set construction strategies on the prediction accuracy of the spectral estimation model of soil total nitrogen content established by partial least squares regression (PLSR) were compared. To study whether spectral transformation can improve the representativeness of calibration sample set based on KS algorithm. The results show that different spectral transformation will affect the construction of calibration sample set. After SG or Haar wavelet transform, the corrected sample set constructed by KS algorithm is the same as that constructed by KS algorithm directly based on reflection spectrum, and the accuracy of the estimation model is invariant. The relative analysis error (RPD) is 1. 41 and 1. 27, respectively. The correction set constructed by combining FD,SNV or MSC transform with KS algorithm is different from that constructed by KS algorithm based on reflectance spectrum. The estimated model RPD is improved from 0.95g 1.48 and 1.42 to 1.131.78 and 2.20 respectively. The results show that the spectral transformation methods such as SNV and MSC can improve the representativeness of the corrected sample set based on the KS algorithm, and can effectively improve the prediction accuracy of the model.
【作者单位】：武汉大学资源与环境科学学院;武汉大学苏州研究院;武汉大学地球空间信息技术协同创新中心;武汉大学教育部地理信息系统重点实验室;浙江大学农业遥感与信息技术应用研究所;中国科学院地理科学与资源研究所;
【基金】：国家自然科学基金项目(41501444) 苏州市应用基础农业项目(SYN201422) 国家基础人才培养计划《武汉大学地理科学理科基地》科研能力训练项目(J1103409)资助
【分类号】：O657.3;S153

【相似文献】