Constructing Representative Calibration Dataset Based on Spectral Transformation and Kennard-Stone Algorithm for VNIR Modeling of Soil Total Nitrogen in Paddy Soil
1. School of Resource and Environment Science，Wuhan University，Wuhan 430079，China
2. Suzhou Institute of Wuhan University，Suzhou 215123，China
3. Collaborative Innovation Center of Geospatial Technology，Wuhan University，Wuhan 430079，China
4. Key Laboratory of Geographic Information System of Ministry of Education，Wuhan University，Wuhan 430079，China
5. Institute of Agricultural Remote Sensing and Information Technology Application，Zhejiang University，Hangzhou 310058，China
6. Institute of Geographic Sciences and Natural Resources Research，Chinese Academy of Sciences，Beijing 100101，China
Abstract：Visible and near infrared (VNIR) has been widely used to estimate various soil properties. The construction of calibration dataset during the VNIR estimation of soil properties not only influences the representative of the calibration dataset, but also on prediction accuracy of models. Current strategy for calibration dataset construction combines VNIR spectra with Kennard-Stone (KS) algorithm. However, this strategy neglects the fact that soil reflectance spectra are a comprehensive reflection of soil properties rather than a specific component. As a result, the constructed dataset from this approach may be not good enough to represent the relationships between soil spectra and the target soil component. Given that different spectral transformations could be helpful to highlight the spectral characteristics of target component, they might be also useful in the selection of samples for model calibration. The aim of the study is to explore the potentials of the combined approach of different spectral transformations and KS algorithm in the construction of calibration dataset and the VNIR estimation of soil total nitrogen (TN). It is hypothesized that the proposed approach could help to better select samples for calibration, which are more representative comparing with those selected by KS algorithm using sample reflectance spectra. A total of 100 samples have been collected from paddy soil in Jianghan Plain of Hubei Province. Five transformation methods, namely the first derivative (FD), Savitzky-Golay (SG), Standard Normal Variate (SNV), Multiple Scatter Correction (MSC) and Harr Wavelet transform have been employed for spectral transformations. Thereafter, KS algorithm is used to construct representative calibration datasets based on the differently transformed spectra. Partial least square regression (PLSR) is then used for model calibration. Whether the different spectral transformation methods can improve the representative of the calibration dataset constructed by KS algorithm is examined. The results illustrate that different spectral transformations can exert different effects on the construction of calibration dataset. The SG and Wavelet spectral transformations do not make a difference for the calibration dataset constructed by KS algorithm using reflectance spectra, with ratio of performance to standard deviate (RPD) of 1.41 and 1.27 respectively. The spectral transformations of FD, SNV or MSC do improve the calibration dataset constructed by KS algorithm, with the RPDs improve from 0.95, 1.48 and 1.42 to 1.13, 1.78 and 2.20 respectively. The study indicates that such spectral transformations as SNV and MSC could change the way that KS algorithm constructs calibration dataset and improve its representative relationships between soil spectra and soil TN. Therefore, we conclude that the proposed strategy for calibration dataset construction holds great pontentials to improve the model prediction capability in the VNIR estimation of soil TN.
Key words：Spectral transformation; Kennard-Stone algorithm; Calibration dataset construction; Model accuracy; Soil total nitrogen
