Evaluation of MLSR and PLSR for estimating soil element contents using visible/near-infrared spectroscopy in apple orchards on the Jiaodong peninsula

• Published in: Catena (Giessen) 2016-02, Vol.137, p.340-349
• Main Authors: Yu, Xiang, Liu, Qing, Wang, Yebao, Liu, Xiangyang, Liu, Xin
• Format: Article
• Language: English
• Subjects: Estimation; Multiple linear stepwise regression; Partial least square regression; Soil element contents; Visible/near-infrared spectroscopy
• ISSN: 0341-8162; 1872-6887
• DOI: 10.1016/j.catena.2015.09.024

Conventional methods for soil element content determination based on laboratory analyses are costly and time-consuming. A soil reflectance spectrum is an alternative approach for soil element content estimation with the advantage of being rapid, non-destructive, and cost effective. Visible/near-infrared spectra (350nm to 2500nm) were measured from 105 soil samples originating from 30 apple orchards on the Jiaodong peninsula. The Savitzky–Golay (FD-SG) technique for spectral data was implemented to reduce the signal noise. Logarithm of the reciprocal of reflectance (logR−1) and the first derivative transformation (DR) were used to accentuate the features and to prepare the data for use in quantitative estimation models. The SI (sum index), DI (difference index), PI (product index), RI (ratio index), and NDI (normalized difference index) were calculated to extract sensitive waveband combinations that are significantly related to soil element contents. Soil element contents were retrieved based on sensitive waveband combinations by multiple linear stepwise regression (MLSR) and partial least square (PLSR) models. The results showed that DR performed better than logR−1 in eliminating the interfering factors of soil particle size and spectral noise. The MLSR and PLSR calibration models based on PI performed better than those based on SI or DI did. The MLSR performed better than PLSR in estimating soil elemental content. The contents of total nitrogen (TN), arsenic (As), and mercury (Hg) could be estimated well using MLSR and PLSR calibration models developed with PI. The MLSR calibration model developed with PI performed well in estimating available potassium (A-K) content. However, the contents of available phosphorus (A-P), ammonium nitrogen (NH4+-N), nitric nitrogen (NO3--N), and soil organic matter (SOM) could not be estimated using MLSR or PLSR calibration models. These outcomes will provide the theoretical basis and technical support for estimations of soil element content using visible/near-infrared spectra. Although they were shown to be useful in apple orchards of the Jiaodong peninsula, these models and methods should be further tested in soil samples from other regions and countries to prove their validity. •Laboratory-based Vis–NIR spectra of soil samples were gained.•The contents of eight soil elements were measured.•Five spectral indices were calculated based on two spectral formats.•Performances of spectral indices in estimating soil element cont