Original scientific paper
https://doi.org/10.5513/JCEA01/20.1.2158
Hyperspectral sensing of soil pH, total carbon and total nitrogen content based on linear and non-linear calibration methods
Ivana Šestak
orcid.org/0000-0002-5802-8669
; University of Zagreb, Faculty of Agriculture, Department of General Agronomy, Svetošimunska cesta 25, 10000 Zagreb, Croatia
Lea Mihaljevski Boltek
; University of Zagreb, Faculty of Agriculture, Undergraduate study program: Agroecology, Svetošimunska cesta 25, 10000 Zagreb, Croati
Milan Mesić
; University of Zagreb, Faculty of Agriculture, Department of General Agronomy, Svetošimunska cesta 25, 10000 Zagreb, Croatia
Željka Zgorelec
; University of Zagreb, Faculty of Agriculture, Department of General Agronomy, Svetošimunska cesta 25, 10000 Zagreb, Croatia
Aleksandra Perčin
; University of Zagreb, Faculty of Agriculture, Department of General Agronomy, Svetošimunska cesta 25, 10000 Zagreb, Croatia
Abstract
Soil properties can be estimated non-destructively by visible and near infrared (VNIR) reflectance spectroscopy. However, results of calibration models differ in dependence of measurement precision, spectral range, variability of soil properties and calibration methods used for prediction. The objective of research was to estimate the ability of hyperspectral VNIR sensing for field-scale prediction of soil pH, total carbon (TC, %) and total nitrogen (TN, %) content in arable Stagnosols. Total of 200 soil samples taken from field experiment (soil depth: 30 cm; sampling grid: 15x15 m; 2016) was scanned in laboratory using portable spectroradiometer (FieldSpec®3, 350-1,050 nm). Partial least squares regression (PLSR) and artificial neural networks (ANN) were used to build prediction models of selected soil properties based on VNIR spectra (P<0.05). Very strong to complete correlation and low root mean squared error was obtained between predicted and measured values for the calibration and validation dataset, and both prediction methods (PLSR validation: TC, %: R2=0.85, RMSE=0.163; TN, %: R2=0.76, RMSE=0.018; soil pH: R2=0.69, RMSE=0.55; ANN validation: TC, %: R2=0.88, RMSE=0.108; TN, %: R2=0.86, RMSE=0.012; soil pH: R2=0.74, RMSE=0.42). ANN models were more efficient in capturing the complex link between selected soil properties and soil reflectance spectra than PLSR. Calibrations defined in this research should help to support sitespecific soil survey as addition to standard laboratory analysis.
Keywords
neural networks; partial least squares regression; reflectance spectroscopy; soil carbon and nitrogen content; soil pH
Hrčak ID:
218137
URI
Publication date:
19.3.2019.
Visits: 1.833 *