On-line Vis-Nir sensor determination of soil variations of sodium, potassium and magnesium

Y Tekin,A M Mouazen,Y Ulusoy,Z Tumsavaş

doi:10.1088/1755-1315/41/1/012011

Abstract

Among proximal measurement methods, visible and near infrared (Vis-Nir) spectroscopy probably has the greatest potential for determining the physico-chemical properties of different natural resources, including soils. This study was conducted to determine the sodium, potassium and magnesium variations in a 10. Ha field located in Karacabey district (Bursa Province, Turkey) using an on-line Vis-Nir sensor. A total of 92 soil samples were collected from the field. The performance and accuracy of the Na, K and Mg calibration models was evaluated in cross-validation and independent validation. Three categories of maps were developed: 1) reference laboratory analyses maps based on 92 points 2) Full-data point maps based on all 6486 on-line points Vis-Nir predicted in 2013 and 3) full- data point maps based on all 2496 on-line points Vis-Nir predicted in 2015. Results showed that the prediction performance in the validation set was successful, with average R2 values of 0.82 for Na, 0.70 for K, and 0.79 for Mg, average root mean square error of prediction (RMSEP) values of 0.02% (Na), 0.20% (K), and 1.32% (Mg) and average residual prediction deviation (RPD) values of 2.13 (Na), 0.97 (K), and 2.20 (Mg). On-line field measurement was also proven to be successful with validation results showing average R2 values of 0.78 (Na), 0.64 (K), and 0.60 (Mg), average RMSEP values of 0.04% (Na), 0.13% (K), and 2.19% (Mg) and average RPD values of 1.57 (Na) 1.68 (K) and 1.56 (Mg). Based on 3297 points, maps of Na, K and Mg were produced after N, P, K and organic fertilizer applications, and these maps were then compared to the corresponding maps from the previous year. The comparison showed a variation in soil properties that was attributed to the variable rate of fertilization implemented in the preceding year.

Highlights

In the last two decades, the number of studies evaluating other visible and near infrared (Vis-Nir) spectroscopy applications in soil science and agronomy has increased rapidly, with a primary focus on measuring various basic properties of soils, such as the organic matter content, clay content and, more recently, chemical properties [1]
The complex relationship between spectral signatures and soil properties can be better modeled via multivariate regression methods, which have an advantage over the simple bivariate relationships, e.g., those based on peak intensity measurements [8]
One of the advantages of Partial least squares (PLS) regression compared to other chemometric methods, such as principal component regression analysis, is the possibility of interpreting the first few latent variables, because these show the correlations between the property values and the spectral features [11]

Summary

Introduction

In the last two decades, the number of studies evaluating other Vis-Nir spectroscopy applications in soil science and agronomy has increased rapidly, with a primary focus on measuring various basic properties of soils, such as the organic matter content, clay content and, more recently, chemical properties [1]. Vis-Nir spectroscopy is one of the main methods that have been explored This can be attributed to the fact that, by using suitable chemometric methods, large sets of spectral information can be extracted from the Vis-Nir spectra of soils. The complex relationship between spectral signatures and soil properties can be better modeled via multivariate regression methods, which have an advantage over the simple bivariate relationships, e.g., those based on peak intensity measurements [8]. One of the advantages of PLS regression compared to other chemometric methods, such as principal component regression analysis, is the possibility of interpreting the first few latent variables, because these show the correlations between the property values and the spectral features [11]. The calibration samples should cover the variability expected in the full sample set, and the future unknown data and the validation (test) set must be independent of the calibration set in order to avoid an optimistic assessment of predictive performance [8, 12, 13]

Objectives

Methods

Results

Conclusion