Application of Support Vector Machine Regression and Partial Least-Square Regression Models for Predicting Sugarcane Leaf Nutrients Content Using Near Infra-Red Spectroscopy

Abstract

ABSTRACT Near infra-red spectroscopy (NIRS) has been suggested as a rapid, cost-effective, and accurate diagnostic tool for leaf nutrient analysis that could replace more traditional laboratory diagnostics. To ease operational workflows, there would advantage in estimating nutrients using a single method, namely NIRS. This study evaluated the potential of NIRS as a diagnostic method for the measurement of key macro and micronutrients in sugarcane leaf samples. Three hundred and fifty-one sugarcane leaf samples used in quality control reference analysis in Fertiliser Advisory Service (FAS) were used for model calibration. About 35% of the samples were from growers within South Africa, while the remainder were from estates across southern and eastern Africa. Dried and milled leaf material was scanned on a Bruker MPA-NIRS instrument, and spectral pre-processing was performed. Support vector machine regression (SVMR) and partial least squares regression (PLSR) were used for calibrating the estimation models with the test validation(Tval) procedure. The results showed both the PLSR and SVMR model resulted in the best calibration for (K, Ca, Mg)(R2 >87% and the ratio of performance to inter-quartile distance(RPIQ) > 2.1). With a test validation dataset, the SVMR yielded relatively high accuracy(R2 >85% and the RPIQ > 2.5, while PLSR yielded poor accuracy. The SVMR also outperforms Zn(R2 = 0.86; and RPIQ = 2.5) for model calibration. NIR Spectrometry combined with the SVMR technique resulted in a practical option to accurately measure leaf nutrient concentrations and evaluate sugarcane mineral contents accurately. These results indicate that NIRS can replace the current analytical methods used in FAS.