Case Studies in Chemical and Environmental Engineering (Dec 2023)
Application of LW-NIR for rapid prediction of primary nutrients cropland by chemometrics: Comparison among preprocessing and machine learning algorithm approaches
Abstract
Rapid analytical approaches are necessary for accurate decision-making on the suitability of soil nutrients and types of plants to be cultivated on its cropland. Therefore, in the present study, the potential of long-wave near-infrared spectroscopy (LW-NIRs) as a rapid and nondestructive tool to predict primary nutrients attributes of soil agriculture including N, P, K was investigated. Five different machine learning algorithms, namely partial least squares regression (PLSR), k-nearest neighbor (kNN), decision tree (DT), adaboosting (AB), and Gaussian process regression (GPR) were used and compared to predict the N, P, K content of cropland from Aceh Province. These prediction models were established based on NIR spectra acquired in the wavelength range from 1100 to 2000 nm. Four different preprocessing transformations including standard normal variate (SNV), multiplicative scatter correction (MSC), first derivative (D1) and second derivative (D2) were applied as spectra enhancement prior to prediction model development. The results obtained show that the SNV preprocessing followed by kNN is better than any other preprocessing approach and machine learning algorithm for the prediction of N, P, K content. The optimal prediction model for N and P content attributes were obtained by kNN with the closest neighbor of 4 components as a the best hyperparameter and the two components for prediction K content attribute prediction. The model's performance using the ratio of prediction to deviation (RPD) to predict N, P, K content attributes was higher than 7. The overall results satisfactorily demonstrate that NIRs technology combined with proper preprocessing and regression approaches has promising results to determine primary nutrients attributes of soil agriculture nondestructively.
