Optimization of Preprocessing Spectra and Modeling Using Machine Learning for Prediction of Agricultural Soil Nutrients

Abstract

Research that addresses soil nutrient prediction with NIR mostly uses PLS algorithms. The advent of machine learning (ML) has resulted in automated learning methods to find the optimal model. Preprocessing in the development of prediction models is an important part. Spectrum preprocessing aims to eliminate uninformative sources of variance. Research on the best preprocessing methods is often determined through trial-and-error. The preprocessing approach compares a number of preprocessing operations but this method is less efficient. This research proposes the application of ML to find the best combination of preprocessing operations quickly and simultaneously. Preprocessing test results using 11 operators resulted in 2,112 combinations. The use of preprocessing techniques can improve the performance of all algorithms (RF, SVR, PLS, LR, and MLP). K soil element testing has the lowest error in LR, Mg, Ca, P, and pH soil element testing using MLP has the best performance and in N soil element testing the best performance in RF.