Forward Feature Selection for Ensembles to Predict Brix Values in Mango Fruits based on NIR Spectroscopy Technique

Abstract

The purpose of this study is applying ensemble models with forward feature selection based on NIR spectrum datasets for predicting the Brix values of mangoes. Spectrum data of 4 groups of 300 mango fruits from NIR spectroscopy technique were used with forward feature selection to create datasets, and then ensemble models were built. Methods used for prediction were linear regressions (LR), neural networks (NN) and k-nearest neighbour (KNN). 112 ensemble models were a combination of methods and datasets. From the experiment, it indicated that lower standard deviation (SD) and root mean square error (RMSE) values were produced by higher harvesting-period mangoes. For the RMSE numbers, the LR ensemble model training with the 120-day harvesting period dataset and selecting features by all 3 methods (3M120) generated the least RMSE value. For the highest performance of predicting Brix values, the LR-NN-KNN ensemble model training with the 120-day harvesting period dataset and selecting features by KNN method performed well by giving the minimum SD value and the RMSE number close to the minimum one.