Hyperspectral Imaging Feature Selection Using Regression Tree Algorithm: Prediction of Carotenoid Content Velvet Apple Leaf

Abstract

Hyperspectral imaging system is an alternative in measuring biological content, especially in plants. Carotenoid content in leaves is one of the ingredients that can be measured using Vis-NIR hyperspectral camera because carotenoids are pigments that are in that range. The combination of spatial and spectral information produces many advantages; one of them is fast measurement time. Spatial and spectral information is extensive data that must be processed in making prediction systems. Spectral information is the wavelength that becomes features in machine learning. A large number of features results in increased computational costs and general rules of machine learning if too many features are used that will result in overfitting. Therefore, this study aims to increase computational costs and reduce overfitting by reducing features not related to the target. The use of supervised learning in selecting features can maintain wavelength information on carotenoid content which the unsupervised method cannot do. The system predicts carotenoid content with MAE and RMSE values obtained at 21.42 and 39.21 using the random forest model with decision tree feature selection.