Wavelength Selection of Persimmon Leafusing Decision Tree Method in Visible Near-Infrared Imaging

Abstract

Phenolic compounds are one of the secondary metabolites in vegetation. In general, total phenolic content can be measured using a biological approach that requires some preparation time and destructive. In this study, total phenolic content was predicted using Visible Near-Infrared (VNIR) Imaging approach. VNIR analysis in the spectral range of 400-1000 nm was used to predict the total phenolic content of velvet apple leaf non-destructively. Spectral features from samples are calculated based on the average reflectances area of leaves with a spatial dimension of 20×20 pixels in 224 spectral features. The optimal feature selection was performed using the Decision Tree (DT) method. Decision Tree Regression (DTR) algorithm was applied to predict measured values based on spectral features. Sample data evaluated with cross-validation to calculated system perform. The best performance of prediction system which has 30 optimal wavelength band with the determination coefficient (R2) of 0.92 and root mean square of the relative error (RMSE) of 3.48 in predicting the total phenolic content in a Velvet apple leaf.