Advance Models for Camera-Vision Based Oil Content Prediction of Intact Oil Palm Fruits (Elaeis Guineensis Jacq) on Trees with Nondestructive Evaluation

Abstract — In this study, the correlation between oil palm fresh fruits bunch (FFB) appearance and its oil content (OC) was explored. The FFB samples were recorded from various distance (2, 7, 10, and 15 m) with different lighting spectrums (Ultraviolet: 280-380nm, visible: 400-700nm, and Infrared: 720-1100nm) and light intensities (600watt and 1000watt lamps). The FFB images were segmented and its color features were subsequently extracted to be used as input variables for modeling the OC of the sample. Twenty models were developed, one for every experiment-setup arrangement, using the MLP-ANN methods. Statistical engineering software was used to create the models. The number of FFB samples in this study was limited. Consequently, only five out of 20 developed models were selected and consider as valid to predict the FFB’s OC. The five models were namely UV-15m, UV-10m, UV-2m, Vis1-7m, and IR1-10m. The coefficient of correlation for each model when validated was 0.962, 0.897, 0.993, 0.993, and 0.401 respectively. For bias and error, these selected models have root-mean-square error (RMSE) of 2.385, 3.707, 2.868, 3.795, and 6.374. The best of five selected models was developed for predicting FFB’s OC in Vis1-7m arrangement, i.e. recording FFB images under visible light with 600watt lamps from 7 m of distance