Oil palm maturity classifier using spectrometer and machine learning
The quality of palm oil depends on fresh fruit bunch (FFB) ripeness level. Ripe bunch has higher oil quantity compared to unripe bunch. It also has less free fatty acid (FFA) compared to overripe bunch which reduces the quality of palm oil to become poor. Therefore, classification and grading of FFB...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2021
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/104094/1/GOH%20JIA%20QUAN%20-IR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The quality of palm oil depends on fresh fruit bunch (FFB) ripeness level. Ripe bunch has higher oil quantity compared to unripe bunch. It also has less free fatty acid (FFA) compared to overripe bunch which reduces the quality of palm oil to become poor. Therefore, classification and grading of FFB into correct categories and process them separately is an important step to avoid loss in quality of the extracted palm oil. Traditionally, the grading of FFB bunches is performed by well-trained graders according to different parameters such as mesocarp color, number of loose fruits on the ground and number of empty sockets on the bunches. This method depends heavily on human eyes which can be subjective and lead to different outcomes of grading between graders. Thus, non-destructive method is another option for tasks of FFB ripeness level classification. In this research, a spectrometer with a wavelength range of 180 to 1100 nm was applied to collect the reflectance data of FFB from unripe, ripe, and overripe classes. The three objectives in this study are (1) to determine the most suitable part of FFB for classifying oil palm ripeness level, (2) to identify the ideal vegetation index as prediction model for FFB classification and (3) To assess the classification accuracies and validate the selected prediction model. Each bunch was scanned at its different parts including apical, front equatorial, front basil, back equatorial and back basil. The reflectance data from these five parts was analyzed using statistical method and machine learning algorithm. Front equatorial was found to have significant difference between the three classes of ripeness, and an overall 92.7% of accuracy in differentiating between the maturity classes. Next, specific bands were extracted to compute vegetation indices for prediction model. Normalized Difference Vegetation Index (NDVI) is selected as the best prediction model with 93.8% classification accuracy. The accuracy assessment showed that NDVI has precision of 0.938, recall of 0.939 and F1-Score of 0.937. This shows a promising result of the NDVI as vegetation index to classify FFB ripeness level. The trained NDVI model was exported as prediction model that can assist in predicting ripeness level of FFB which can be applied by researcher and graders from the industry. The model was validated by predicting ripeness level for another FFB reflectance dataset. The prediction was able to produce 100% accuracies by using Linear and Weighted KNN as classification testing algorithm. An application was built by using the NDVI prediction model. It allows users to enter red and NIR reflectance values of FFB for the prediction of FFB ripeness level. Furthermore, the average accuracies of each classifier were compared. Fine KNN had the highest average accuracy of 68.6% whereas Coarse KNN had the lowest average accuracies of 36.0%. These findings provide valuable information to future researchers in this field to develop automatic oil palm FFB classifier. |
|---|