Classification of Palm Oil Fresh Fruit Bunch using Multiband Optical Sensors

Abstract

This study investigated optical sensor system consist of sixteen light emitting diode (LED) in visible/near infrared region to detect palm oil fresh fruit bunch (FFB) quality. Practically, experience grader assessed FFB quality by its ripeness based on external features such as colour and number of detached fruitlets. However, different seed and plantation management resulting in FFB quality variation. Same external features not linearly correlate with FFB oil content that corresponding with industrial needs. The 660 nm LED is choosen to be used to estimate the oil content of FFB. Using linear discriminant analysis (LDA) with Mahalanobis distance, the accuracy of the systems is 79.8% and 88.2%. From 33 FFB oil content measurement, grader misclassified 4 out of 17 FFB as ripe FFB but with low oil content (<17.5%) and misclassified 7 out of 16 FFB as unripe but with high oil content (>=17.5%). Classifying model build from FFB from main plantation then tested to evaluate FFB from smallholder. Classification model generated from FFB oil content data showed more accurate result compared to model generated from visual inspection 66.7% compared to 52.1%. Model accuracies attained by Discriminant Analysis (DA) and k-Nearest Neighbors (k-NN) were 79.8% and 80.7%, respectively based on grader evaluation. Model accuracies based on FFB oil content was 88.2% for both classifying algorithms.

Classification of healthy and white root disease infected rubber trees based on relative permittivity and capacitance input properties using LM and SCG artificial neural network

“…From Figure 2(a), for single input relative permittivity measurement, the selected optimised model from SCG had a value of sensitivity equal to 79.00% and specificity equal to 69.00% compared with the optimised model from LM, which had a value of sensitivity equal to 77.00% and specificity equal to 73.67%. From Figure 2 Accuracy and Area under Curve (AUC) are important factors in determining the best model in the ANN classification process and it has been proven that these two factors have the ability in determining the measurement for evaluating the best model [23][24][25]. The evaluation factors contributing to the performance of the selected models were further discussed towards accuracy and AUC.…”

Classification of healthy and white root disease infected rubber trees based on relative permittivity and capacitance input properties using LM and SCG artificial neural network

“…From Figure 2(a), for single input relative permittivity measurement, the selected optimised model from SCG had a value of sensitivity equal to 79.00% and specificity equal to 69.00% compared with the optimised model from LM, which had a value of sensitivity equal to 77.00% and specificity equal to 73.67%. From Figure 2 Accuracy and Area under Curve (AUC) are important factors in determining the best model in the ANN classification process and it has been proven that these two factors have the ability in determining the measurement for evaluating the best model [23][24][25]. The evaluation factors contributing to the performance of the selected models were further discussed towards accuracy and AUC.…”

Oil Palm Fresh Fruit Bunch Ripeness Detection Methods: A Systematic Review

“…Color vision Fresh fruit bunches [8,44,45] Oil palm fruitlets [4] LiDAR sensor Fresh fruit bunches [11,13] Oil palm fruitlets [12] NIR Fresh fruit bunches [17] Oil palm fruitlets [18] Hyperspectral camera Fresh fruit bunches [10] Optical sensor Fresh fruit bunches [35] Optical spectrometer Fresh fruit bunches [46] Multiband optical sensor Fresh fruit bunches [14] Optical sensor system Fresh fruit bunches [15] Raman spectra Oil palm fruitlets [21,47,48] Fluorescence sensor Fresh fruit bunches [22,23] Inductive sensor Oil palm fruitlets [28,29] Fresh fruit bunches [30] Laser-light backscattering and computer vision Fresh fruit bunches [38] Thermal vision Fresh fruit bunches [39] Thermal imaging Fresh fruit bunches [40,49] Fruit battery Oil palm fruitlets [41,42] Fruit battery and computer vision Oil palm fruitlets [43]…”

“…In our studies, the hyperspectral camera is not the only appliance for building an optical sensor system. Others researchers used several LEDs with different wavelengths as their source light to capture images of the FFBs for determining the maturity stage [14]. Moreover, the FFB image data were collected using a handheld four-band active sensor [15].…”

Fresh Fruit Bunch Ripeness Classification Methods: A Review

“…The statistical results from the study demonstrated significant differences in ripeness at a wavelength of 660 nm. The captured images were processed with linear discriminant analysis (LDA) with Mahalanobis distance, yielding an average accuracy of 88.2% (Setiawan et al, 2019). Bensaeed et al (2014) used a hyperspectral-based system to classify FFB ripeness by recording the reflectance at various wavelengths.…”

“…Linear regression -Simple and interpretable -Assumes linearity, may not capture complex patterns, sensitive to outliers -Low computational demand (Cherie et al, 2015) 2019 LDA -Efficient in dimensionality reduction, provides optimal linear discriminants -Assumes normality and equal covariance matrices, sensitive to outliers -Moderate computational demand (Setiawan et al, 2019) Raman spectroscopy 2018 ANN -Offers flexibility in capturing complex relationships in data -Suffer from overfitting, especially with limited data -Training can be time-consuming (Dan et al, 2018)…”

Classification of healthy and white root disease infected rubber trees based on relative permittivity and capacitance input properties using LM and SCG artificial neural network

“…From Figure 2(a), for single input relative permittivity measurement, the selected optimised model from SCG had a value of sensitivity equal to 79.00% and specificity equal to 69.00% compared with the optimised model from LM, which had a value of sensitivity equal to 77.00% and specificity equal to 73.67%. From Figure 2 Accuracy and Area under Curve (AUC) are important factors in determining the best model in the ANN classification process and it has been proven that these two factors have the ability in determining the measurement for evaluating the best model [23][24][25]. The evaluation factors contributing to the performance of the selected models were further discussed towards accuracy and AUC.…”

Oil Palm Fresh Fruit Bunch Ripeness Detection Methods: A Systematic Review

“…Color vision Fresh fruit bunches [8,44,45] Oil palm fruitlets [4] LiDAR sensor Fresh fruit bunches [11,13] Oil palm fruitlets [12] NIR Fresh fruit bunches [17] Oil palm fruitlets [18] Hyperspectral camera Fresh fruit bunches [10] Optical sensor Fresh fruit bunches [35] Optical spectrometer Fresh fruit bunches [46] Multiband optical sensor Fresh fruit bunches [14] Optical sensor system Fresh fruit bunches [15] Raman spectra Oil palm fruitlets [21,47,48] Fluorescence sensor Fresh fruit bunches [22,23] Inductive sensor Oil palm fruitlets [28,29] Fresh fruit bunches [30] Laser-light backscattering and computer vision Fresh fruit bunches [38] Thermal vision Fresh fruit bunches [39] Thermal imaging Fresh fruit bunches [40,49] Fruit battery Oil palm fruitlets [41,42] Fruit battery and computer vision Oil palm fruitlets [43]…”

“…In our studies, the hyperspectral camera is not the only appliance for building an optical sensor system. Others researchers used several LEDs with different wavelengths as their source light to capture images of the FFBs for determining the maturity stage [14]. Moreover, the FFB image data were collected using a handheld four-band active sensor [15].…”

Fresh Fruit Bunch Ripeness Classification Methods: A Review

“…The statistical results from the study demonstrated significant differences in ripeness at a wavelength of 660 nm. The captured images were processed with linear discriminant analysis (LDA) with Mahalanobis distance, yielding an average accuracy of 88.2% (Setiawan et al, 2019). Bensaeed et al (2014) used a hyperspectral-based system to classify FFB ripeness by recording the reflectance at various wavelengths.…”

“…Linear regression -Simple and interpretable -Assumes linearity, may not capture complex patterns, sensitive to outliers -Low computational demand (Cherie et al, 2015) 2019 LDA -Efficient in dimensionality reduction, provides optimal linear discriminants -Assumes normality and equal covariance matrices, sensitive to outliers -Moderate computational demand (Setiawan et al, 2019) Raman spectroscopy 2018 ANN -Offers flexibility in capturing complex relationships in data -Suffer from overfitting, especially with limited data -Training can be time-consuming (Dan et al, 2018)…”

Classification of healthy and white root disease infected rubber trees based on relative permittivity and capacitance input properties using LM and SCG artificial neural network

“…From Figure 2(a), for single input relative permittivity measurement, the selected optimised model from SCG had a value of sensitivity equal to 79.00% and specificity equal to 69.00% compared with the optimised model from LM, which had a value of sensitivity equal to 77.00% and specificity equal to 73.67%. From Figure 2 Accuracy and Area under Curve (AUC) are important factors in determining the best model in the ANN classification process and it has been proven that these two factors have the ability in determining the measurement for evaluating the best model [23][24][25]. The evaluation factors contributing to the performance of the selected models were further discussed towards accuracy and AUC.…”

Oil Palm Fresh Fruit Bunch Ripeness Detection Methods: A Systematic Review

“…Color vision Fresh fruit bunches [8,44,45] Oil palm fruitlets [4] LiDAR sensor Fresh fruit bunches [11,13] Oil palm fruitlets [12] NIR Fresh fruit bunches [17] Oil palm fruitlets [18] Hyperspectral camera Fresh fruit bunches [10] Optical sensor Fresh fruit bunches [35] Optical spectrometer Fresh fruit bunches [46] Multiband optical sensor Fresh fruit bunches [14] Optical sensor system Fresh fruit bunches [15] Raman spectra Oil palm fruitlets [21,47,48] Fluorescence sensor Fresh fruit bunches [22,23] Inductive sensor Oil palm fruitlets [28,29] Fresh fruit bunches [30] Laser-light backscattering and computer vision Fresh fruit bunches [38] Thermal vision Fresh fruit bunches [39] Thermal imaging Fresh fruit bunches [40,49] Fruit battery Oil palm fruitlets [41,42] Fruit battery and computer vision Oil palm fruitlets [43]…”

“…In our studies, the hyperspectral camera is not the only appliance for building an optical sensor system. Others researchers used several LEDs with different wavelengths as their source light to capture images of the FFBs for determining the maturity stage [14]. Moreover, the FFB image data were collected using a handheld four-band active sensor [15].…”

Fresh Fruit Bunch Ripeness Classification Methods: A Review

“…The statistical results from the study demonstrated significant differences in ripeness at a wavelength of 660 nm. The captured images were processed with linear discriminant analysis (LDA) with Mahalanobis distance, yielding an average accuracy of 88.2% (Setiawan et al, 2019). Bensaeed et al (2014) used a hyperspectral-based system to classify FFB ripeness by recording the reflectance at various wavelengths.…”

“…Linear regression -Simple and interpretable -Assumes linearity, may not capture complex patterns, sensitive to outliers -Low computational demand (Cherie et al, 2015) 2019 LDA -Efficient in dimensionality reduction, provides optimal linear discriminants -Assumes normality and equal covariance matrices, sensitive to outliers -Moderate computational demand (Setiawan et al, 2019) Raman spectroscopy 2018 ANN -Offers flexibility in capturing complex relationships in data -Suffer from overfitting, especially with limited data -Training can be time-consuming (Dan et al, 2018)…”