Prediksi Kandungan Kimia Buah Kelapa Sawit secara Non-Destruktif Menggunakan Portable Spectrometer dengan Metode Machine Learning dan PLSR

Abstract

Indonesia merupakan salah satu negara penghasil minyak kelapa sawit terbesar di dunia. Kandungan buah kelapa sawit sangat menentukan mutu minyak yang dihasilkan. Penentuan kandungan buah kelapa sawit saat ini masih dilakukan secara konvensional dengan cara visual yang subjektif dan dengan laboratorium yang memakan waktu serta biaya tinggi. Metode yang efektif dibutuhkan, salah satunya dengan menggunakan spektrometer portabel yang lebih efisien dan tidak merusak buah. Penelitian ini bertujuan untuk mengembangkan metode prediksi kandungan kadar air, kadar minyak, dan kadar asam lemak bebas buah kelapa sawit dengan menggunakan instrumen SCiO NIR portabel dengan metode SVR dan PLSR. Sebanyak 408 sampel buah kelapa sawit diukur reflektansinya pada panjang gelombang 740-1070 nm, selanjutnya data kimia buah ditentukan melalui analisis kimia laboratorium. Spektra reflektan hasil pengukuran diolah menggunakan pra-pemrosesan spektra smoothing savitzky-golay, detrending, normalize, standard normal variate (SNV), dan kombinasi SNV+smoothing SG yang diolah menggunakan software The Unscrambler X kemudian dilakukan kalibrasi dan validasi untuk mengukur performa metode. Hasil terbaik untuk memprediksi kadar air dan minyak masing-masing adalah metode SVR menggunakan pre-treatment normalize dan SNV dengan akurasi (R2 = 0,99; SEC = 2,75%; SEP = 3,37%; RPD = 6,72), (R2 = 0,92; SEC = 5,02%; SEP = 5,64%; RPD = 3,12). Sedangkan hasil terbaik untuk prediksi kadar ALB adalah metode PLSR dengan 15 factor component tanpa pre-treatment (R2 = 0,51; SEC = 1,38%; SEP = 1,37%; RPD = 1,42). Metode kalibrasi SVR dengan pre-treatment normalize dan SNV dari pengukuran portable spectrometer dapat digunakan untuk prediksi kadar minyak dan air secara akurat, namun metode tidak dapat digunakan untuk prediksi kadar asam lemak secara akurat.

Kata kunci : kelapa sawit, machine learning, NIR, PLSR, SVR

Indonesia is one of the world’s largest producers of palm oil. The composition of the oil palm fruit is a key determinant of the quality of the oil produced. Currently, the composition of the oil palm fruit is still determined using conventional methods, such as subjective visual inspection and laboratory testing, which are time consuming and costly. An effective method is needed, such as the use of a portable spectrometer that is more efficient and does not damage the fruit. This study aims to develop a method for predicting the moisture content, oil content, and free fatty acid content of oil palm fruit using a portable SCiO NIR instrument with SVR and PLSR methods. The reflectance of 408 oil palm fruit samples was measured at wavelengths of 740-1070 nm, and the fruit’s chemical data were subsequently determined through laboratory chemical analysis. The measured reflectance spectra were processed using Savitzky-Golay (SG) smoothing, detrending, normalization, standard normal variate (SNV), and a combination of SNV and SG smoothing via The Unscrambler X software. Calibration and validation were then performed to assess the method’s performance. The best results for predicting moisture and oil content, respectively, were obtained using the SVR method with normalization pre-processing and the SNV method, with an accuracy of (R2 = 0,99; SEC = 2,75%; SEP = 3,37%; RPD = 6,72), (R2 = 0,92; SEC = 5,02%; SEP = 5,64%; RPD = 3,12). Meanwhile, the best results for predicting ALB content were obtained using the PLSR method with 15 principal components without pre-treatment (R² = 0.51; SEC = 1,38%; SEP = 1,37%; RPD = 1,42). The SVR calibration method with normalization and SNV pre-treatment of portable spectrometer measurements can be used to accurately predict oil and water content, but the method cannot be used to accurately predict fatty acid content.

Keywords: machine learning, NIR, oil palm, PLSR, SVR