Prediksi Kandungan Kimia Biji Kopi Robusta dengan Near Infrared Spectoscopy dan Artificial Neural Network

Abstract

Pengembangan metode cepat dan hemat biaya untuk memprediksi kandungan kimia kopi robusta sangat diperlukan, mengingat analisis konvensional membutuhkan waktu lama dan biaya yang mahal. Penentuan mutu kopi robusta di Pagar Alam selama ini hanya mengandalkan analisis sensori, yang hasilnya kurang konsisten. Oleh sebab itu, dibutuhkan metode alternatif yang lebih objektif dan andal mengingat penentuan mutu kopi merupakan suatu hal yang penting. Berdasarkan permasalahan tersebut penggunaan near infrared spectroscopy (NIRS) dapat menjadi solusi yang potensial karena mampu mengukur kandungan kimia secara cepat, nondestruktif dan akurat. Penelitian ini bertujuan untuk memprediksi kandungan kimia biji kopi robusta Pagar Alam menggunakan NIRS dan kalibrasi partial least square (PLS) dan principal component analysis-artificial neural network (PCA–ANN). Pada penelitian ini reflektansi biji kopi robusta diukur menggunakan alat NIRFlex N-500 spektrometer dengan panjang gelombang 1000–2500 nm. Selanjutnya sampel diukur kandungan kimianya secara konvensional di laboratorium. Data reflektansi diproses menggunakan pre-treatment spektra. Pre-treatment yang digunakan yaitu normalization (No1), turunan pertama savitzky-golay (SG1), multiple scatter correction (MSC), kombinasi No1 dan SG1 (No1SG1) dan kombinasi MSC dan SG1 (MSCSG1). Selanjutnya dilakukan kalibrasi dan validasi dengan kandungan kimia menggunakan PLS dan PCA–ANN. Kinerja model dievalusi menggunakan parameter kinerja statistik, yaitu koefisien korelasi (r), standar error of calibration set (SEC), standar error of prediction set (SEP), coefficient of variation (CV), Ratio of performance to deviation (RPD) dan konsistensi. Hasil penelitian menunjukkan bahwa model terbaik untuk kadar air diperoleh dengan pre-treatment No1SG1 pada metode PLS menggunakan 8 faktor PLS (r = 0,92; RPD = 2,42 dan konsistensi = 94,00%), sedangkan pada metode PCA–ANN dengan 8 PC diperoleh (r = 0,95; RPD = 2,95 dan konsistensi = 90,24%). Prediksi protein dicapai menggunakan pre-treatment No1SG1 dengan 8 faktor PLS (r = 0,91; RPD = 2,26 dan konsistensi = 97,56%), sementara pada metode PCA–ANN dengan 8 PC diperoleh (r = 0,92; RPD = 2,38 dan konsistensi = 97,44%). Prediksi lemak optimal diperoleh dengan pre-treatment No1SG1 pada metode PLS menggunakan 9 faktor (r = 0,91; RPD = 1,92 dan konsistensi = 80,77%), metode PCA–ANN dengan 10 PC diperoleh (r = 0,93; RPD = 2,27 dan konsistensi = 81,82%). Kadar abu terbaik pada metode PLS diperoleh melalui pre-treatment MSCSG1 dengan 7 faktor PLS (r = 0,91; RPD = 2,05 dan konsistensi = 86,36%), sementara pada metode PCA–ANN dengan pre-treatment SG1 dan 8 PC didapatkan hasil (r = 0,93; RPD = 2,25 dan konsistensi = 85,00%). Prediksi karbohidrat optimal menggunakan pre-treatment No1SG1 dengan 10 faktor PLS (r = 0,91; RPD = 2,29 dan konsistensi = 94,81%), sedangkan metode PCA–ANN dengan 5 PC diperoleh (r = 0,93; RPD = 2,76 dan konsistensi = 109,38%). Pada prediksi kafein, hasil terbaik dicapai dengan pre-treatment MSCSG1 dengan 7 faktor PLS (r = 0,90; RPD = 2,42 dan konsistensi = 98,24%), sementara pada metode PCA–ANN dengan 10 PC diperoleh (r = 0,94; RPD = 3,45 dan konsistensi 109,88%). Hasil ini menunjukkan bahwa NIRS dapat digunakan untuk memprediksi kandungan kimia biji kopi robusta Pagar Alam secara akurat.

The development of a rapid and cost-effective method for predicting the chemical composition of robusta coffee is essential, considering that conventional analysis is time-consuming and costly. The quality assessment of robusta coffee in Pagar Alam has so far relied solely on sensory analysis, which often produces inconsistent results. Therefore, an alternative method that is more objective and reliable is necessary, as quality determination is a crucial aspect in coffee evaluation. Based on these issues, the use of near-infrared spectroscopy (NIRS) presents a promising solution, as it is capable of measuring chemical content in a nondestructive, rapid, and accurate manner. This study aims to predict the chemical composition of robusta coffee beans from Pagar Alam using near-infrared spectroscopy (NIRS) and calibration based on partial least squares (PLS) and principal component analysis–artificial neural network (PCA–ANN). In this study, the reflectance of robusta coffee beans was measured using the NIRFlex N-500 Spectrometer within a wavelength range of 1000–2500 nm. Subsequently, the chemical content of the samples was measured conventionally in the laboratory. The reflectance data were processed using spectral pre-treatment methods. The pre-treatments applied included normalization (No1), first derivative Savitzky-Golay (SG1), multiple scatter correction (MSC), a combination of No1 and SG1 (No1SG1), and a combination of MSC and SG1 (MSCSG1). Calibration and validation were then performed using chemical content data through PLS and PCA–ANN. The model performance was evaluated using statistical performance parameters, namely the correlation coefficient (r), standard error of calibration (SEC), standard error of prediction (SEP), coefficient of variation (CV), ratio of performance to deviation (RPD), and consistency. The results of the study showed that the best model for moisture content was obtained using the No1SG1 pre-treatment with the PLS method employing 8 PLS factors (r = 0.92; RPD = 2.42 and consistency = 94.00%), while with the PCA–ANN method using 8 PCs, the results were (r = 0.95; RPD = 2.95 and consistency = 90.24%). Protein prediction was achieved using the No1SG1 pre-treatment with 8 PLS factors (r = 0.91; RPD = 2.26 and consistency = 97.56%), whereas with the PCA–ANN method using 8 PCs, the results were (r = 0.92; RPD = 2.38 and consistency = 97.44%). The optimal fat prediction was obtained using the No1SG1 pre-treatment with the PLS method employing 9 factors (r = 0.91; RPD = 1.92 and consistency = 80.77%), while the PCA–ANN method with 10 PCs produced (r = 0.93; RPD = 2.27 and consistency = 81.82%). The best ash content prediction using the PLS method was achieved with the MSCSG1 pre-treatment and 7 PLS factors (r = 0.91; RPD = 2.05 and consistency = 86.36%), whereas the PCA–ANN method using the SG1 pre-treatment and 8 PCs yielded (r = 0.93; RPD = 2.25 and consistency = 85.00%). The optimal carbohydrate prediction was obtained using the No1SG1 pre-treatment with 10 PLS factors (r = 0.91; RPD = 2.29 and consistency = 94.81%), while the PCA–ANN method with 5 principal components yielded (r = 0.93; RPD = 2.76 and consistency = 109.38%). For caffeine prediction, the best results were achieved using the MSCSG1 pre-treatment with 7 PLS factors (r = 0.90; RPD = 2.42 and consistency = 98.24%), while the PCA–ANN method with 10 PCs produced (r = 0.94; RPD = 3.45 and consistency = 109.88%). These results indicate that NIRS can be effectively used to accurately predict the chemical composition of Coffea robusta beans from Pagar Alam.