Analisis Mutu Vanili (Vanilla planifolia) Secara Cepat dan Portabel Berbasis Near Infrared Spectroscopy dan Machine Learning

Abstract

Penentuan mutu vanili secara cepat, akurat, dan non-destruktif menjadi tantangan penting dalam industri vanili. Spektroskopi Near-Infrared (NIR) portabel dikombinasikan dengan algoritma machine learning menawarkan pendekatan inovatif yang efisien dan praktis. Penelitian ini bertujuan mengembangkan model prediksi kadar air dan vanilin serta klasifikasi mutu (grade) vanili menggunakan data spektra NIR portabel. Dua rentang panjang gelombang dikaji, yaitu short wave NIR (SW-NIR: 740–1070 nm) dan long wave NIR (LW-NIR: 1350–2550 nm), untuk mengevaluasi pengaruh spektral terhadap akurasi model. Algoritma Random Forest (RF), Support Vector Regression (SVR), dan Partial Least Square (PLS) digunakan untuk prediksi, sementara klasifikasi dilakukan menggunakan RF, Decision Tree, k-Nearest Neighbors, dan Naïve Bayes. Data diproses dengan berbagai teknik preprocessing, termasuk Min-Max normalization, SNV, MSC, dan First Derivative serta kombinasinya. Evaluasi model mencakup R², RMSE, RPD, RER, LOD, LOQ, dan untuk klasifikasi digunakan F1-score serta AUC. Hasil menunjukkan bahwa RF memberikan performa terbaik secara konsisten untuk semua tujuan. Model klasifikasi terbaik (F1-score 0,85; AUC 0,91) diperoleh dari kombinasi RF dan first derivative–SNV pada spektra SW-NIR. Prediksi kadar air terbaik dicapai pada spektra LW-NIR (R² 0,97; RPD 7,75), sedangkan prediksi kadar vanilin terbaik diperoleh pada SW-NIR dengan Min-Max normalization (R² 0,94; RPD 4,58). Temuan ini menunjukkan potensi besar teknologi NIR portabel berbasis machine learning dalam mendukung pengujian mutu vanili yang cepat, non-destruktif, dan efisien.

Rapid, accurate, and non-destructive quality assessment of vanilla remains a critical challenge in the vanilla industry. Portable Near-Infrared (NIR) spectroscopy combined with machine learning algorithms offers an innovative, efficient, and practical approach. This study aimed to develop predictive models for moisture and vanillin content, as well as classification models for vanilla quality grades, using spectral data obtained from portable NIR devices. Two wavelength ranges were examined—short-wave NIR (SW-NIR: 740–1070 nm) and long-wave NIR (LW-NIR: 1350–2550 nm)—to evaluate their influence on model accuracy. Random Forest (RF), Support Vector Regression (SVR), and Partial Least Squares (PLS) were employed for prediction tasks, while classification was performed using RF, Decision Tree, k-Nearest Neighbors, and Naïve Bayes. Spectral data were processed using various preprocessing techniques, including Min-Max normalization, Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), First Derivative, and their combinations. Model performance was evaluated using R², RMSE, RPD, RER, LOD, and LOQ for prediction models, and F1-score and AUC for classification models. The results demonstrated that RF consistently delivered the best performance across all objectives. The best classification model (F1-score 0.85; AUC 0.91) was achieved using RF combined with first derivative–SNV preprocessing on SW-NIR spectra. The best moisture prediction was obtained from LW-NIR spectra (R² 0.97; RPD 7.75), while the best vanillin prediction was achieved using SW-NIR with Min-Max normalization (R² 0.94; RPD 4.58). These findings highlight the strong potential of portable NIR technology integrated with machine learning for rapid, non-destructive, and efficient vanilla quality assessment.