Abstract:
Analytical tests are commonly performed in laboratories to analyze and ensure food quality due to concerns about food adulteration. However, traditional analytical methods that rely on chemicals or equipment are often time-consuming and expensive. Therefore, we propose an efficient method for detecting starch adulterants in turmeric, which is clean, green, inexpensive, and rapid. Near-infrared (NIR) spectroscopy meets all these criteria and has a high potential for conducting routine assessments. To acquire reflectance spectra from the 900 nm-1700 nm range, we used the compact TI DLPNIRscan Nano module instead of a traditional bulky and costly spectrophotometer. Turmeric samples were adulterated with starch, ranging from 0 percent to 50 percent, and the Savitzky-Golay (SG) filter was applied to the recorded spectra. Various machine learning (ML) models were used to train and test these spectra, and the PCA approach was used to reduce the dimensionality of the data and assess its effectiveness. We have used several metrics, including R sup(2), Root Mean Square Error (RMSE sub(V)), Mean Absolute Error (MAE sub(V)), and Leave-one-out Cross-Validation (LOOCV), to evaluate the performance of the ML models. The Extra Tree Regressor (ETR) outperformed the other models, achieving an R sup(2) of 0.995, an RMSEV of 1.056 mg (W/W), an MAE of 0.597 mg (W/W), a LOOCV R sup(2) of 0.994, and a LOOCV RMSE sub(V) of 1.038 mg (W/W).
