Most of the contemporary research published in field of visible-short wave near-infrared (Vis- SWNIR) fruit spectroscopy is 'derivative' in nature as they primarily showcase the application of existing spectroscopic devices coupled with traditional multivariate analysis techniques subjected to different fruits. The results of such studies often remain theoretical due to lack of integration mechanisms to incorporate developed models back into these proprietary devices. In present study, an open-source portable spectrophotometer has been proposed using commercially available multi-spectral sensor chipset integrated with microcontroller, housed in an ergonomically designed cabinet for spectral data acquisition in reflectance mode. During a 15-day post-harvest storage study of over 100 samples of raw tomatoes, the spectral data was acquired at 18 different wavelengths ranging from 410 nanometer (nm) to 940 nm, along with laboratory estimation 14 physicochemical attributes. Statistical and chemometrics analysis revealed that blue (380-440 nm) and green (440-600 nm) spectra varied with attributes associated with loss in water content, while red (600-750 nm) and SWNIR (750-1100 nm) spectra varied with attributes associated with carotenoid content, such as lycopene, antioxidant activity, and colour pigmentation. In order to predict physicochemical attributes using spectral wavelengths as regressor variables, various linear models including multiple linear regression (MLR), principal component regression (PCR), partial least squares regression (PLSR) and non-linear models including random forest regression (RF), support vector machine regression (SVM), and artificial neural network (ANN) were developed using 10-fold cross-validation on 80-20% train-test split of the dataset. MLR was observed to have exhibited a linear relationship between lycopene and wavelengths 560 nm, 645 nm and 730 nm with highest coefficient of determination (R-2) among linear models at 0.6361 (p < 0.05) but SVM outperformed all models with root mean squared error on test dataset, RMSE (Test) at 0.087 (p < 0.05). MLR was found to be best suitable for prediction of red colour pigmentation in tomatoes with 93.17% accuracy and lowest RMSE (Test) at 0.0608 (p < 0.05) for wavelengths 560 nm, 645 nm, 730 nm and 810 nm. While non-linear models demonstrated competence in capturing the underlying relationship between physicochemical attributes and spectra the optimised MLR models, albeit time and computationally dearer due to extensive pre-processing and criterion-based selection strategies, showed outstanding performance with high R-2 and competitive RMSE (Test). Additionally, these MLR models offered interpretability to facilitate better understanding of spectral features that were influential in making predictions of carotenoid associated attributes, which is often absent in contemporary research papers. The present work augments technological advancement in portable fruit spectroscopy for rapid and non-invasive estimation of physicochemical attributes through chemometrics-machine learning framework.
