Smartphone-based Near-Infrared Food Quality Monitoring System for Fresh Produce

Thesis Abstract

This study addresses the challenge of rapid, non-destructive assessment of fruit and vegetable quality to reduce post-harvest losses and improve supply chain transparency by leveraging smartphone-based near-infrared (NIR) spectroscopy integrated with machine learning. The aim is to develop and validate a portable, user-friendly quality monitoring system that provides real-time estimates of soluble solids content (SSC), firmness, and internal defect indicators for common fresh produce. Specific objectives include (1) assembling a diverse dataset of 1,200 fruit and vegetable samples across three commodity groups (apples, tomatoes, and strawberries) collected from commercial markets, (2) capturing calibrated NIR spectral data (900–1700 nm) using a smartphone-mounted accessory and recording reference laboratory measurements (refractive index, refractometry for SSC, texture analyzer for firmness, and defect grading by expert panels), (3) developing calibration models using partial least squares regression (PLSR) and support vector regression (SVR) with spectral preprocessing (SNV, derivative transforms), (4) exploring deep learning approaches (1D-CNN) for end-to-end quality prediction, and (5) validating the system under field conditions with 150 independent samples to assess robustness across devices and ambient lighting. The methodology adopts a mixed-methods, applied research design anchored in predictive analytics and user-centered evaluation. The population comprises commercially sourced fresh produce in a temperate agro-food corridor. A stratified random sampling strategy yields 1,200 fruit and vegetable samples for model development and 150 samples for external validation. Data collection involves (a) smartphone-based NIR spectral acquisition using a standardized attachment and a controlled 3D-printed cradle to minimize geometry effects, (b) laboratory reference measurements for SSC (refractometry), firmness (universal testing machine), and visual/defect scores by trained graders, and (c) user feedback from 30 produce supply-chain professionals via structured interviews. Instrument validity and reliability are established through calibration checks, inter-rater reliability for defect scoring (Cohen’s kappa > 0.80), and cross-validation of spectral sensors with a benchtop spectrometer. Data analysis proceeds in three tiers spectral preprocessing and conventional calibration (PLSR, SVR with RMSE and R^2 evaluation), machine learning enhancement (1D-CNN with spectro-temporal features), and cross-commodity transferability analysis using domain adaptation with a subset of apples, tomatoes, and strawberries. Model performance is assessed against independent test sets and industry-relevant thresholds (SSC ±0.5 Brix, firmness ±0.8 N, defect probability error < 5%). Expected findings indicate that smartphone-based NIR spectra, when coupled with robust preprocessing, can predict SSC with R^2 > 0.85 and RMSE < 0.6 Brix, firmness with R^2 > 0.80 and RMSE < 0.9 N, and detect internal defects with sensitivity > 0.75 and specificity > 0.85. The study anticipates that PLSR and SVR provide stable baselines across commodities, while the 1D-CNN approach enhances accuracy for samples with complex surface textures or moisture variations. The research also expects that domain adaptation techniques will improve cross-commodity generalization, enabling a single smartphone app to calibrate for multiple produce types with minimal additional data. The contribution to knowledge includes (i) a validated, scalable framework for smartphone-based NIR quality assessment in fresh produce, (ii) comparative evidence on traditional regression versus deep learning for spectral quality prediction in field conditions, and (iii) practical guidelines for device calibration, ambient-light mitigation, and user training in real-world supply chains. The study’s theoretical contribution integrates the Technology Acceptance Model with predictive analytics performance theories to explain adoption drivers and the Linkage Theory of Spectral Informatics to elucidate how spectral features map to physicochemical quality attributes. Practical implications encompass actionable protocols for app-based quality scoring, integration with supply-chain traceability systems, and policy-relevant recommendations for reducing post-harvest losses. Potential limitations include sensor heterogeneity across smartphone models and variable cropping seasons, which are mitigated through device calibration protocols and robust cross-validation. The principal conclusion anticipates that a smartphone-integrated NIR system can deliver accurate, rapid, non-destructive quality assessments compatible with routine handling by farmers, traders, and retailers, with recommendations emphasizing standardized measurement routines, periodic recalibration, and user training to sustain performance over time.