Smartphone-based NIR imaging for real-time food quality grading

Thesis Abstract

Smartphone-based near-infrared (NIR) imaging offers a scalable, low-cost solution for non-destructive assessment of food quality at the point of processing and sale. This study addresses the persistent challenge of rapid, accurate quality grading in heterogeneous agricultural commodities by leveraging consumer-grade smartphones augmented with a compact NIR sensing module and machine learning-based calibration. The aim is to develop a robust, real-time pipeline that maps NIR-derived spectral features to key quality attributes (e.g., sugar content, moisture, internal bruising) across multiple produce types. Specific objectives include (i) to characterize the spectral response of apples, avocados, and berries within the 900–1700 nm range using a 5-megapixel smartphone-integrated NIR sensor; (ii) to construct calibration models that predict firmness, soluble solids content, and internal defects with cross-validated accuracy above 85%; (iii) to evaluate the influence of ambient lighting and device geometry on model performance and implement a normalization strategy to mitigate these effects; (iv) to compare linear and non-linear machine learning approaches and identify the optimal algorithm for each quality attribute; and (v) to develop a user-friendly mobile application workflow for farmers and small-scale processors. A cross-sectional study design will be employed, enrolling 600 fruit samples (200 apples, 200 avocados, 200 berries) collected from three commercial orchards and two markets across a growing season. Spectral data will be captured with a smartphone-mounted NIR sensor, accompanied by reference laboratory measurements soluble solids content (BZ refractometry), moisture content ( oven-dried method), firmness (penetrometry), and internal defects (CT or ultrasound for a subset). Data collection instruments will include the smartphone NIR sensor, a calibrated reference spectrometer for spectral alignment, a digital caliper for size normalization, and standard lab protocols for ground-truth attributes. Model development will involve preprocessing steps (Savitzky–Golay smoothing, baseline correction), feature extraction (principal component analysis, successive projections algorithm), and predictive modeling using partial least squares regression (PLSR), support vector regression (SVR), random forest, and gradient boosting. Validation will follow a nested cross-validation scheme and external test sets to assess generalizability across varieties and origins. The analytical framework will incorporate a theoretical basis grounded in the Theory of Perceived Quality and Surface-to-Interior Correlation in spectroscopic sensing, as well as instrumental learning theory for model generalization under device variability. Statistical performance metrics will include RMSE, R^2, and mean absolute error (MAE) for quantitative attributes, alongside confusion matrices and F1 scores for discretized quality classes. Sensitivity analyses will quantify the impact of ambient light (100–1000 lux) and device-to-sample distance (2–8 cm) on predictive accuracy, guiding the design of normalization and data augmentation procedures. Ethical considerations will address data provenance, farmer consent, and equitable access implications. Expected findings indicate that smartphone-based NIR imaging can predict soluble solids content and moisture with R^2 > 0.85 and RMSE within industry-accepted tolerances for apples and berries, and R^2 > 0.80 with RMSE comparable for avocados, after normalization for lighting and geometry. Non-linear models are anticipated to outperform linear approaches for complex internal defect detection, while multi-attribute fusion will enhance overall grading accuracy by 8–12% relative to single-attribute models. The study will contribute to knowledge by demonstrating the feasibility of portable, ICT-enabled spectroscopic sensing for real-time quality grading across diverse fruit types, outlining a scalable calibration framework, and delivering a prototype mobile app with decision support features for end users. Conclusions are expected to endorse smartphone-based NIR imaging as a viable, cost-effective tool for on-farm quality control, with recommendations including standardization of accessory kits for consistent spectral acquisition, development of transferable calibration models via domain adaptation techniques, and policy guidance to support adoption in small to medium-scale agricultural enterprises. Potential limitations include dependence on sample diversity and the need for continued calibration maintenance; future work will explore additional commodities, broader spectral ranges, and integration with blockchain-based traceability for supply-chain transparency.