Spectroscopic sensor-based machine learning framework for in situ evaluation of fruit internal quality

Non-destructive assessment of fruit internal quality is central to harvest management, postharvest sorting, and supply-chain optimization, yet destructive reference assays remain slow and unsuitable for large-scale in situ use. Spectroscopic sensing offers a practical alternative because light absorption, scattering, photon migration, and Raman responses can be linked to sugars, water status, acidity-related chemistry, and texture-associated structural changes without cutting the fruit. This critical review examines spectroscopic sensor-based machine-learning frameworks for in situ evaluation of fruit internal quality, with emphasis on visible and near-infrared spectroscopy, mid-infrared methods, Raman spectroscopy, and hyperspectral imaging. Particular attention is given to how acquisition geometry, penetration depth, spectral specificity, validation design, and model complexity interact from classical chemometrics to deep and multimodal learning. Literature across 2000–2026 was prioritized when studies reported whole-fruit sensing, external validation, and operationally relevant performance metrics. To make cross-study interpretation more rigorous, reported R, R 2 , RMSEP/SEP, sample scale, validation type, and deployment context are synthesized in a standardized benchmarking framework. The synthesis shows that soluble solids and dry matter remain the most reliable targets, whereas firmness, acidity, and thick-rind commodities still present major robustness challenges. The review also highlights calibration transfer, cultivar effects, ambient variability, model compression, uncertainty control, data leakage, and domain shift as decisive translational issues. By integrating sensor physics, quantitative benchmarking, data-analysis strategy, and deployment constraints, the article identifies the research gaps that most strongly limit reliable field and packinghouse adoption.