REVIEW PAPER TOPIC - SMART SENSING FOR NON DESTRUCTIVE FRUIT AND VEGETABLE ASSESSMENT: A COMPREHENSIVE REVIEW OF NIR SPECTROSCOPY, MACHINE VISION AND YIELD-QUALITY PREDICTION UNDER A CHANGING CLIMATE

Abstract

Rapid, non-destructive, quantitative evaluation of quality and yield of fresh produce has become a core activity in horticultural production, postharvest handling, and supply-chain management. This is increasingly important as climate variability weakens the relationships between external appearance, internal content, and final performance of produce each year. This review compiles the present knowledge on smart sensing technologies for fruits and vegetables, focusing on near-infrared (NIR) and visible/NIR spectroscopy, hyper-spectral and multi-spectral imaging, machine vision coupled with deep-learning-based grading, and other related technologies including electronic nose, fluorescence imaging, X-ray computed tomography and acoustic sensing. It also reviews the remotely and proximally sensed instruments for estimating yields, including UAVs, satellite vegetation indices, wireless sensor networks, and the Internet of Things.  The discussion delves to some extent into the physics, instrumentation and chemometric or machine-learning pipelines preprocessing, wavelength selection, partial least squares regression, support vector machines, random forests, and convolution neural network architectures (e.g., YOLO, Mask R-CNN) all turning raw spectral, image or sensor data into quantitative predictions of soluble solids content, titratable acidity, firmness, dry matter, pigmentation, internal disorders and fruit count. Case studies of apple, grape, citrus, mango, blueberry, tomato, potato and many other important crops are compared to demonstrate the current level of maturity and the limitations of each technology.  Ambient temperature, high CO2, drought and UV/ozone stress modify biochemical composition of the fruit and canopy reflectance, so we focus on the influence of these changing environmental factors on sensor calibration stability, model transferability and the potential to predict yield quality reliability. Corresponding methodologies to keep the prediction performance stable under changing climatic baselines calibration transfer, domain adaptation and multi-sensor data fusion are reviewed as well. The review ends by describing the challenges that still need to be: transfer of calibration models between cultivars, seasons and instruments; cost and computational complexity of hyper- spectral and deep learning pipelines; and lack of standardised climate- resilient validation protocols. It concludes by suggesting avenues for running smart sensing in precision horticulture, climate-adaptive breeding and supply-chain decision making.