Non-destructive prediction and visualisation of multiple loquat quality indicators using hyperspectral imaging and machine learning

Various indicators are used for the post-harvest quality assessment and grading of loquats, including weight loss rate, firmness, pH, soluble solids content, carotenoids, total flavonoid content, and colour index. In this study, hyperspectral imaging (HSI) and machine learning (ML) were performed to non-destructively predict multiple loquat quality indicators and visualise their spatial distributions across the fruit surface. Optimal feature variables for predicting quality indicators were selected from the full spectrum (256 bands in range of 390–1030 nm) using four feature variable extraction methods, and three ML algorithms were used to construct multiple prediction models. The competitive adaptive reweighted sampling-multiple linear regression (CARS-MLR) model was optimal, exhibiting favourable residual prediction deviation (RPD) values of 3.6763, 2.1821, 2.5544, 2.3928, 2.3450, 2.8073, and 3.2616 when predicting the seven quality indicators. Finally, pseudo-colouring was employed to input pixel information from hyperspectral images into the CARS-MLR model to visualise the distribution and compare loquats of different quality grades. Our findings provide a theoretical basis for non-destructively determining loquat quality based on HSI and ML. • Non-destructive detection for multiple loquat quality indicators. • Used hyperspectral imaging (400–1000 nm) and machine learning models. • Feature variables selected with CARS, GA, and RF algorithms. • CARS-MLR model achieved best prediction performance for seven indicators. • Visualising the spatial distributions of seven quality indicators in loquats.