Updating apple Vis-NIR spectral ripeness classification model based on deep learning and multi-seasonal database

Judicious assessment of ripeness is crucial for ensuring the quality and commercial value of apples. However, when it comes to detecting apples spectrally under different seasonal variations, there are limitations in the application of calibration models that are built for a single season. Therefore, it is necessary to implement model updating. In this study, a large dataset was acquired of apple visible and near-infrared spectra spanning four seasons and assessed the ripeness of the samples based on computer vision tools. After completing a series of data processing and parameter optimisation, a one-dimensional convolution neural network was built on the initial seasonal dataset. Subsequently, model transfer between seasons was completed using deep transfer learning. Further, multi-seasonal model updating of apple ripeness classification models was achieved in two scenarios with and without historical data. The results indicated that by retraining the network’s convolution layer, the classification accuracies for the three new seasons improved by 4%, 18%, and 15% respectively, while remaining stable for the original season. Combining 5%–20% new season samples with cumulative historical data, the model’s classification performance improves by up to 54% and 55% on the two new seasons. This study contributes to the updating of the multi-seasonal spectral database model for fruit quality control.