Hyperspectral imaging combined with a universal hybrid deep network for identifying early chilling injury in kiwifruit across varieties

Constructing an accurate model is crucial for applying hyperspectral imaging (HSI) to identify early diseases in agricultural products such as fruits. Chilling injury is a physiological disease of kiwifruit that is challenging to identify by the naked eye before it reaches a severe stage. However, different varieties of fruit own different physicochemical properties and spectral characteristics, causing that the model established for one variety is not applicable to another, while refactoring the model requires a significant amount of time and effort. Given this, the feasibility of applying HSI for identifying early chilling injury in kiwifruit across varieties was investigated. A universal hybrid deep network of CNN-DotGRU-SelfAttention (CDGSA-Net) was developed for feature extraction, feature diversity capture, and feature enhancement to establish the identification model with high accuracy and generalization. Ultimately, the pelican optimization algorithm (POA) was proposed to optimize hyperparameters of CDGSA-Net to improve the efficiency. The performance of the POA-CDGSA-Net model was superior to other machine learning (ML) and deep learning (DL) models, yielding results of 99.17 %, 99.17 %, 99.26 %, 99.79 %, and 99.20 % in terms of acc, prec, rec, spec, and F1 on the test set, respectively. Therefore, HSI coupled with POA-CDGSA-Net offers a viable approach for identifying early chilling injury in kiwifruit across varieties.