Detection of Early Subtle Bruising in Strawberries Using VNIR Hyperspectral Imaging and Deep Learning

Abstract

Detecting early surface bruising in strawberries during postharvest storage is crucial for maintaining product quality and reducing waste. In this paper, we combined visible-near infrared hyperspectral imaging (VNIR-HSI) technology with deep learning methods to efficiently detect early surface bruising in strawberries. Specifically, we created a hyperspectral image dataset of strawberries, captured in the 454–998 nm wavelength range at five intervals: 1, 12, 24, 36, and 48 hours after applying four levels of bruising: none, slight, moderate, and severe. To address the challenges of a limited sample size and redundant hyperspectral data, we employed data augmentation and two feature wavelength extraction techniques: Uninformative Variable Elimination (UVE) and Competitive Adaptive Reweighted Sampling (CARS). We then developed several classification models, including SVM, CNN, CNN-LSTM, and CNN-BiLSTM. Experimental results showed that the CNN-BiLSTM model, which used feature wavelengths selected by CARS, achieved a 97.8 % classification accuracy for detecting slight bruising 12 hours post-treatment, with an average bruised area of 24.09 ± 6.38 mm². This performance surpassed the SVM, CNN, and CNN-LSTM models by 14.7, 10.5, and 4.5 percentage points, respectively. This study effectively classified early bruising in strawberries and visualized bruised areas, demonstrating significant improvements in detection and classification of early bruising, particularly for smaller areas.