Non-Destructive Prediction of Fruit Ripeness and Firmness Using Hyperspectral Imaging and Lightweight Machine Learning Models

A new study from researchers at arXiv (Konrad et al.) systematically evaluated 20 classical machine learning algorithms for non-destructive fruit quality assessment using hyperspectral imaging. The team tested models across five fruit species for simultaneous ripeness classification and firmness prediction, employing cross-validated experimental design with Bayesian hyperparameter optimization. Their results showed that tree-based ML models can outperform state-of-the-art deep learning approaches like Fruit-HSNet, while requiring significantly less computational resources.

The study's most impactful finding was that only three visible-range wavelengths are needed to recover over 94% of full-spectrum accuracy. This means low-cost multispectral sensors combined with lightweight ML models can serve as practical alternatives to expensive hyperspectral cameras and complex deep learning systems for fruit quality sorting. The researchers also noted that data preprocessing strategies—particularly class balancing and spectral transformations—contribute as much to prediction accuracy as algorithm choice, making these methods more accessible for real-world agricultural settings without GPU resources or large-scale training data.