Towards Sustainable Food Systems: Multi-Channel Capsule Neural Network for Fruit Freshness and Spoilage Timeline Analysis

Food waste is a major contributor to global greenhouse gas emissions, water overuse, and ecological degradation, making its reduction a critical priority in the fight against climate change. This study presents a lightweight deep learning-based multi-channel classification framework using Capsule Neural Networks (CapsNet) to detect fruit freshness and predict spoilage timelines, with bananas as a case study. The proposed model consists of two parallel branches: the first classifies banana images into six ripeness categories, namely: fresh ripe, fresh unripe, ripe, overripe, unripe, and rotten. The second branch predicts the number of days until spoilage. The model was evaluated on an open-source dataset consisting of over 7000 banana images and their corresponding spoilage timelines, and it achieved a classification accuracy of $91.31 \%$, an F1-score of $\mathbf{9 1. 2 8 \%}$, and an AUC of $\mathbf{9 9. 1 4 \%}$, alongside a low mean absolute error (MAE) of 0.0981 days for spoilage prediction. These results highlight the model’s strong generalization performance and reliability. The integration of classification and forecasting within a single architecture provides actionable insights for optimizing supply chains, minimizing waste, and informing consumer choices. This research demonstrates the potential of AI-powered approaches to support climate-resilient food systems by reducing wasterelated methane emissions and resource inefficiencies at the point of consumption and distribution.