Hybrid loss for robust ripeness classification under noisy labels

The reliability of supervised machine learning in agricultural classification is often undermined by noisy labels, which arise from subjective human annotations and environmental variability. This study introduces a noise-robust framework that integrates dual-network architectures to create a novel Hybrid Loss function. The Hybrid Loss integrates cross-entropy with consistency regularization to tackle the problem of noisy labels. The Modified Hybrid Loss further enhances robustness for small-scale datasets by updating dual networks independently. We validated these approaches on an avocado microwave image dataset and photographic avocado dataset. Across noise rates of 10%, 20%, and 50%, the proposed methods consistently outperformed baseline loss functions, including Cross-Entropy, Co-Teaching, and JoCoR. ANOVA statistical analysis confirmed significant improvements (p ' 0.00001) in classification accuracy. Beyond benchmark validation, we demonstrated practical application by tracking avocado ripening profiles over time with microwave imaging. These results establish Hybrid and Modified Hybrid Loss as effective strategies for mitigating label noise in agricultural machine learning, advancing the development of reliable, scalable, and non-destructive fruit ripeness assessment systems in Agricultural artificial intelligence (AI).