Explainable deep learning for grading of Edible Bird's Nest (EBN)

Edible Bird’s Nest (EBN), an agricultural product derived from the solidified saliva of swiftlets, presents a classification challenge due to the subtle differences between nest grades. Accurate grading is essential for ensuring product fair pricing and supporting farm-scale production efficiency. This study utilized deep learning–based computer vision techniques to improve the accuracy and consistency of EBN grading across eight categories: A-High, A-Wash, A-White, AB, B, B3, Tiao-A, and Tiao-B, advancing the theoretical application of computer vision in precision agriculture. A total of 4,326 images were used for model training and evaluation, comprising 1,442 original images and 2,884 augmented images. This dataset was utilized to train and evaluate five Convolutional Neural Network (CNN) models: EfficientNetV2-B0, EfficientNet-B7, ResNet-50, Inception-v3, and MobileNetV3. During training, the models achieved classification accuracies of 96% to 99%. When evaluated against the unseen testing dataset, all models attained a testing accuracy of 99%. MobileNetV3 achieved an overall accuracy of 99% across most grades; however, it recorded a slight decline for the A-Wash and A-White grades. To improve interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed, confirming that the models focused on essential nest features such as shape and impurity level, rather than background noise. Consequently, this study validates the robustness of the proposed deep learning framework, demonstrating that explainable AI can provide a scientifically rigorous and standardized methodology for EBN quality assessment. • Deep learning architectures analyzed for EBN quality grading. • Dataset expanded to 4,326 images via robust augmentation. • CNN models achieved 96–99% accuracy with high generalization. • Grad-CAM validates feature fidelity and morphological focus. • Establishes a robust framework for objective quality assessment.