PriBeL-Net: Extending betel leaf dataset with CNN-based image classification

Deep learning is core to precision agriculture. In this research, the authors compare four deep-learning frameworks: MobileNetV2, EfficientNetB0, ResNet50V2, and DenseNet121, on a custom dataset under controlled and on-field environments. Whereas MobileNetV2 and ResNet50V2 gave the best results on the controlled setup, robust to light variations, backgrounds, and different leaf orientations, DenseNet121 gave better results in field environments with good accuracy and F1-score. However, EfficientNetB0 was not up to the mark, implying the restrictions of light-weight models while working with noisy, real-world datasets. With such implications, DenseNet121 is reported to be the most dependable candidate for application in agriculture. In the next phase, adaptation and rearrangement of DenseNet121's architecture and parameters will be taken up to better its performance and its ability for adaptation under diverse agricultural conditions.