Comparative analysis of CNN-based approaches for flower classification
Abstract
For different agricultural tasks such as robotic pollination, vision systems must perform a dual classification to determine both the flower species and their corresponding bloom stage to ensure suitable and directed pollen delivery. In this study, we evaluate multiple deep learning models, including VGG16, ResNet50, EfficientNetV2, MobileNetV3, and an adapted YOLOv10 backbone, where the YOLO architecture is repurposed as a feature extractor. These models are tested under a dual-output image classification framework where each model simultaneously predicts flower species and four bloom stages (bud, partially bloom, fully bloom, and wilt) from a single input image. The models were trained and evaluated on a floral dataset of roses and sunflowers with four different bloom stages. Bloom-stage classification proved more challenging due to the subtle visual differences between adjacent stages. All the models performed relatively well in terms of flower species classification. However, VGG16 achieved the highest performance with 82.8% accuracy for bloom-stage prediction whereas ResNet50 showed stable and well-balanced performance across both tasks. Meanwhile as a lightweight alternative, MobileNetV3 provided a moderate accuracy. In case of EfficientNetV2 and YOLOv10, they showed reduced performance in bloom-stage prediction. These findings highlight how these different models perform in the classification tasks, which gives us potential architectures for agricultural vision systems.
Authors: Samarth Agrawal, RituparnaDatta, Abhishek Rudra Pal, Claudio Angione, K. Karunamurthy
Published in: Smart Agricultural Technology (2026)