JRViT : a multi-scale feature fusion framework integrating convolution encoder and vision transformer for robust jackfruit leaf health and growth stage prediction

Abstract

This research introduces a robust cutting-edge ResNet50 with Vision Transformer (ViT) hybrid model for jackfruit leaf health and growth stage classification: which is the first hybrid model addressing both jackfruit leaf health and growth-stage classification jointly. The hybrid model is based on the advantages of Convolutional Neural Networks (CNNs) and Transformer-based networks, which integrates ResNet50's local in-depth feature extraction and ViT's global context and longrange relationship learning abilities, offering a more stronger and more robust leaf classification. The novelty of this research lies in introducing a unified multi-scale feature fusion framework built upon a newly curated real-field jackfruit leaf dataset that jointly addresses health and growth-stage prediction through a hybrid CNN-Transformer representation model, a combination that has not been covered in jackfruit leaf research previously. Our Experimental results shows that the proposed hybrid model outperforms other recent typical models like Xception, VGG16, and ResNet50 with a 98.67% accuracy rate in several leaf classes, (i.e. dried, healthy, leaf Miner, senescence, and young). The model also exhibits high precision, recall, and F1-score, as well as high AUC values when checked through ROC curve analysis, verifying its performance in real agricultural practice. Proposed hybrid approach provides a strong platform for plant health observation and pre-emergence growth stage prediction, with immense potential for scalable AI based agricultural solutions.