Multimodal Maple Plant Disease Detection Using EfficientNet and Transformer-Based Semantic Fusion

Abstract

This paper presents a novel multimodal deep learning framework for maple plant disease detection by integrating visual and semantic information. Traditional plant disease detection systems rely primarily on visual features extracted from leaf images, which often leads to misclassification in cases of visually similar disease symptoms. To address this limitation, the proposed approach combines EfficientNet-based convolutional neural networks for visual feature extraction with transformerbased language models, including BERT and FLAN-T5, for semantic feature encoding. A Multilayer Perceptron (MLP)-based fusion mechanism is employed to integrate visual and textual features, enabling effective cross-modal learning. The proposed model is evaluated on a balanced dataset of 2,000 maple leaf images and associated disease descriptions. Experimental results demonstrate that the multimodal framework achieves an accuracy of 94.8%, outperforming vision-only and text-only models by a significant margin. Ablation studies and comparative analysis confirm the effectiveness of multimodal fusion and transformerbased semantic encoding. The proposed framework provides a robust and scalable solution for intelligent plant disease detection and has potential applications in smart agriculture systems.

Code and data availability

The paper describes a multimodal maple disease detection framework evaluated on 2,000 maple leaf images with textual descriptions, but no blocks contain any data availability statement, public dataset link, repository URL, or code deposit. No paper-specific public asset is identified.

No evidence-backed public reproduction asset is currently recorded.