Adaptive Attention‐Enhanced Multiscale Vision Transformers for Real‐Time Leaf Disease Detection in Smart Agriculture
International Journal of Intelligent Systems · 2026-01 · 2026 (1)
DOI: 10.1155/int/5563178 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Intelligent Systems, online erschienen am 14.07.2026. Verfasst von Xuan Yang und 5 weiteren Autor:innen. Die Arbeit nennt 35 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Natural Science Foundation of Shandong Province, Weifang University of Science and Technology.
Abstract (Originalsprache)
Plant diseases represent a major challenge to global agricultural productivity and require precise and rapid diagnostic tools for effective intervention. This study presents an adaptive attention‐enhanced multiscale vision transformer (MS‐ViT) framework for plant disease classification. The proposed framework integrates RGB and multispectral imaging through multiscale feature extraction and adaptive attention‐based refinement to improve the discrimination of disease‐relevant patterns. It is further optimized for edge AI deployment to support real‐time inference in resource‐constrained agricultural environments. The proposed method was evaluated across six crop species and multiple disease categories, and it outperformed several competitive baseline models, including YOLOv7 and a hybrid CNN–ViT framework, in terms of accuracy, precision, recall, and F 1‐score. MS‐ViT achieved an overall accuracy of 97.8% with an average inference latency of 5.4 ms. To contextualize these results, the framework was evaluated on an integrated dataset comprising six crop species and 19 health/disease classes (82,560 images in total), using a 70:15:15 train/validation/test split and 5‐fold cross‐validation, while edge latency was measured on Raspberry Pi and NVIDIA Jetson Nano devices after INT8 post‐training quantization and structured pruning. In addition, ablation and robustness analyses confirmed the contribution of the adaptive attention‐based refinement mechanism, particularly under challenging conditions such as illumination variation and partial occlusion. Unlike conventional fusion approaches with fixed weighting, the proposed framework dynamically prioritizes the most informative spectral and visual features, thereby improving robustness and classification reliability across diverse agricultural scenarios. Overall, this work provides a unified and practical framework that bridges advanced AI‐based plant disease recognition with real‐time edge deployment for smart agriculture.
Förderung
- Natural Science Foundation of Shandong Province
- Weifang University of Science and Technology
Zitieren
APA
Yang, X., Wang, X., Li, D., Bai, Z., Li, Y., & Khosravi, M. (2026). Adaptive Attention‐Enhanced Multiscale Vision Transformers for Real‐Time Leaf Disease Detection in Smart Agriculture. International Journal of Intelligent Systems, 2026(1). https://doi.org/10.1155/int/5563178
Vancouver
Yang X, Wang X, Li D, Bai Z, Li Y, Khosravi M. Adaptive Attention‐Enhanced Multiscale Vision Transformers for Real‐Time Leaf Disease Detection in Smart Agriculture. International Journal of Intelligent Systems. 2026;2026(1). doi:10.1155/int/5563178
BibTeX
@article{Yang2026,
author = {Yang, Xuan and Wang, Xuewei and Li, Dejuan and Bai, Zhiqing and Li, Yujie and Khosravi, M.},
title = {Adaptive Attention‐Enhanced Multiscale Vision Transformers for Real‐Time Leaf Disease Detection in Smart Agriculture},
journal = {International Journal of Intelligent Systems},
year = {2026},
volume = {2026},
number = {1},
doi = {10.1155/int/5563178}
}
EUVIMED