Hybrid EfficientNet and Vision Transformer architecture for cross-platform agricultural pathology detection

Abstract

The early diagnosis of foliar diseases is crucial for food security; however, there is an applicability gap (reality gap) that limits the operational viability of Deep Learning models in real agricultural environments due to their high computational complexity and sensitivity to visual noise. The objective of this study is to design, validate, and deploy a lightweight, hybrid architecture (AgroScan) capable of operating as a real-time decision support tool against 49 diseases distributed across 10 high-impact crops. The methodology proposes a two-stage inference process: a binary filtering "gatekeeper" model to mitigate visual background noise, followed by a classifier trained on 98,000 images, which fuses the parametric efficiency of EfficientNetB0 with the spatial correlation capability of a Transformer module (Multi-Head Self-Attention). Trained and evaluated on 98,000 images. The hybrid architecture achieved an overall accuracy of 94.69% (95% CI: [94.28%, 95.11%]) and an F1-Score of 94.68% (95% CI: [94.22%, 95.08%]). Confidence intervals were calculated using non-parametric resampling on the test set. The main contribution of the study lies in the empirical demonstration that global attention increases accuracy in disease diagnosis while maintaining an inference latency (9.45 ms per image). Finally, the implementation of the architecture through a client-server platform, accessible via web and mobile applications, has demonstrated its operational robustness as a viable agricultural support tool.