Eff-swin-hgso: attention-driven and optimized plant leaf disease diagnosis using efficientNetV2B0 and swin transformer with HGSO feature selection

Abstract

Abstract Agriculture is essential to human civilization, providing food and raw materials. Plant diseases significantly threaten agricultural productivity, making early and accurate detection essential. Despite Recent advances of deep learning in making automatic plant leaf diseases diagnosis systems, some of them depend on simple features fusion methods and lack an efficient method to exploit complementary information. Therefore, this paper proposes a system for diagnosing plant leaf diseases by fusing two powerful deep learning models: EfficientNetV2B0 and Swin Transformer via attention-based feature fusion that adaptively weights each model’s contribution with features. These models extract complementary features: EfficientNetV2B0 extracts fine-grained local features, and the Swin Transformer extracts global contextual information, producing highly and complementary expressive fused features. The high-dimensional fused features demand High-Performance Computing (HPC) resources for efficient parallel processing and accelerated training. Moreover, the Henry Gases Solubility Optimization (HGSO) metaheuristic is applied to select the most discriminative and related features. Unlike previous methods that diagnose diseases affecting only one plant, the proposed approach handles multiple plant species simultaneously, further increasing computational demand. Finally, RBF-kernel SVM is applied for a classification step. The system was implemented on a GPU-based high-performance computing environment using CUDA acceleration to enhance computational efficiency. Experimental evaluation on the PlantVillage benchmark dataset with seven classes achieved a high classification accuracy of 99.2%, outperforming other state-of-the-art methods. These results enhance the model’s practical capability for application in real-world agricultural decision-support systems.

Plant phenotyping relevance

葉画像から植物病害を診断する深層学習システムの開発とベンチマーク評価が研究の中心であり、植物の病害状態を直接推定しているため。

Code and data availability

The paper's sole experimental input is the PlantVillage leaf-disease image dataset, which the authors explicitly state is publicly available online with a Kaggle URL in the Data availability statement. No author code, models, or other paper-specific assets are disclosed.