Enhanced wheat crop leaf disease classification using multi-level contrast enhancement and modified vision transformers.

Abstract

The integration of advanced tools and techniques has significantly boosted agricultural productivity. Wheat crops, which are vital for global food security, are often susceptible to various bacterial and viral diseases, considerably impacting both yield and quality. Efficient disease detection is crucial for effective treatment and yield optimization. This study presents an innovative approach that combines a multi-level contrast enhancement framework with a novel transformer-based architecture for the rapid and precise diagnosis of wheat crop leaf diseases. By employing contrast enhancement techniques, we enhance the visual quality of wheat crop images, facilitating improved feature extraction. Incorporating Vision Transformers (ViTs) enhances computational efficiency, enables multi-scale feature extraction, and reduces dimensionality. We implement and evaluate our proposed models on two publicly available wheat datasets, utilizing three variants of ViT: a modified ViT with seven-block transformers, a pre-trained ViT-16-Tiny, and a modified ViT with seven transformer blocks and skip connections. Our results demonstrate classification accuracies of 98.90% for the modified seven-block ViT, 97.50% for the ViT-16-Tiny model, and 97.90% for the ViT seven-block with skip connections. A comparative analysis with state-of-the-art techniques reveals that our proposed techniques outperform existing methods in terms of accuracy, precision, sensitivity, False Negative Rate, and the total number of learnable parameters.

Plant phenotyping relevance

コムギ葉の病徴を画像から分類する画像処理・Vision Transformer手法の開発と評価が中心であり、植物の病害状態を直接推定するため、方法論文として含める。

Code and data availability

The paper uses public Kaggle wheat/cotton leaf image datasets (cited as refs 40–43), but the supplied blocks contain no authors' public URLs, code deposit, or availability statements for their models, scripts, or enhanced data. The datasets are cited prior public resources, not paper-specific deposited assets.

No evidence-backed public reproduction asset is currently recorded.