Explainable transformer with optimized attention for industrial oilseed groundnut leaf stress recognition

Groundnut (Arachis hypogaea L.) is a crucial industrial oilseed for oil-based value chains. Foliar diseases and nutrient stress reduce pod yield, weaken plant growth, and make harvested seed lots less uniform. These effects can hinder oil supply and processing stability. In many fields, diagnosis still relies on visual inspection, which is slow, labor-intensive, subjective, and difficult to scale across large farms. To address these limitations, this study proposes ResmECAMViTNet, a lightweight deep learning model that integrates residual convolutional neural networks (CNNs), a modified efficient channel attention (mECA) module, and a mobile vision transformer (MViT). The proposed model was trained on two publicly available groundnut leaf stress datasets: Groundnut Plant Leaf Images (Dataset 1) and the Groundnut Leaf Dataset (Dataset 2), each containing 10,361 and 1720 images, respectively. The model accurately classifies six stress categories in Dataset 1 and five stress categories in Dataset 2. ResmECAMViTNet achieved 99.50% accuracy, 99.50% F1-score, and 0.999 specificity on Dataset 1, while on Dataset 2 it achieved 99.42% accuracy, 99.61% F1-score, and 0.998 specificity. Compared to the baseline ECA module, the proposed mECA improves F1-score and accuracy by 6.03% and 6.5%, respectively. With only 2.06 million parameters and a 7.86 MB memory footprint, the model outperforms five classical machine learning models, seven deep CNNs, five attention-based variants, and five transformer-based architectures. Ablation studies confirm the efficacy of mECA in enriching spatial and channel-wise feature representation. Integrated XAI methods (Grad-CAM, LIME) further elucidate model decisions. ResmECAMViTNet offers an accurate, efficient, and interpretable solution for real-time deployment on resource-constrained devices, including UAVs and smartphones. It enables early field-level stress monitoring for targeted crop management. This can help protect pod and seed biomass production, reduce oilseed feedstock variability, and support more predictable downstream oil processing and bio-based product development.