Plant disease detection is vital for agricultural sustainability and food security. While Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have achieved high accuracy in this domain, CNNs often require millions of parameters and substantial computation. ViTs suffer from the quadratic time and space complexity of self-attention (SA), limiting their use on resource-constrained devices. Although SA is capable of modelling long-range dependencies when symptoms are dispersed, many plant diseases exhibit small, localized lesions or texture changes; therefore, Neighborhood Attention (NA) offers a more efficient and targeted alternative by focusing on nearby regions rather than the entire image.

This work proposes a custom Localized NA block implemented in TensorFlow/Keras that operates directly on CNN feature maps, bypassing patch embedding and transformer modules. A lightweight CNN is then developed by combining depth-wise separable convolutions with the proposed localized NA block. In addition, a 100-category plant disease dataset covering 16 crops is presented. The dataset is curated, class-balanced, and made publicly available to support reproducibility and encourage further research.

The proposed 9-layer CNN, with just 1.7M parameters and a size of 6.74 MB, achieved a favorable balance between accuracy, model size, and computational efficiency, compared with MobileNetV1, MobileNetV2, DenseNet121, InceptionV3, MobileViT-XXS, and EfficientViT-M0, achieving 98.97%± 0.33% accuracy on PlantVillage and 93.36%± 0.28% on the proposed dataset. The ablation study showed that the NA block improved test accuracy by approximately 2–3%, while Grad-CAM visualizations indicated more precise targeting of diseased areas in the leaf image.