A deep straw network for strawberry leaf disease detection

Timely and accurate identification of leaf diseases in strawberries is essential for improving crop health, reducing chemical usage, and ensuring the efficiency of horticultural supply systems. This study introduces the Deep Straw Network (DSN), a novel hybrid deep learning architecture that integrates an Image Guided Module (IGM) and a Patch Guided Module (PGM) for precise detection of eight major strawberry leaf diseases. The DSN architecture combines local spatial feature extraction with global contextual modeling, enhancing robustness under natural variabilities such as lighting, occlusion, and background clutter. To support real-world deployment, we present KUStraw, a newly curated dataset comprising 2,464 annotated images captured in greenhouse environments in the United Arab Emirates. When compared with existing methods, the performance of DSN was rigorously assessed using the KUStraw dataset in combination with three publicly accessible datasets: PlantVillage, PlantDoc, and SDDD, DSN attained an average Top-1 accuracy of 96.25%, highlighting its performance. Grad-CAM visualizations further enhance the interpretability of model predictions. Comprehensive ablation studies confirm the contribution of hybrid design choices and patch-based processing. The proposed DSN framework offers a scalable and data-efficient solution for precision disease management, making a significant contribution to the advancement of sustainable practices in strawberry agriculture and enhancing the efficiency of the horticultural supply chain management. • Developed a novel Deep Straw Network combining guided local and global modules. • Created a labeled dataset of 2,464 images from operational UAE greenhouses. • DSN achieved an average Top-1 accuracy of up to 96.25%. • Performed detailed ablation and validation to verify the proposed model's efficacy.