Abstract
INTRODUCTION: Accurate identification of crop diseases is crucial for ensuring crop quality and yield. However, existing deep learning models for crop disease identification lack robustness in complex field environments and suffer from large model parameter sizes, which makes them difficult to deploy on resource-constrained devices. This gap between laboratory models and practical field applications necessitates the development of a lightweight and robust identification model. METHODS: To address these challenges, this paper proposes a lightweight YOLO-CGA model for sunflower disease identification and deploys it on a Raspberry Pi for field application. The model incorporates three key improvements based on YOLOv8n-cls: (1) A CBAM_ADown module is designed, which integrates attention mechanisms with asymmetric downsampling to enhance feature extraction and noise suppression in complex image backgrounds; (2) The C2f module of YOLOv8n-cls is replaced with the C3Ghost module, which utilizes ghost convolution to reduce parameter count while preserving fine-grained features; (3) An AFC_SPPF module is constructed, which aggregates multi-scale disease features through a multi-branch adaptive fusion structure to improve recognition performance for diverse lesions. RESULTS: Experimental results on three major datasets show that the proposed YOLO-CGA model achieves high identification accuracy: 98.48% on the BARI-Sunflower dataset, 98.32% on the Cotton Disease Dataset, and 91.11% on the FGVC8 dataset. Meanwhile, the model maintains a lightweight property with only 0.92M parameters, which is significantly fewer than that of other comparative models. DISCUSSION: The deployment of the YOLO-CGA model on the Raspberry Pi end device effectively bridges the gap between laboratory models and field applications, fulfilling the demand for real-time and on-site crop disease identification. The integration of attention mechanisms, ghost convolution, and multi-scale feature fusion enables the model to balance accuracy, robustness, and lightweight performance, making it suitable for resource-limited field scenarios.