基于改进SAS-YOLOv8的水稻病害检测

Rice Disease Detection Based on Improved SAS-YOLOv8

摘要: 水稻病害的快速准确检测对保障粮食安全和实现智慧农业具有重要意义。本文提出一种基于改进YOLOv8n的SAS-YOLOv8n模型。首先引入SE注意力增强关键病害特征的表达。其次使用轻量化自适应空间特征融合模块ASFFLite替代原有独立特征精炼层，在提升多尺度融合能力的同时显著降低参数量。最终采用SPDConv (空间–深度卷积)替代部分步长卷积，以无损下采样方式保留小目标的细粒度空间信息。在水稻病害数据集上的实验结果与基线对比，所提出的SAS-YOLOv8n模型的mAP@50，mAP@50:95和R分别提升1.9%，1.1%和2%。实现了检测精度的有效提升，在水稻病害的田间实时检测中具有良好的可靠性。

Abstract: Rapid and accurate detection of rice diseases is of great significance for ensuring food security and realizing smart agriculture. This paper proposes an improved YOLOv8n-based model, termed SAS-YOLOv8n. First, SE attention is introduced to enhance the expression of critical disease features. Second, a lightweight adaptive spatial feature fusion module, ASFFLite, is adopted to replace the original independent feature refinement layer, which improves multi-scale fusion capability while significantly reducing the number of parameters. Finally, SPDConv (Spatial-Depth Convolution) is employed to replace certain strided convolutions, preserving fine-grained spatial information of small targets through lossless downsampling. Experimental results on a rice disease dataset show that, compared with the baseline, the proposed SAS-YOLOv8n model improves mAP@50, mAP@50:95, and Recall by 1.9%, 1.1%, and 2%, respectively, achieving effective enhancement in detection accuracy and demonstrating good reliability.