Abstract: In order to achieve rapid identification and accurate classification of tobacco diseases, directed at the challenges of complex diseases distribution in tobacco fields, varying degrees of leaf overlap and obstruction, a novel tobacco disease detection method based on improved YOLOv8s targeting at tobacco brown spots, tobacco mosaic virus, tobacco cucumber mosaic virus, tobacco weather fleck, and tobacco wildfire was proposed. The study shows: 1) Based on the YOLOv8s object detection framework, a multi-scale feature module (C2f_MSBlock) was constructed. This module reduced the number of model parameters while ensuring the detection accuracy through hierarchical feature fusion and heterogeneous convolution kernel selection, thereby achieving the light weighting of the model. 2) The introduction of SEAM (Spatially Enhanced Attention Module) attention mechanism combined with the depth-wise separable convolution and residual connections which compensated for the response loss in areas obscured by leaves, thereby effectively addressed the issue of disease detection omission caused by leaf overlap and occlusion. 3) Spatial Pyramid Pooling with Efficient Layer Aggregation Network (SPPELAN) that replaced the original structure, enhanced the model's capability to detect small target lesions in real tobacco field environments through multi-scale pooling and efficient feature aggregation. The results indicated that the mean average precision (mAP) of the improved model reaches 92.5%, representing a 9.2% increase compared to the original YOLOv8s model, while the model weight reduced by 20.4%. Compared to other mainstream models such as SSD, Faster R-CNN, RT-DETR, YOLOv5s, and YOLOv12s, the improved model demonstrated significant advantages in both the detection accuracy and model light weighting.