The precise monitoring and localization of industrial exhaust smoke emissions play a crucial role in environmental management. Existing methods encounter challenges like intricate detection environments, small-scale targets, and extensive model parameters. This study presents an advanced drone smoke detection model, YOLOv8n-BWG, building on YOLOv8. It introduces a novel BC2f structure into the backbone network, leveraging an adaptive query mechanism to minimize computational and storage demands while boosting feature extraction efficiency. Additionally, the study employs a dynamic sample allocation strategy to refine the loss function, enhancing the model's sensitivity to small targets. It also integrates a lightweight convolution, GSConv, in place of traditional convolution techniques. GSConv employs a channel grouping approach, streamlining model parameters without sacrificing accuracy. Results on a specialized dataset reveal that YOLOv8n-BWG outperforms YOLOv8n by increasing the mean Average Precision (mAP) by 4.2%, boosting recognition speed by 21.3% per second, and decreasing both the number of floating-point operations (FLOPs) by 28.9% and model size by 26.3%. Significantly, deploying YOLOv8n-BWG on drones yielded promising outcomes in smoke detection, offering innovative approaches and insights for effective smoke monitoring practices.