In the field of UAV smoke detection, due to the significant variations in different detection scenes, existing smoke detection algorithms often suffer from issues such as low detection accuracy and slow speed. To address these issues, in this paper we constructed a new UAV smoke dataset (USD) in multiple scenes, and proposed an improved YOLOx UAV smoke detection algorithm in multiple scenes. Firstly, we introduced an improved attention module into the YOLOx network to improve the extraction process of channel features and spatial features respectively, which can extract more representational smoke features. Then, we presented a two-way fusion network to enhance the fusion ability of multi-scale feature fusion module for small smoke target features. Finally, we utilized a Focal-EIOU loss function to address the issues such as the imbalance of positive and negative samples in the training process, and the distance and coincidence degree of two frames cannot be reflected when the prediction frame and real frame do not intersect. Experimental results show that the proposed algorithm has good robustness when applied to UAV smoke detection tasks in multiple scenarios. Compared with several classical smoke detection algorithms, the accuracy of the proposed smoke detection method on different data sets has been improved respectively. For instance, compared with the original YOLOx-s model, the accuracy was improved by 2.7%, the recall rate was improved by 3%, and the speed reached 73.6 frames per second.