In order to improve the recognition accuracy of Unexploded Ordnance (UXO) targets on the ground by UAVs in complex environments,a UXO target detection method based on the improved YOLOv5 is proposed.On the basis of YOLOv5,the method improves the loss function of the original YOLOv5 network to improve the recognition accuracy of UXO targets.At the same time,the method adds an attention mechanism,improves mosaic data enhancement,and improves the prediction frame screening mechanism to improve the recognition efficiency of UXO targets,and realizes the detection of UXO targets in air-to-ground scenarios with better accuracy and speed.Experimentally,multiple UXO datasets in different complex backgrounds are selected,labeled and trained to obtain UXO target models.Then,the correctness of the algorithm and model is evaluated from the perspectives of model training results and target detection results.The experimental results show that:1) The model obtained by NGG-YOLOv5 has a significant improvement in detection accuracy and detection speed in comparison with that obtained by the original YOLOv5,with an increase in accuracy from 78% to 91% and an increase in mean Average Precision (mAP) from 50% to 56%;and 2) It can effectively detect UXO targets in four kinds of complex backgrounds,with a low missed alarm rate.