Managing mosquito breeding grounds is a significant global public health issue. In this case, using drones in conjunction with automatic image detection is a viable tool. However, the task faces significant challenges due to the varying altitudes at which drones operate, leading to substantial variations in object scale that hinder network optimization. Furthermore, high-speed flight and low-altitude manoeuvres introduce motion blur, making object distinction even more challenging. We propose an enhanced approach to tackle these issues by integrating a Transformer Prediction Head into the YOLOv7 framework (specifically, the Swin Transformer) using pre-trained weights from the VISDRONE dataset. We conducted extensive experiments using the Mosquito Breeding Grounds (MBG) Database, which was pre-processed and augmented. The proposed method demonstrates remarkable performance and interpretability in drone-captured scenarios. The average precision (AP) result of YOLOv7 across all classes is 88.4%, with particularly notable performance in the bucket and water tank classes, achieving a mAP@.5 of 96.9%. The results suggest that YOLOv7 with Transformer Prediction Head can be used to automate the detection of mosquito breeding grounds.