Existing detection algorithms for underwater object in images of unmanned aerial vehicles present challenges such as suboptimal accuracy for small objects and enormous computational demands. This work proposes a lightweight detection network with an enhanced pyramid for small objects based on YOLOv11, named SOEP-YOLO. To tackle the small object detection challenge, a feature enhancement network is proposed, which improves the processing capability of small target features. Instead of the approach of adding a P2 detection layer, the P2 layer feature map is processed through space-to-depth convolution to obtain enriched features containing small target information, which reduces the number of the model parameters and inference time. To enhance detection efficiency and reduce model complexity, the cross-stage partial concept is adopt to optimize the omni-kernel module, resulting in a lightweight module for feature integration. This module is designed to effectively learn feature representations from global to local scales, ultimately improving the detection accuracy of small objects. The experimental results on the SeaDroneSee dataset indicate that SOEP-YOLO enhances the mean average precision by 3.03% over the original YOLOv11s, while reducing the number of parameters by 47.9%.