Abstract: To address the problems of missed and false detections of surface targets by unmanned vessels operating in complex multi-scenario water environments, a surface object detection model based on improved YOLOv8s is proposed. First, the small-target detection layer is reconstructed by introducing low-level feature details to reduce the number of model parameters and improve the model's perception of small targets. Second, partial convolution (PConv) is introduced to replace the traditional Conv and construct the feature extraction module P-C2f, aiming to reduce redundant features and computation, which further compress the model size. Subsequently, the reparameterized generalized feature pyramid network (RepGFPN) is used to fuse features, aiming to enhance the full interaction and fusion of low-level detail information and high-level semantic information, thereby improving the multiscale target detection ability of the model. Finally, transfer learning is used to fine-tune the model and further improve detection performance. When tested on the WSODD dataset, the reference number of the improved model decreased by nearly 67.5% compared to that of the original model, whereas recall rate R increased by 4%, and mAP@0.5 increased by 2.1%, reaching 81.4%. Compared with other mainstream detection models, the improved model has obvious advantages and can help unmanned vehicles perform better surface detection tasks.