Performance comparison among different models.

In response to the urgent need for water environment protection, this study proposes an improved algorithm for detecting floating objects on the surface of water: You Only Look Once version 8- Small Surface Targets (YOLOv8-SST). This algorithm aims to address the impacts of illumination variations and water surface distortion on floating object detection, as well as missed and false small object detections in complex aquatic scenarios. First, to mitigate the noise interference introduced during the downsampling process of the backbone network in complex aquatic environments, a C2fBF (C2f-BiFormer) module, based on the BiFormer dual-layer routing attention mechanism, was developed. This module effectively preserves fine-grained contextual feature information during feature extraction. Then, the conventional loss function was replaced with a more effective Inner-Complete Intersection over Union (Inner-CIoU) loss under auxiliary bounding boxes, allowing the model to adjust the size of auxiliary boxes more flexibly during detection and thereby improving detection accuracy. Finally, the adaptive moment estimation (Adam) optimizer in the original algorithm was replaced with the second-order clipped stochastic optimization (Sophia) optimizer to improve the generalizability of the model. On a combined dataset integrating FloW-Img, WSODD, and our self-collected data, YOLOv8-SST outperformed the baseline YOLOv8n, achieving a 3.1% increase in mean average precision (mAP)@0.5 and a 5.0% increase in mAP@0.5:0.95. These results demonstrate the effectiveness and robustness of the proposed method for small object detection in challenging natural water environments.