Floating object detection on water surfaces is crucial for waterway security and ecological protection. Current systems face significant performance degradation in adverse weather conditions due to precipitation noise and low visibility. To address this, we propose an enhanced YOLOv12n framework incorporating three key improvements: a synthetic weather-robust dataset for training, an edge-preserving spatial-channel convolution module for better feature extraction, and a frequency-domain attention mechanism for noise suppression. Our experiments on a self-built dataset demonstrate the effectiveness of these enhancements, achieving 87.1% mAP@ 0.5 while maintaining real-time performance at 476 FPS. The results show reliable detection capability across various weather conditions, making the system practical for real-world monitoring applications where weather resistance is essential.