Surface litter target detection plays an important role in timely detection and early warning of river pollution, providing strong support for the realisation of round-the-clock uninterrupted monitoring and providing data support for river litter clean-up. River rubbish has the characteristics of small target and easy to cover, which brings difficulties for target detection. In order to solve the above problems and improve the detection accuracy, the river trash detection algorithm based on improved YOLOv9 is proposed. Aiming at the problem of small targets of water surface rubbish, a multi-scale spatial pyramid module is introduced to extract finer granularity rubbish features; aiming at the problem that river rubbish will have mutual occlusion in the actual river environment, the SEAM module is added to highlight the rubbish region information in the image and weaken the background region information; the shapeIoU replaces the CIoU in the original model to reduce the rubbish bounding box regression of the small targets The IoU value of the sample is affected by the shape of the manually labelled box to further improve the detection accuracy. Experimental validation is conducted on the self-constructed visible water surface litter dataset DSGD and infrared water surface litter dataset ISGD as well as the dataset FLOW-IMG, which is publicly available for inland river floating object detection, respectively, and the results show that the detection accuracies reach 85.8%, 71.1%, and 86.8%, which are 4.4%, 3.5%, and 0.8% higher than those of the original YOLOv9 model, respectively. Compared with the existing mainstream target detection models, SSM-YOLOv9 all show better water surface trash recognition and meet the real-time requirements.