Water surface object detection in river channels based on feature cross-layer fusion and reconstruction

Water surface object detection in river channels is essential for effective river monitoring systems. However, existing object detection techniques are frequently inadequate for perceiving objects of varying sizes in complex and varying backgrounds. To address this issue, firstly, we propose a Feature Cross-Layer Fusion and Reconstruction Module, which effectively fuses multi-scale features through adaptive weights (used to dynamically adjust the importance of features at different layers) and employs a spatial-channel reconstruction mechanism (by separately learning and reconstructing spatial and channel dimension features) to reduce background feature redundancy, achieving a 5.1% improvement in Precision over the baseline model. Furthermore, we introduce a Feature Extraction Module based on structural reparameterization, which enhances the feature representation capability while maintaining computational efficiency, resulting in a 1% improvement in mAP @0.5 compared to the baseline. Building on these improvements, we develop a water surface object detection algorithm that incorporates an improved loss function for better accuracy. To comprehensively evaluate its performance, we constructed a dedicated UARODD dataset, which includes 16 categories of commonly observed water surface objects in river channels. The dataset consists of 9500 images collected from real aerial photography and the internet, with a total of 24534 annotated instances, covering a wide range of river scenes worldwide. Experimental results indicate that the proposed algorithm achieves a mean Average Precision (mAP@ 0.5) of 79.6% on this real-world dataset, representing a 5% improvement in mAP@ 0.5 compared to the YOLOv11 baseline. The detailed program code and weight files have been made publicly available at ht tps://github.com/zhangenze1016/FCFR-yolo.