YOLO-MRAF: A multi-scale reinforced attention fusion network for waterbody segmentation

Abstract

Accurate waterbody segmentation is important for environmental monitoring, water surface analysis, and water resource management. However, waterbody extraction from complex remote sensing images remains challenging due to irregular water morphology, shadows, illumination variations, reflection interference, and visually similar backgrounds. In addition to pixel-level mask extraction, practical waterbody monitoring also requires region-level localization and prediction confidence for separated water areas. To address these requirements, this paper proposes YOLO-MRAF, a YOLO-based instance-aware waterbody segmentation framework. The proposed model integrates three modules: a Multi-Scale Reinforced Feature (MSRF) module for preserving spatial details during hierarchical feature extraction, a Direction-Sensitive Attention Fusion (DSA-Fuse) module for enhancing horizontal and vertical spatial dependencies, and a Feature Recalibration Module (FRM) for strengthening feature interaction through additive, multiplicative, and transposed-convolution coupling. Experiments on the Water Detection & Segmentation Dataset under three random seeds show that YOLO-MRAF improves mask-level mAP50 and mAP50-95 from 0.594 and 0.398 to 0.644 and 0.448, respectively, compared with YOLOv11n-seg. The results also reveal a Precision–Recall trade-off, as YOLO-MRAF improves mask-level Precision but shows lower box-level Recall. Comparisons with semantic segmentation models and direct external evaluations on RIWA and LuFI-RiverSnap further demonstrate its competitive pixel-level performance and improved external mIoU and Dice. These results indicate that YOLO-MRAF is a practical YOLO-based solution for instance-aware waterbody segmentation.