Enhanced Deep Recurrent Optical Flow With Efficient Feature Encoding and Channel Attention for River Surface Velocimetry

Accurate estimation of river-surface flow velocity is critical for water-resources management, flood early-warning systems, and ecological modeling. Although unmanned aerial vehicle (UAV) image velocimetry offers high efficiency and noncontact operation, its robustness and accuracy degrade under complex conditions-such as illumination variability, low-texture surfaces, and nonrigid water-surface motion. In particular, detecting subtle surface displacements in low-texture regions remains challenging, thereby affecting flow-field reconstruction. To address these issues, we propose RSV-RAFT, a deep optical-flow framework for river-surface velocity assessment. First, we enlarge the receptive field of the feature-extraction module and introduce the Leaky ReLU activation to enhance sensitivity to weak textures and small-scale motion. Second, we design a Decoupled Enhanced Channel-Attention Flow Head (DECA-Flow Head) based on a spatial-channel decoupling architecture to improve modeling of minute displacements and boundary motion. Finally, we present a Deformable Coordinate-Aware Upsampling Mask Generator (DCAU-MaskGen) that integrates deformable convolutions with coordinate attention to enhance boundary perception and structural reconstruction. Experiments on multiple optical-flow benchmarks and real UAV river imagery show that RSV-RAFT outperforms existing methods in both accuracy and efficiency while maintaining a mean absolute percentage error (MAPE) of 10.5%-15.5%, demonstrating its potential for engineering applications.