Lightweight Deep Learning for Real-Time River Surface Velocity Estimation From Low-Altitude Drone Remote Sensing

Drone-based low-altitude remote sensing has become an important means for hydrological monitoring, where real-time on-device perception of both waterbody extent and surface flow velocity is urgently required. However, existing algorithms struggle to simultaneously achieve high monitoring accuracy and low inference latency, due to motion-feature interference induced by complex field environments and the computational bottleneck on embedded devices. To address these challenges, we propose AquaEdge, a lightweight deep learning framework for remote sensing inversion of river surface flow velocity. AquaEdge integrates a segmentation branch built upon a Frequency-augmented Star (FAStar) backbone and a lightweight feature extraction module, Cross Stage Partial with Partial Multi-Dilated Depthwise Convolution and Channel Attention (CSPPC-MDCA). The segmentation branch generates a water-mask prior that suppresses non-water spurious motion and specular-reflection noise propagation. This water-mask prior is then used as a constraint in a lightweight optical flow model, NeuFlow-lite, enabling water-surface motion estimation and velocity inversion under constrained computational resources. Experimental results show that AquaEdge achieves a mean Average Precision (mAP@50–95) of 97% for waterbody segmentation. In field tests across multiple cross-sections in the Heihe River Basin, AquaEdge achieves a mean absolute percentage error (MAPE) of 13.6%–14.2% relative to a propeller current meter and 12.0%–13.9% relative to a handheld radar velocimeter for surface flow velocity. Furthermore, AquaEdge delivers an end-to-end inference latency of approximately 200 ms on an NVIDIA Jetson Orin NX, validating its feasibility for real-time integrated hydrological sensing in real-world natural river scenarios. This study establishes an on device deployable pathway for low altitude remote sensing based surface velocity retrieval and provides ground based validation and algorithmic references for hydrodynamic inversion using remote sensing platforms with high temporal resolution.