Deep Learning based Surface Velocity Estimation in Contaminated Rivers: A Case Study of Varuna River

Accurate river discharge measurement is essential, yet UAV-assisted methods often face challenges such as stability corrections, obstructed views, and algae-rich contamination. This paper presents a deep learning approach for surface velocity estimation, using the Varuna River as a case study. A custom drone with an NVIDIA Jetson Nano and GoPro Hero 7 captured a 2 -minute, $30-\mathrm{fps}$ video of an algae-contaminated section. A branched CNN model processes sequential image pairs to produce PIV outputs, using a filtering layer to manage contamination. Notably, certain algae sizes enhanced PIV analysis by acting as natural particles. This approach offers improved accuracy in complex river environments, integrating deep learning with traditional PIV for better water resource management.