Enhanced River Connectivity Assessment Across Larger Areas Through Deep Learning With Dam Detection

Abstract

Monitoring river connectivity across large regions is essential for understanding hydrological processes and environmental management. However, comprehensive assessments of river connectivity are often hindered by inaccurate dam databases, which are biased towards larger dams while overlooking smaller or low-head dams. To enhance the accuracy of river connectivity assessments, we developed three advanced convolutional neural networks (CNNs; YOLOv5, Advance-You Only Look Once [YOLO], and Faster R-CNN) to accurately classify dams and evaluate river connectivity using high-resolution (1 m) remote sensing imagery. The evaluation results showed that Advance-YOLO performs best with an average mean average precision (mAP) of 86.6%, while Faster R-CNN performs mediocrely with an average mAP of 77.9%. Applying the well-trained model in the Tarim River Basin (China), one of the largest inland river basins around the globe, we found that there are currently 135 dams in total on the Tarim River and its sources. Conversely, the existing public dam database underestimates 85.9% of the dams. Notably, we found a 14.3% decline in river connectivity of the Tarim River over the past decade, and the current dam density of the Tarim River and its four source rivers is 1.12 per 10 000 km². However, the existing public dam database overestimated river connectivity by 83.9%. The model developed here enhances river connectivity assessment across larger areas over a long period, thereby fostering more advanced research on hydrological processes and effective water resource management.