Large woody debris (LWD), a key indicator of riparian vegetation disturbance and river corridor dynamic, plays a crucial role in habitat complexity, geomorphic dynamics and river management. Accurate mapping and monitoring of LWDs are therefore essential for river process analysis and ecosystem assessment, particularly in highly dynamic braided river systems. However, mapping and monitoring LWD remains challenging due to its variable morphology, spectral similarity, and dynamics of braided river. Advancements in artificial intelligence (AI) and unmanned aerial vehicle (UAV) remote sensing offer promising opportunities for addressing these applied geoscience challenges. In this study, we evaluate different AI techniques for the accurate detection of LWD in braided rivers. Specifically, using RGB-UAV imagery, we test two DL models, U-Net and DeepLabv3+, and compare them to other classifiers to identify the most accurate and transferable approach. The results indicate that the DeepLabv3+ method effectively captures the actual spatial distribution of LWD, and two-class classifications were more efficient than multi-class ones. Furthermore, the DL model demonstrated strong transferability when applied to a different spatiotemporal area, highlighting its utility for applied geoscience investigations and river management.
Han et al. (Sun,) studied this question.