91´ó»ÆѼ

Large wood (LW; >10 cm in diameter and >1 m in length) within river corridors - Ìýincluding channels and adjacent floodplains - plays a key role in shaping hydraulic conditions, sediment deposition and erosion, nutrient cycling, and habitat availability for diverse aquatic and terrestrial species. Thus, to fully understand how a river system functions, we must understand when, how, and why LW is stored or transported.Ìý

Here, we present the results of initial tests that use a CNN called You Only Look Once (YOLO) v26 to track wood pieces in videos of a set experiments investigating wood transport in forested river valleys. To create training data, we extracted and annotated wood in 800 video frames from 16 representative experiments. We then retrained the out-of-box YOLOV26n model to detect wood pieces. We investigated different data augmentation schemes, such as altering image color and image mosaicing. Once this model was trained and validated, we used a tracker algorithm called BOTsort to relate wood detections across frames to create traces for each piece of wood.Ìý

We found that image mosaicing and color augmentations improved detection precision, recall, and mean average precision by 10-15%. Additionally, BOTsort produced fragmented traces, which could be stitched together in a post processing step to create much more reliable traces. These results represent a step towards using advanced machine learning techniques to support LW research. Successful wood trace extraction will likely result in new insights into how wood is transported in complex environments and how wood jams accumulate. Understanding these complex processes will improve our understanding of wood in river systems and help river managers make informed decisions about wood in their systems.Ìý