Integrating deep learning detection and hydrological monitoring for driftwood flux and carbon stock estimation in a steep tropical basin

Global warming has increased the amount of deadwood in forests due to wildfires, insect outbreaks, and droughts. Deadwood and fresh wood are mobilised by erosion into river systems as driftwood, forming the largest organic carbon sink due to its slow decomposition rate. Thus, event-based driftwood transport is crucial for disaster management and assessing carbon storage. Here, we applied YOLOv8 to detect driftwood using images from surveillance cameras and a drone. Three types of driftwood, i.e., instream, riverbank, and nearshore, were used from the database of Swiss, Arctic Data Center, and our own drone surveys to train the model of object detection and instance segmentation. To estimate the volume of driftwood, we compared the detected image areas with radio-frequency identification (RFID) tagged logs of known dimensions, establishing an area-to-volume conversion. Our models achieved an mAP50 of 0.96 for in-stream object detection. Applying this model to Typhoon Kong-rey in the Liwu River, we estimated an in-stream driftwood volume of 3.5×10⁵ m³, with a carbon stock of 8.24×10¹⁰ g C, representing 0.11% of Taiwan’s annual carbon export. Furthermore, we observed that driftwood flux increases nonlinearly with river discharge. Our analysis suggests that driftwood accumulation along the outer bends of the riverbank may lead to pulsed driftwood flux. These findings highlight the significance of event-scale driftwood transport as a quantifiable component of green carbon and demonstrate the feasibility of integrating deep learning-based detection with hydrological monitoring for carbon budget assessments.

Keywords: Driftwood flux, YOLOv8, RFID, drone, green carbon