Abstract

In-channel wood constitutes a key component of the riverine environment, yet it poses certain risks. Traditionally, methods for the identifying of in-channel wood have been labor-intensive, relying predominantly on field-based visual inspection. Accurate detection of in-channel wood is essential for understanding river processes, yet conventional field surveys remain costly, time-consuming, and spatially limited. Here, we present an unmanned aerial vehicle (UAV)-based deep learning framework that systematically evaluates and compares four state-of-the-art semantic segmentation architectures, U-Net, Feature Pyramid Network (FPN), DeepLabV3+, and Attention U-Net, for automatic mapping of in-channel wood. Using a novel high-resolution dataset of over 10,000 manually annotated in-channel wood from three Chilean rivers, we designed experiments to evaluate: (i) the influence of additional topographic layers (elevation and roughness), (ii) sensitivity to multiple spatial resolutions (0.05, 0.1, 0.15, and 0.30 m/pixel), and (iii) transferability to independent river reaches. Results demonstrate that the Attention U-Net trained solely with RGB imagery at 0.1 m/pixel achieved the best overall performance (Overall Accuracy = 97%, Dice = 0.82), surpassing all other models. Adding topographic layers did not improve accuracy, confirming that the attention mechanism effectively captures spatial features from RGB inputs. Model performance decreased slightly at coarser resolutions, but Attention U-Net consistently outperformed alternatives across all tests. Transfer learning results confirmed the model's robustness, which correctly mapped >65% of in-channel wood surfaces in validation reaches, including both single logs and wood jams. This work advances current practice by providing the first systematic benchmarking of deep learning architectures for in-channel wood detection, coupled with controlled experiments on additional channels inputs and spatial resolution effects. The efficacy of the trained model highlights the significant potential of UAV-derived orthomosaics in the automatic detection and delineation of diverse in-channel wood configurations. This study establishes a foundation for the advancement of UAV and satellite-assisted monitoring of in-channel wood, carrying significant implications for the progression of ecological and geomorphological research.