A robust, user-friendly tool for accurate fluvial grain-size/substrate class estimation

Accurate estimation of sediment size and substrate classes in fluvial remote sensing is pivotal for habitat modeling and hydrodynamic applications. While recent advancements have adopted computer vision based approaches (i.e. deep learning), the complexity of setting up these algorithms, along with the requirement of dedicated hardware, and the lack of readily available tools, hinder their wider adoption. This study presents a novel, user-friendly two-step tool tailored for precise substrate class estimation in clear-water river environments from ultra high-resolution orthoimagery, typically coming from UAVs. Leveraging image texture properties (evaluated with the co-occurrence matrix), image color channels (typically Red, Blue, and Green bands) and machine learning classificators (i.e. Random Forest, Support Vector Machine), the proposed methodology is able to accurately identify substrate classes ranging from fine sediments (e.g. sand and lime), various size gravel and cobbles, and boulders, both submerged (wet) and above water. It is a 2-step methodology that involves (a) manual labeling of homogeneous substrate class patches within any Geographic Information System (GIS) platform, followed by (b) streamlined data input. Validation across three reaches of gravel-bed rivers —Aurino, Piave, and Brenta rivers in NE Italy— with differing sizes and morphologies, and substrate ranging from fine sediments to boulders, yielded F1 scores of 0.86, 0.97, and 0.938, respectively. Some challenges still arise when classifying substrate in areas where visibility and light conditions are significantly altered, such as in very deep water, within tree canopy shadows, or due to strong sun reflections. Finally, this tool enables easy and accurate substrate class estimations in riverine environments, offering a significant contribution to fluvial studies and applications.