Abstract. The sediment of alluvial riverbeds plays a significant role in river systems both in engineering and natural processes. However, the sediment composition can show great spatial and temporal heterogeneity, even on river reach scale, making it difficult to representatively sample and assess. Indeed, conventional sampling methods in such cases cannot describe well the variability of the bed surface texture due to the amount of energy and time they would require. In this paper, an attempt is made to overcome this issue introducing a novel image-based, Deep Learning algorithm and related field measurement methodology with potential for becoming a complementary technique for bed material samplings and significantly reducing the necessary resources. The algorithm was trained to recognise main sediment classes in videos that were taken underwater in a large river with mixed bed sediments, along cross-sections, using semantic segmentation. The method is fast, i.e., the videos of 300–400 meter long sections can be analysed within minutes, with very dense spatial sampling distribution. The goodness of the trained algorithm is evaluated mathematically and via intercomparison with other direct and indirect methods. Suggestions for performing proper field measurements are also given, furthermore, possibilities for combining the algorithm with other techniques are highlighted, briefly showcasing the multi-purpose of underwater videos for hydromorphological adaptation. The paper is to show the potential of underwater videography and Deep Learning through a case study.