Quantitative Characterization of Gravel Bed Structures Using 2DSSFs and Machine Learning‐Based Surrogate Surface Generation

Abstract Accurate characterization of gravel structures is essential for understanding riverbed roughness, flow resistance, and sediment transport. However, quantifying the definitive characteristics of these gravel structures remains challenging due to the inherent complexity of particle spatial arrangement. This study presents an improved statistical method for characterizing gravel structures, utilizing two‐dimensional second‐order structure functions (2DSSFs) and a more robust surrogate surface generation methodology for significance testing. The method is first validated using synthetic surfaces composed of regularly arranged spherical particles, demonstrating its ability to capture the geometric features of grain structures. To address the lack of null hypothesis data sets for natural gravel beds, a machine learning–based technique is proposed for generating randomized surrogate surfaces from down‐sampled mosaic images of actual riverbeds. This method preserves particle‐scale morphology while disrupting spatial correlations between particles. Two independent models were trained on artificial and mixed (artificial and natural) gravel surfaces, allowing the generation of surrogate surfaces that replicate either artificially prepared or natural gravel bed features. The application of 2DSSFs significance testing to natural gravel surfaces revealed spatial correlation patterns related to gravel clustering, orientation, and scale, while minimizing interpretative subjectivity. Our findings highlight the potential of integrating spatial correlation analysis with machine learning–based surrogate surface generation to advance the quantitative analysis of complex gravel‐bed structures.