Answers are grounded in this paper and cite their sources.
Granular segregation is widely observed in geohazards and particulate transport, where it influences both dynamic behavior and structural evolution. However, continuous and quantitative measurement of segregation degree remains challenging in physical experiments because image observations are often affected by occlusion, illumination variation, boundary effects, and the lack of reliable true-value references. This study introduces SegVis, a vision-based framework for automatic estimation of granular segregation degree. The method uses front-view and top-view images as inputs, extracts centroid-based structural indicators and visible area-ratio features through color segmentation, and maps these visual proxies to the DEM-derived true segregation degree using a Random Forest model. A total of 27 binary segregation scenarios were simulated using the three-dimensional Discrete Element Method (DEM), generating 5400 image-label pairs covering different volume ratios, density ratios, size ratios, and particle shapes. For the DEM-derived dataset, SegVis achieved high predictive accuracy and stable generalization, with Random Forest outperforming single visual indicators and linear regression. Additional flume-like numerical simulations with non-cyclic boundaries were conducted to evaluate transferability under more experiment-like granular-flow configurations. The results show that prediction deviations mainly occur in the flow head and tail, where thin particle layers and image-quality variations weaken the reliability of visual proxies; removing these boundary-affected samples improved the prediction accuracy from R 2 = 0.82 to R 2 = 0.94. Overall, SegVis provides a non-invasive, automated, and scalable approach for dynamic segregation analysis, and its open-source implementation supports reproducible application to granular systems with visually distinguishable particle groups.
No takes yet. Share an insight, caveat, or question.
J Jieheng Guo (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: