Scour Around River Spur Dikes: A Review of Head-Shape Classification, Governing Factors, and Predictive Modelling Approaches

Spur dikes are among the most widely used river-training structures for bank protection, flow alignment, and land reclamation. However, the same obstruction to flow that makes them effective also triggers localised erosion, or scour, at their base and nose. If left unchecked, this scour undermines the foundation of the structure and is the most common cause of its failure. As a result, the prediction and control of scour depth around spur dikes has remained an active area of hydraulic research for more than seven decades. This paper reviews the literature on scour around spur dikes, tracing the subject from its origins as an empirical, practice-driven field of river engineering to its current state, which combines physical modelling, computational fluid dynamics, and machine learning. The review is organised around five themes: (i) the historical development of spur dikes as river-training works in different regions; (ii) the classification of spur dikes by their head shape, namely I-, L-, and T-shape, and the comparative performance of each; (iii) the mechanism of local scour, centred on horseshoe-vortex formation and downflow impingement; (iv) the governing factors, including spur geometry, sediment characteristics, and approach-flow conditions, that control the magnitude of scour; and (v) the evolution of predictive tools, from empirical regression equations to numerical simulation, including recent machine-learning and hybrid ensemble models. Across these themes, the review finds that I-shaped spur dikes remain the most extensively studied configuration, but L- and T-shaped spur dikes redistribute flow more favourably and reduce scour. Machine-learning models, particularly ensemble and hybrid frameworks, have generally been reported to outperform both empirical and single-algorithm approaches, although the comparisons rarely use common datasets or validation protocols, and field-scale performance remains untested. The review closes by identifying gaps in the literature, including the limited exploration of composite or curved head-shape geometries, unresolved questions regarding optimal spacing for such geometries, and the need for physics-informed hybrid machine-learning frameworks validated on independent field-scale data.