The scouring effect of the flowing water around bridge piers may undermine the stability of the structure, leading to extremely high direct and indirect costs and, in extreme cases, the loss of human lives. The use of Artificial Neural Network (ANN) models has been recently proposed in the literature for estimating the maximum scour depth around bridge piers: this study aims at further investigating the potentiality of the ANN approach and, in particular, at analysing the influence of the experimental setting (laboratory or field data) and of the sediment transport mode (clear water or live bed) on the prediction performances. A large database of both field and laboratory observations has been collected from the literature for predicting the maximum local scour depth as a function of a parsimonious set of variables characterizing the flow, the sediments and the pier. Neural networks with an increasing degree of specialization have been implemented – using different subsets of the calibration data in the training phase – and validated over an external validation dataset. The results confirm that the ANN scour depths' predictions outperform the estimates obtained by empirical formulae conventionally used in the literature and in the current engineering practice, and demonstrate the importance of taking into account the differences in the type of available data – laboratory or field data – and the sediment transport mode – clear water or live bed conditions.