Two-dimensional convolutional neural network outperforms other machine learning architectures for water depth surrogate modeling

Abstract

Rapid prediction of spatially distributed hydrological variables, such as water depths in rivers, is an important but challenging task. This study proposes a novel matrix-based deep learning approach for predicting spatial distribution of water depths in rivers. The proposed approach was constructed based on a two-dimensional (2D) convolutional neural network (CNN) with a new architecture that was specifically designed for providing spatial distribution maps. A numerical dataset was established based on a field cruise and two-dimensional hydraulic modeling for different scenarios, and numerical experiments were designed to predict spatial distribution of water depths for different scenarios using the adaptive neuro-fuzzy inference system (ANFIS), support vector machine (SVM), genetic programming (GP), multi-gene genetic programming (MGGP), one-dimensional CNN (1D-CNN), and the proposed CNN algorithms. The results showed that the proposed CNN approach captured both the large-scale and small-scale spatial patterns remarkably well, and it outperformed the other approaches. This study shows that the 2D CNN algorithm is better than the classical machine learning (ML) algorithms for inundation modeling. The proposed approach is thus a promising tool for providing rapid predictions of spatial distribution of water depths in river systems and can potentially be leveraged to predict other spatially distributed hydrological variables.