Emulating the estuarine morphology evolution using a deep convolutional neural network emulator based on hydrodynamic results of a numerical model

Abstract

Coastal and estuarine areas present remarkable environmental values, being key zones for the development of many human activities such as tourism, industry, fishing, and other ecosystem services. To promote the sustainable use of these services, effectively managing these areas and their water and sediment resources for present and future conditions is of utmost importance to implement operational forecast platforms using real-time data and numerical models. These platforms are commonly based on numerical modelling suites, which can simulate hydro-morphodynamic patterns with considerable accuracy. However, in many cases, considering the high spatial resolution models that are necessary to develop operational forecast platforms, a high computing capacity is also required, namely for data processing and storage. This work proposes the use of artificial intelligence (AI) models to emulate morphodynamic numerical models results, allowing to optimize the use of computational resources. A convolutional neural network was implemented, demonstrating its capacity in reproducing the erosion and sedimentation patterns, resembling the numerical model results. The obtained root mean squared error was 0.59 cm, and 74.5 years of morphological evolution was emulated in less than 5 s. The viability of surrogating numerical models by AI techniques to forecast the morphological evolution of estuarine regions was clearly demonstrated.

A Comprehensive Review on Sediment Transport, Flow Dynamics, and Hazards in Steep Channels

“…In this regard, a study emphasized the importance of managing coastal and estuarine areas for sustainable use. It introduced a convolutional neural network to efficiently emulate complex numerical models, accurately forecasting morphological changes in estuaries in seconds, optimizing computational resources, and enhancing operational forecast platforms (de Melo et al 2022). The limitations of current sediment transport studies utilizing ML and AI approaches include challenges related to data quality and availability.…”

Machine Learning in Coastal Engineering: Applications, Challenges, and Perspectives

“…Images of hydrodynamic numerical simulations were fed into the network for morphology computation. The network architecture, adapted from Melo et al [100], is composed of U-nets and recursively deconvolutional branched networks to map the features of input images. These input images were the discretized spatial domain of mean velocity and bed shear stresses resulting from the hydrodynamic simulation.…”