Modelling the Hydro-morphodynamics of Mangrove Environments Using Surrogate Machine Learning Methods

Abstract

Mangrove ecosystems play a crucial role in mitigating the impact of climate change, acting as natural buffers against coastal erosion and storm surges. Traditional numerical models have been instrumental in simulating the dynamics of these environments necessary for implementing protection and restoration projects. Nonetheless, their computational needs in terms of resources have limited their applicability for large and complex regions. The advancements of machine learning models, augmented by the abundance of data, have helped tackle the computational burden by learning complex relationships from the underlying physical processes. In particular, an emerging field in machine learning models is hybrid physics-informed neural networks, where existing knowledge of the physical system is integrated into the machine learning training pipeline to improve on computational efficiency and prediction reliability. Furthermore, uncertainty quantification using probabilistic machine learning models has gained significant attention due to its ability in easing the computational burden. This thesis investigates how using machine learning models can improve the understanding of complex mangrove dynamics. First, focusing on the Sundarbans, the largest mangrove forest in the world, a novel hydro-morphodynamic model is introduced. This model provides the first comprehensive simulation of hydro-morphodynamics of mangrove environments in real-world scenarios incorporating tidal waves. Through validation against real-world tidal gauge data, the model showed that mangrove environments are able to decrease water elevations and velocities by more than 97%, and prevent almost any sediment erosion when compared with the experiment with no mangroves. Furthermore, the model outperforms existing numerical models with up to 118% reduction in errors compared against actual tidal data. Secondly, a novel high-resolution hybrid physics-informed neural network model is presented as a potential replacement for the computationally expensive state-of-art numerical model. Leveraging temporal causality, this machine learning framework adeptly captures temporal dynamics while significantly enhancing data efficiency and reducing computational complexity. Validation against the numerical model showed an average error between O−2 and O−6 confirming its reliability and robust performance. Furthermore, training the model was 5 times faster than running the numerical model, and inference took just 15 seconds, whereas the numerical model had to run the full simulation again. Finally, to address uncertainties inherent in the outputs of the numerical model, a novel probabilistic uncertainty quantification framework using deep Gaussian process is presented.

A comparative analysis is performed to assess the improvements of this method against standard Gaussian processes, explaining their unique use-cases and implications for the field. In terms of accuracy, the deep Gaussian process was over 5 orders of magnitude better than the Gaussian process model. These advancements mark a significant leap forward for researchers and stakeholders invested in assessing and mitigating coastal risks arising from climate change using advanced machine learning methods.