Mahmoud al Najar’s thesis defense

On November 30th, Mahmoud al Najar succesfully defended his PhD, entitled Modelling coastal evolution with machine learning (abstract below) ¹. Mahmoud was my first student, in the sense that we were working together as early as 2019, when he did an internship that led to his thesis. He tackled the challenging problem of interdisciplinary research with grace over his PhD, becoming an expert in both AI and coastal science. Mahmoud will be continuing his research as a postdoc at IRIT, working with Ehouarn Simon on ML for oceanography. I’m looking forward to working with Mahmoud in the future on this important domain of accelerating the understanding of our changing climate with AI.

Modelling coastal evolution with machine learning

The ever-growing density of human populations in coastal areas, coupled with the increased frequency and intensity of coastal hazards resulting from climate change, create serious safety hazards and present a number of challenges to coastal managers and engineers. The ability to predict the evolution of coastal systems is a requirement for effective coastal management including risk assessment and mitigation, and has been a fundamental goal of coastal research for decades. However, the process of coastal evolution is complex by nature, and predicting its development through time remains challenging using current methodologies. The absence of representative datasets which accurately track the state and evolution of coastal systems greatly limits our ability to predict coastal change in response to different natural and anthropogenic pressures. Traditional field surveys are used extensively in the Coastal Science literature, and have served as important assets in advancing our knowledge of these systems. However, a number of operational costs and constraints limit the use of these techniques to a few areas of the world, and to sparse temporal resolutions which are insufficient to effectively track different climatic modes, storm impact and recovery, for instance. Satellite-based Remote Sensing (RS) provides a series of technologies and techniques for frequently monitoring coastal areas around the globe at high temporal resolutions and scales. While space-borne RS unlocks the potential for studying coastal environments around the globe, it requires the development of novel data processing methodologies for handling large streams of Earth Observation data. Machine Learning (ML) is a subfield of Artificial Intelligence which aims at constructing algorithms able to leverage large amounts of example data in order to construct predictive models, and has been a critical component of many scientific advancements in recent years. Given the continuous stream of high-dimensional data recorded by multiple Earth Observation satellite constellations, in addition to the impressive performance of modern ML across diverse scientific and industrial domains, ML is often adopted in EO data processing pipelines in order to augment or replace more conventional signal and image processing-based analysis. This thesis aims to examine the potential and capability of modern ML in two important problems in Coastal Science where the potential of ML remains relatively unexplored. In this thesis, Deep Learning and Interpretable Machine Learning are applied to the problems of satellite-derived bathymetry and shoreline evolution modelling. We demonstrate that ML is competitive with current physics-based baselines on both tasks. Furthermore, we show the potential of ML in automating many of our large-scale coastal data analysis, towards gaining a global understanding of coastal evolution. We conclude our work by discussing the difficulties encountered, the limitations of both ML methodologies and how they can be improved, in addition to the long term perspectives that can be built upon this work.