Machine learning models for satellite-based coral reef mapping

<p dir="ltr">The ongoing crisis of climate change necessitates the development of effective methods for monitoring and mapping environmental features and species to ensure their preservation. This thesis explores the application of machine learning algorithms to efficiently map coral reefs using multispectral satellite images. The Maupiti lagoon in French Polynesia serves as a case study. The research led to the production of an automated tool capable of generating coral reef maps from satellite images. Moreover, the tool can be adapted to map other ecosystems, such as forests or ice sheets, provided that the model is retrained with relevant data.</p><p dir="ltr">To begin, a comprehensive literature review investigates current methods and trends in utilizing machine learning algorithms for coral reef mapping. Then, the attempts to develop the tool led us to face the special case of compositional data, which are data carrying relative information and lying in a mathematical space known as simplex. Adaptations of conventional methods are required to address the specific characteristics of this space.</p><p dir="ltr">First, in response to data imbalance, an oversampling technique is developed specifically for compositional data. Additionally, a spatial autoregressive model based on the Dirichlet distribution is formulated to account for spatial effects that may arise in the mapping process.</p><p dir="ltr">Finally, we present the implementation of our final mapping tool. To achieve the desired objective, a two-staged classification process is implemented, combining pixel-based and object-based approaches. This technique enables the tool to achieve an accuracy exceeding 85% with 15 classes.</p><p dir="ltr">The research contributes novel solutions for handling compositional data and delivers a high-performing mapping tool for coral reef ecosystems, aiding in environmental management and conservation efforts.</p>