Room 109, Building 9, St Eloi campus Machine Learning in Montpellier, Theory & Practice The acceleration of human-mediated species transfers between biogeographical regions increases the frequency of biological invasions, with significant costs to the environment and our societies. Spatial dynamic models based on established and interpretable ecological mechanisms are a crucial tool for understanding and anticipating biological invasions. However, we lack standardized spatio-temporal data to measure species life history traits that determine these mechanisms, such as their species fecundity, mortality, movement strategies and speed, and their dependence on life stages or the environment. Since my postdoc at Stellenbosch University, I have been working to integrate available massive but heterogeneous biodiversity observations to estimate species’ life history traits using Bayesian state-space models, including observation models adapted to the many sampling biases that affect data distribution relative to actual population distribution. I will present a proof of concept of the approach where we integrated presence-only data from three datasets to reconstruct the past spatial dynamics of an invasive bush in South Africa, as well as general limitations to introduce ongoing and future work. Machine Learning in Montpellier, Theory & Practice

Room 109, Building 9, St Eloi campus Machine Learning in Montpellier, Theory & Practice The field of remote sensing (RS) has witnessed remarkable advancements, but the scarcity of labeled data remains a challenge. This talk focuses on leveraging transfer learning and self-supervised learning (SSL) methods to benefit downstream tasks with limited data by learning representations from large datasets. However, most pre-trained models in RS are based on ImageNet or MS COCO, which may not capture the spectral and spatial characteristics of RS data. To address this, SSL provides a solution by learning useful feature representations from unlabelled RS data. By considering multimodal data, particularly optical and radar images, SSL models can improve accuracy and robustness in RS analysis. The talk aims to shed light on the potential of transfer learning and SSL methods in RS, emphasising the importance of considering the multi-modal nature of RS data. Machine Learning in Montpellier, Theory & Practice