Claire Robin

Forecasting the impact of climate change on local vegetation is essential for effectively mitigating global warming effects on both human populations and biodiversity. This thesis focuses on the development and implementation of machine learning methods to predict the impact of climate extremes on vegetation.
First, we propose to develop a database of extreme events that considers the diversity of vegetation response to climate extremes, and their spatio-temporal extents to improve model training and validation.
Second, we aim to enhance current state-of-the-art models by using transformer-based approaches, addressing their quadratic time and memory complexity using domain science knowledge to drive the relationship between meteorological and environmental variables.
Third, we view unseen areas, climate shifts, and novel climate extremes as a data distribution shift problem. We propose to substitute space for time to address the data sparsity issue of climate shift and emphasize estimating the area of applicability of our model using various test sets representing different distribution shifts. Finally, we would like to evaluate the reliability of our method using weather forecasts and different climate projections for vegetation impact estimation at short and long term.
Accurately predicting vegetation response to climate extremes is essential for making informed decisions to mitigate climate change, protect ecosystems, and build more resilient and sustainable societies.