Egor Vikhrev on Unsplash

Carolina collaborated with colleagues from the Chair for Artificial Intelligence in Climate and Environmental Sciences on a newly published study in Machine Learning: Earth. They benchmarked seven deep learning models against two baseline approaches for next-day wildfire danger prediction in the Mediterranean region. They also applied explainable AI techniques to evaluate whether the models learned physically meaningful wildfire relationships rather than relying solely on predictive accuracy.

Carolina Natel de Moura, KIT

Carolina led the review study “Deep Learning for Satellite-Based Forest Disturbance Monitoring: Recent Advances and Challenges,” published in WIREs, in collaboration with colleagues across Germany and the US. In this work, the authors identify three main technical avenues to address current limitations in forest disturbance detection, attribution, and training data scarcity: spatiotemporal architectures, embeddings and geospatial foundation models, and learning approaches designed for limited labelled data. The review further emphasizes that progress toward large-scale, reliable forest monitoring will require improved benchmark datasets, stronger interdisciplinary collaboration, and more open and standardized data-sharing practices.

Jens Krause, KIT

April 28-30 the LPJ-GUESS community met for their annual hybrid meeting, discussing latest developments and having fun at KIT-Campus Alpin (IMKIFU) in Garmisch-Partenkirchen. The photo shows the in-Person participants.