Julian Kuehnert
Speaker at AMLD Africa 2024
Julian is a Research Scientist at IBM Research Africa in Nairobi, Kenya. He is part of the Climate & Sustainability initiative, working with climate impact models for improved decision making. His current research interests focus around boosted calibration of model parameters using Bayesian optimization frameworks and uncertainty estimation in dynamical systems using Machine Learning approaches.
Prior to joining IBM Research in 2020, Julian studied Geophysics and received his PhD in 2019 from the IPGP in Paris, France, where he worked at the interface of computational and environmental seismology.
If you want to get a personal glimpse into Julian's life at the research lab in Nairobi, you can listen to his conversation with Jon Lenchner and Victor Akinwande on the "On Not Knowing" podcast, available on Spotify and Audible.
At AMLD Africa
United States International University Africa, Nairobi, Kenya
Geospatial AI foundation models for Climate change discovery
Foundation models are artificial intelligence (AI) models that are pre-trained on large unlabeled datasets through self-supervision and then fine-tuned for different downstream tasks. There is increasing interest in the scientific community to investigate whether this approach can be successfully applied to domains beyond natural language processing and computer vision to effectively build generalist AI models. Here, we introduce Prithvi, a geospatial AI foundation model pre-trained on large source of multispectral satellite imagery from the NASA Harmonized Landsat-Sentinel 2 (HLS). Prithvi is a Temporal Vision Transformer that includes positional and temporal embeddings, which was trained on IBM Cloud Vela cluster (NVIDIA A100 GPUs) using a Masked Auto Encoder approach and Mean Absolute Error loss function for a total of 10k GPUs hours. Benchmarking downstream tasks such as flood mapping and burn scar identification, Prithvi could successfully be fine-tuned to produce state-of-the-art AI models for Earth observation tasks with the potential to achieve peak performance on test data quicker and with less training. As example use cases, we consider the applicability of geospatial AI foundation models to monitor reforestation activities in Kenya’s National Water Towers. The pre-trained model and fine-tuning workflows are available open-source on Hugging Face (https://huggingface.co/ibm-nasa-geospatial).
