Forecasting the Future with EarthPT

Explore the groundbreaking EarthPT, an open-source Earth Observation pretrained transformer, in this 30-minute conference talk at EuroPython 2024. Delve into the capabilities of this 700 million parameter decoding transformer foundation model, specifically developed for Earth Observation use-cases. Learn how EarthPT, trained on satellite imagery time series, accurately predicts future pixel-level surface reflectances across a wide spectral range. Discover its impressive performance in forecasting the Normalised Difference Vegetation Index (NDVI) evolution, outperforming traditional models. Gain insights into the semantic richness of EarthPT's embeddings and their potential applications in dynamic land use classification, crop yield prediction, and drought forecasting. Understand the exciting scalability prospects of EarthPT, drawing parallels with Large Language Models' neural scaling laws. Access the MIT-licensed EarthPT on GitHub and explore its potential to revolutionize Earth Observation data analysis and forecasting.

Forecasting the future with EarthPT — Mike Smith