Following on from our halfway point update on the ESL 2026 challenges last month, we can now bring you the coverage from the Live Showcase event held in August. This acts as a round-up of the eight week sprint period, prior to the teams’ writing up their findings in full technical papers.
MoEarths
Can we blend foundation models to leverage their complementary strengths?
As a reminder, the first challenge, MoEarths (Mixture of Earths), proposed to build framework for combining geospatial foundation models to better address challenging tasks arising from the interlinked nature of the Earth’s climate systems. This was to try and address the nature of the existing 100+ models that were built with unique strengths: temporal specialists, modality specialists, high-resolution specialists, etc., that perform well on one task and fail on another.
By fusing elements of many of these models together, they hoped to create a combined geospatial foundation model (FM) that works robustly for any application (flood, fire, forest) anywhere on Earth. Watch their Live Showcase presentation below to learn more.
It’s clear from the presentation that the team has successfully combined multiple models. Not a giant model that lumps in all existing ones, but rather a smarter one drawing on the best components of each. We’ll be back with a full detailed case study of this work later in the year.
Living Forest
Can we make use of the newest BIOMASS satellites to create a foundation model with a full complementary view of the forest?
The second ESL 2026 challenge, Living Forest, sought to build another multimodal foundational model, this time focused on the Earth’s forests and biomass. It was previously shown it was possible to classify a forest from space to query its carbon sequestration capabilities with around 80% accuracy with a view to applying it to biomass P-band data.
This year’s team now wanted to better understand how to measure living forests to help track and manage their health and resilience over time, attempting to develop the first Dynamic Forest Foundation Model designed to learn the underlying physics of forests.
The presentation details how the team developed Biomass-FM, a single model built from combining data from ESA’s Biomass mission with several other satellite and space station-based sources. It offers up to 34% greater accuracy and detail, in tree mapping, height and biomass, than existing separate models, and the team intends to release Biomass-FM as a baseline for others to build further on. Check back here for a full detailed case study of this work later in the year.
Tropical Cyclone Dynamics
Can we understand the dynamics of cyclones as they approach landfall?
The final ESL 2026 challenge, Tropical Cyclone Dynamics, looked to address the gaps in our knowledge when it comes to predicting the intensity of tropical cyclones. Together they represent critical threats whena single event can incur billions in damages and long-lasting impacts on human lives.
Despite advances in numerical weather prediction, a persistent resolution gap remains between global simulations and the fine-grained spatial structure of storms remains. The team sought to build a novel multimodal AI model that fused 3D cloud morphology with temperature, precipitation and synthetic aperture radar (SAR) wind speeds for the first time.
The team’s presentation shows how their model and corresponding dashboard improves on current state-of-the-art results by tracking the rapid intensification event that previous models failed to predict. The hope is that this work can be further fine-tuned to develop better early warning forecasting, specifically the intensification of wind speeds. Check back later in the year to read a full detailed case study of this work.