IBM has released its most powerful geospatial foundational model to date, with Prithvi-EO-2.0 boasting 600 million parameters.
The model is trained on NASA's HLS V2 product which provides 30m granularity, with Prithvi 2.0 capable of picking out individual tree-species and crop types from satellite images, and cattle and solar panels from drone footage.
IBM’s range of geospatial models are allowing communities experiencing the firsthand effects of climate change to identify risk-areas, and warn populations ahead of time.
Battling energy, flooding, and urban heat islands
Speaking on the development of Prithvi 2.0 at IBM Research Zurich, Juan Bernabe-Moreno, Director IBM Research Europe and Ireland & UK, explained the model is trained on both Harmonized Landsat Sentinel-2 (HLS) data and Modern-Era Retrospective analysis for Research and Applications (MERRA-2) data.
Using HLS data allows users to fine tune specialized models capable of detecting flooding, wildfires, and landslides, as well as quantifying biomass, classify specific crops, measure canopy heights and track land usage.
The MERRA 2 model is capable of tracking weather, hurricanes, and turbulence, alongside specific capabilities in climate model precipitation, weather modelling, and temperature mapping. Bernabe-Moreno notes IBM hopes to have a climate prediction model released in 2025.
What's more, Prithvi 2.0 has expanded capabilities such as Multi-Temporal Cloud Gap Imputation that allow the model to reconstruct areas of satellite imagery that are covered by clouds, which helps improve accuracy when monitoring environments, planning agricultural changes, and estimating crop yields.
IBM also hopes to take the lessons learned from the development of Prithvi 2.0 and use them to create an open source model aimed at managing power infrastructure and sustainable energy called GridFM. The project, in collaboration with LF Energy and Hydro Quebec, looks to build a model capable of predicting power loads and renewable energy generation, track distribution and transmission of power, and simulate future power usage to help forecast pricing and demand.
The future of climate prediction
“The global south is at the forefront of climate change,” says Ambassador Philip Thigo, Special Envoy on Technology for the Republic of Kenya, noting it is especially hard for areas affected by climate change in Africa to address the challenges they face without the right technology.
By being able to access tools such as Prithvi 2.0, it is now possible for countries with less resources to address the effects of climate change and help to preserve their ecosystems.
In 2023, IBM announced a partnership with the Government of Kenya with the aim of assisting reforestation efforts in areas that have seen large levels of biomass decline caused by human activity. Ambassador Thigo states that IBM’s geospatial models have allowed the Kenyan Government to not only accurately track their reforestation efforts, but also attract project funding by being able to provide detailed information on their efforts.
Additionally, Prithvi 2.0 and its data has been used to track and predict flooding patterns, allowing the Kenyan Government to issue 48 hour warnings during periods of heavy rain for communities at risk of flooding.
For Matthew Chersich, Research Professor in the Climate and Health Directorate at Wits RHI, heat is a significant area of concern. Chersich focuses on urban heat islands - urban areas that are markedly hotter than surrounding areas due to infrastructure, human activity, and reduced green space.
Geospatial AI has been a key tool in mapping variations of heat in African urban areas, and has allowed researchers to model health risks based on heat islands, flooding, and other climate issues - being especially useful as an early warning system. The 30m x 30m resolution provided by Prithvi 2.0 allows for highly localized temperature readings which can be cross referenced with drought data to provide specified warnings to residents in and around heat islands and drought-risk areas.
GridFM could make the path to sustainability more sustainable
“By 2050, power production will need to double in order to meet decarobonization goals,” François Mirallès, Researcher at Hydro Québec, points out, further stating by 2035 around 60tWhs of sustainable energy will be added to the energy grid. The difficulty in managing and predicting supply and demand is exacerbated by the lack of consistency, when it comes to sustainable energy sources such as solar, wind, and tidal.
The long term goal in the development of GridFM is to provide the energy industry with a model that can be trained to perform highly specialized tasks to address and handle these multiple variables to not only provide long-term energy forecasting, but also hourly predictions and anomaly detection.
Dan Brown, Head of Marketing at LF Energy Foundation, adds to Mirallès’ point stating that addressing climate change and decarbonization “requires change from the ground up.” The complexity of sustainable energy makes it difficult to scale, but open source models and research are helping to fuel innovation and address the key challenges of creating a sustainable future, especially for tasks such as load forecasting and expansion prediction.
One particular challenge faced in the training of AI to predict energy usage and peak loads is that energy data is regulated creating governance issues in its usage. Brown notes that tools such as OpenSynth are especially useful in these circumstances, as it is trained on real data but provides synthetic data in its outputs.
IBM and its partners in the AI Alliance are demonstrating that open source is the future of research and innovation, allowing trust in AI models to be built and demonstrated in transparency and collaboration.
You might also like
- “If we want our data to remain confidential, we have to act now”: Quantum computing can crack all our encryption, and that is everyone’s problem
- These are the best AI tools and best AI writers
- “Making the types of problems that were impossible, possible” - How IBM is shaping the future of quantum computing
