IBM and NASA have jointly announced the release of the Watsonx.ai geospatial foundation model on Hugging Face, a significant development that aims to harness the potential of vast amounts of satellite imagery to advance climate science and enhance life on Earth. The model was initially disclosed in February and is based on NASA’s Harmonized Landsat Sentinel-2 satellite data (HLS). It has undergone additional fine-tuning using labeled data for specific use cases like burn scar and flood mapping.
One of the key advantages of the geospatial foundation model lies in its utilization of enterprise technologies from IBM’s watsonx.ai initiative. Both organizations anticipate that the innovations introduced through this model will prove beneficial for scientific and business applications.
The foundation model’s most notable feature is its ability to address the challenge of data labeling at scale. Traditionally, AI training required extensive sets of labeled data. However, with foundation models, the AI is pre-trained on a large dataset of unlabeled data, and then fine-tuned using a smaller amount of labeled data for a specific use case. This approach allows for highly customized models and has demonstrated faster training and improved accuracy compared to models solely built with labeled data.
For example, when applied to flood prediction, the new foundation model achieved a 15% improvement in prediction accuracy using only half the amount of labeled data compared to a state-of-the-art model. Similarly, for the burn scar use case, the IBM model required 75% less labeled data than the current state-of-the-art model, resulting in significant performance enhancements.
IBM and NASA chose to make the geospatial foundation model available on Hugging Face due to the platform’s reputation as a leading community for open AI models. By doing so, they aim to foster its adoption and hope to gather insights and feedback from the community to further enhance the model’s capabilities over time.
In addition to benefiting scientists who work with satellite data, the model is expected to have implications for enterprise use cases of AI. IBM’s environment intelligence suite, which aids organizations with sustainability efforts, will eventually integrate the new model. Moreover, the experience gained from scientists fine-tuning the foundation model could lead to improvements in other areas of IBM’s AI development efforts through ‘meta learning.’
Overall, the release of the geospatial foundation model on Hugging Face represents a significant step in advancing AI applications for geospatial data analysis and holds promise for scientific and business communities alike.