HighFM: A Foundation Model for High-Frequency Earth Observation Data
Researchers have introduced HighFM, a novel foundation model designed to process high-temporal-resolution Earth Observation (EO) data for real-time disaster monitoring. Addressing the limitations of existing models that rely on low-revisit-rate satellite imagery, HighFM leverages over 2 TB of SEVIRI imagery from the Meteosat Second Generation platform. The model adapts the SatMAE masked autoencoding framework, enhanced with fine-grained temporal encodings to capture short-term variability in fast-evolving phenomena. Pretrained on this extensive dataset, HighFM was fine-tuned for cloud masking and active fire detection tasks. Benchmarking against traditional baselines and recent geospatial foundation models demonstrated consistent improvements in balanced accuracy and Intersection over Union (IoU) metrics. This development highlights the potential of temporally dense geostationary data for enhancing early warning systems and emergency response capabilities. By enabling robust spatiotemporal representation learning, HighFM offers a scalable pathway toward more effective foundation models for tracking climate-related disasters, marking a significant advancement in the application of machine learning to environmental monitoring and safety.
Wire timeline
HighFM: A Foundation Model for High-Frequency Earth Observation Data
Researchers have introduced HighFM, a novel foundation model designed to process high-temporal-resolution Earth Observation (EO) data for real-time disaster monitoring. Addressing the limitations of existing models that rely on low-revisit-rate satellite imagery, HighFM leverages over 2 TB of SEVIRI imagery from the Meteosat Second Generation platform. The model adapts the SatMAE masked autoencoding framework, enhanced with fine-grained temporal encodings to capture short-term variability in fast-evolving phenomena. Pretrained on this extensive dataset, HighFM was fine-tuned for cloud masking and active fire detection tasks. Benchmarking against traditional baselines and recent geospatial foundation models demonstrated consistent improvements in balanced accuracy and Intersection over Union (IoU) metrics. This development highlights the potential of temporally dense geostationary data for enhancing early warning systems and emergency response capabilities. By enabling robust spatiotemporal representation learning, HighFM offers a scalable pathway toward more effective foundation models for tracking climate-related disasters, marking a significant advancement in the application of machine learning to environmental monitoring and safety.
cs.AI updates on arXiv.org