Brief

Researchers have introduced TerraMind, a novel multimodal foundation model designed for Earth observation tasks. This model uniquely combines token-level and pixel-level data representations, allowing it to capture both high-level contextual information and fine-grained spatial details. TerraMind demonstrates strong zero-shot and few-shot learning capabilities, introduces a new technique called "Thinking-in-Modalities" (TiM) for data augmentation during fine-tuning and inference, and achieves state-of-the-art performance on benchmarks like PANGAEA. The model, its pretraining dataset, and associated code are publicly available under a permissive license. AI

Researchers have introduced RAMEN, a novel multimodal encoder designed for Earth observation data. This encoder is unique in its ability to handle diverse spatial, spectral, and temporal resolutions across various sensors without requiring sensor-specific adjustments. RAMEN treats resolution as a controllable parameter, allowing users to balance detail with computational cost. The model was trained on masked multimodal Earth observation data and has demonstrated effective transfer learning to new sensor configurations, outperforming existing state-of-the-art models on the PANGAEA benchmark. AI