TerraMind: Large-Scale Generative Multimodality for Earth Observation

Researchers have introduced TerraMind, described as the first any-to-any generative, multimodal foundation model for Earth observation, capable of processing nine geospatial data types simultaneously [1]. The model, detailed in a paper submitted to arXiv in April 2025 and last revised in June 2026, uses a dual-scale pretraining approach that combines token-level and pixel-level data [1][3]. On a token level, TerraMind encodes high-level contextual information to learn cross-modal relationships, while on a pixel level, it leverages fine-grained representations to capture critical spatial nuances [3]. This early fusion method unlocks a range of zero-shot and few-shot applications for Earth observation [5]. TerraMind was pretrained on a custom global-scale dataset called TerraMesh, which contains nine million samples aligned spatiotemporally and across modalities [3][4]. The dataset includes radar and optical satellite images from the Copernicus Sentinel-1 and Sentinel-2 missions, along with task-specific modalities such as land use and land cover maps, normalized difference vegetation index maps, digital elevation models, geographic coordinates, and natural language captions [3][4]. The model introduces a capability called "Thinking-in-Modalities," or TiM, which generates additional artificial data during finetuning and inference to improve output quality [1][2]. In benchmark testing, TerraMind achieved beyond state-of-the-art performance on the PANGAEA community-standard benchmark for Earth observation, outperforming 12 other geospatial models [1][5]. The pretraining dataset, model weights, and code have been open-sourced under a permissive license and are available on Hugging Face and GitHub [1][2]. The authors note that TerraMind's ability to integrate heterogeneous data sources suggests future expansions to multi-temporal, multi-resolution, and hyperspectral data [5]. The paper was authored by Johannes Jakubik and collaborators, with the initial submission totaling 36,600 KB and subsequent versions reaching approximately 39,925 KB [1].