Earth-OneVision: Extending Remote Sensing Multimodal Large Language Models to More Sensor Modalities and Tasks

Earth-OneVision is a 2 billion-parameter remote sensing multimodal large language model that unifies six sensor modalities (optical, SAR, infrared, multispectral, temporal, and video) and performs nine task categories through a single framework. The model achieves competitive or superior performance compared to larger models (4B-72B parameters) on multiple benchmarks, supported by a new 34M QA pair dataset spanning cross-sensor fusion applications.

Earth-OneVision addresses a critical fragmentation problem in remote sensing AI where existing models support only narrow sensor types and task categories, limiting geoscientific applications. The research demonstrates that architectural innovations—Full-Granularity Vision-Language Alignment, Spatial-Linguistic Isomorphic Serialization, and Progressive Cross-Modality Adaptation—can effectively handle heterogeneous sensor data within a single efficient framework.

This work builds on the broader trend of unified multimodal AI systems that consolidate previously siloed capabilities. Remote sensing traditionally relied on specialized models for different sensor types, creating inefficiencies and preventing cross-modal knowledge exploitation. The introduction of MMRS-OneVision dataset with 34M QA pairs represents substantial progress in training infrastructure, nearly doubling available annotated resources for this domain.

The practical implications extend across agriculture, climate monitoring, disaster response, and urban planning sectors where multi-sensor Earth observation drives decision-making. Organizations currently deploying multiple specialized models can potentially consolidate infrastructure while maintaining performance levels. The 2B parameter efficiency is particularly valuable for deployment in resource-constrained environments or real-time processing scenarios.

Industry observers should monitor adoption patterns among Earth observation providers and whether similar unified approaches emerge in other multi-sensor domains like medical imaging or autonomous systems. The benchmark results suggest parameter efficiency and architectural design may outweigh model scale, challenging assumptions about necessary model sizes for complex spatial reasoning tasks.

• →Earth-OneVision unifies six sensor modalities and nine task categories in a single 2B-parameter model, outperforming larger 4B-72B parameter alternatives.
• →The MMRS-OneVision dataset containing 34M QA pairs substantially expands training resources for cross-modal remote sensing applications.
• →Three novel mechanisms address domain gaps between heterogeneous sensor types, enabling effective multi-modal fusion within an autoregressive framework.
• →Performance benchmarks show 87.52% P@0.5 on optical visual grounding and 80.68% on SAR VQA tasks, demonstrating competitive capability across modalities.
• →Parameter efficiency enables practical deployment in resource-constrained environments while maintaining state-of-the-art performance on Earth observation tasks.