多模態與遙測
LongEarth-R1 Reasons Over Satellite Sequences of Up to 30 Frames, Outperforming Comparison Models Across All 12 Spatiotemporal Tasks
LongEarth-Bench organizes 117,000 remote-sensing images into approximately 120,000 long-sequence question-answer pairs, requiring models to track changes in landforms, anomalies, and missing states. LongEarth-R1, built on Qwen2.5-VL-7B, adds frame numbering, structured reasoning, and spatiotemporal rewards, but its data, weights, and code have not yet been released.
Remote-sensing vision-language models typically process a single image, a before-and-after pair, or a very short time series, making it difficult to answer multi-stage questions such as “When did the flooding begin, how did the affected area expand, and did it subsequently recover?” LongEarth-Bench therefore integrates five datasets, including SpaceNet 7, DynamicEarthNet, and PASTIS-R, to create 120,367 question-answer pairs. Each pair contains an average of 15.14 frames and up to 30 frames, while 21.1% contain at least 24 frames. The benchmark covers 12 tasks across evolution summarization, spatial reasoning, anomaly recognition, and logical prediction.
The team used Qwen2.5-VL-7B as the base model, froze the vision encoder, and trained rank-128 LoRA adapters only in the language layers. In the first stage, it added explicit sequence numbers to each image and then used 30,000 structured reasoning traces to teach the model to identify key frames, regions of change, and temporal order. In the second stage, GRPO compared multiple responses to the same input and assigned separate rewards for formatting, temporal localization, and spatial consistency, rather than evaluating only the final answer.
LongEarth-R1 ranked first across all 12 long-sequence tests reported in the paper. For example, it achieved 90.28% on duplicate-frame detection, compared with 39.78% for the base Qwen2.5-VL model; on temporal-order violation detection, performance rose from 12.38% to 50.80%. Ablation studies also showed that all three reward types were necessary. However, the results come from a benchmark created and used for training by the authors themselves, so data from the same sources may disadvantage comparison models. Some reasoning annotations were initially generated by Qwen3-VL-8B-Thinking, which may also have introduced its preferences into the evaluation. As of publication, the paper did not include public data, weights, or training code. Engineering teams should wait for a reproducible release and specifically validate generalization across regions and sensors, as well as under seasonal interference.