LongEarth-R1: Benchmarking and Aligning Vision-Language Models for Long-Horizon Earth Observation Reasoning
LongEarth-Bench contains approximately 120k question-answering samples derived from 117k unique images. Its sequences average 15.14 frames and extend to 30 frames, covering 12 tasks across evolution summarization, spatial reasoning, anomaly identification, and logical prediction. A 30k-sample subset provides structured reasoning traces linking key frames and changed regions to final answers. LongEarth-R1 applies group relative policy optimization with format, temporal, and spatial rewards and achieves the best results on all 12 long-sequence tasks while remaining competitive on standard remote sensing benchmarks.
Researchers introduced LongEarth-Bench, a benchmark for long-horizon Earth observation reasoning with about 120k QA samples from 117k unique images, sequences averaging 15.14 frames and up to 30 frames, covering 12 tasks. A 30k subset includes structured reasoning traces. They developed LongEarth via supervised fine-tuning with explicit sequence identifiers and chain-of-thought supervision, then LongEarth-R1 using group relative policy optimization with format, temporal, and spatial rewards. LongEarth-R1 achieved the best results on all 12 long-sequence tasks and remained competitive on standard remote sensing benchmarks.
The approach combines supervised fine-tuning with explicit sequence identifiers and structured chain-of-thought supervision, followed by reinforcement learning with group relative policy optimization and three reward types (format, temporal, spatial). This design targets long-horizon reasoning by grounding answers in key frames and changed regions, addressing limitations of prior models that focused on isolated images or short sequences.
The benchmark and model address a gap in remote sensing vision-language models, which have largely been limited to single images or short sequences. By enabling reasoning over extended temporal sequences, this work could support applications such as environmental monitoring, urban planning, and disaster response where multi-stage geographic evolution and anomaly detection are critical.
The benchmark and model could lower the barrier for developing AI systems that analyze long-term Earth observation data, potentially enabling new products in climate monitoring, agriculture, infrastructure inspection, and defense. The structured reasoning traces may also improve interpretability and trust in automated geospatial analysis.
Observable next signals include adoption of LongEarth-Bench by other research groups, publication of follow-up models that improve on LongEarth-R1, and potential integration of long-horizon reasoning capabilities into commercial remote sensing platforms. Further scaling of sequence length and task diversity may also be explored.