Reproduce “LongEarth-R1: Benchmarking and Aligning Vision-Language Models for Long-Horizon Earth Observation Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yupan Ding, Jing Xiao, Zhenyuan Zhang, Chaofeng Chen, Liang Liao, Gui-Song Xia, Mi Wang
The LongEarth-R1 framework enables vision-language models to perform long-horizon Earth observation reasoning over extended image sequences, moving beyond the short temporal spans of existing remote sensing models. It achieved the highest performance across 12 cognitive tasks on the new LongEarth-Bench dataset, demonstrating scores like 90.28% and 89.12% in anomaly identification tasks, which often saw competitor models score below 50%.
Key insights
- Long-horizon Earth observation reasoning can be effectively structured into four distinct cognitive dimensions: Evolution Summarization, Spatial Reasoning, Anomaly Identification, and Logical Prediction.
- A combination of explicit sequence identifiers, Chain-of-Thought supervision, and Group Relative Policy Optimization (GRPO) with complementary format, temporal, and spatial rewards is crucial for robust multi-stage spatiotemporal reasoning.
- GRPO, particularly, yields substantial improvements (e.g., +12.61% on Anomaly Identification), demonstrating its effectiveness in strengthening evidence selection and spatiotemporal consistency in long temporal contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.