LongEarth-R1: Benchmarking and Aligning Vision-Lang... | AI Research

LongEarth-R1 is a framework designed to improve how vision-language models reason about long-term geographic changes. While existing models often focus on static images or short sequences, this research addresses the need for models to interpret multi-stage processes—such as urban construction, disaster recovery, and ecosystem changes—by analyzing extended sequences of Earth observation imagery.

Benchmarking Long-Horizon Reasoning

The researchers introduced LongEarth-Bench, a dataset containing approximately 120,000 question-answering samples derived from 117,000 unique images. The sequences in this benchmark average 15.14 frames and extend up to 30 frames. The benchmark organizes reasoning into four cognitive dimensions: evolution summarization, spatial reasoning, anomaly identification, and logical prediction. To support evidence-based analysis, a subset of 30,000 samples includes structured reasoning traces that explicitly link key frames and spatial regions to the final answer.

The Two-Stage Training Approach

The authors developed a two-stage framework to enable long-horizon reasoning:

1. Supervised Fine-Tuning (LongEarth): The model is trained using explicit sequence identifiers for each frame to establish temporal anchors. It also uses structured chain-of-thought (CoT) supervision, which teaches the model to perform visual scanning, identify key features, and integrate evidence across the entire sequence before reaching a conclusion. 2. Reinforcement Learning (LongEarth-R1): Building on the supervised model, the researchers applied Group Relative Policy Optimization (GRPO). This stage uses rewards based on format validity, temporal grounding (ordering and frame selection), and spatial consistency (agreement on changed regions and extents). This process shifts the model from simple answer imitation toward more structured and consistent spatiotemporal reasoning.

Performance and Capabilities

According to the study, LongEarth-R1 achieved the highest results across all 12 long-sequence tasks defined in the benchmark. It also maintained competitive performance on standard remote sensing benchmarks, including single-image, bi-temporal, and short-sequence tasks. The researchers note that the model’s ability to handle long-horizon reasoning is supported by its capacity to identify chronological violations, detect redundancies, and predict missing states within extended temporal trajectories.

Considerations for Implementation

The effectiveness of the model relies on the integration of both supervised reasoning traces and reinforcement learning rewards. The authors’ ablation study indicates that while sequence identifiers and CoT supervision significantly improve performance, the addition of GRPO further enhances the model's ability to handle complex reasoning tasks. The researchers emphasize that the benchmark’s diversity—covering various geographic regions and land-cover categories—is intended to prevent models from relying on simple shortcuts, such as sequence length, to solve tasks.