LongEarth-R1: Benchmarking and Aligning Vision-Language Models for Long-Horizon Earth Observation Reasoning

• URL: http://arxiv.org/abs/2608.13344v1
• Date: Thu, 13 Aug 2026 15:14:59 GMT
• Title: LongEarth-R1: Benchmarking and Aligning Vision-Language Models for Long-Horizon Earth Observation Reasoning
• Abstract summary: Long-horizon Earth observation reasoning requires models to organize multi-stage geographic evolution, localize spatial changes, detect temporal anomalies, and infer future from extended image sequences.<n>We develop LongEarth through supervised fine-tuning with explicit sequence identifiers and structured chain-of-thought supervision.<n>LongEarth-R1 achieves the best results on all 12 long-sequence tasks while remaining competitive on standard remote sensing benchmarks.
• Score: 63.001172313432114
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Long-horizon Earth observation reasoning requires models to organize multi-stage geographic evolution, localize spatial changes, detect temporal anomalies, and infer future from extended image sequences. However, existing remote sensing vision-language models mainly focus on isolated images, image pairs, or short sequences, limiting reliable grounding in the relevant frames and regions. We introduce LongEarth-Bench, a benchmark containing 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 further provides structured reasoning traces linking key frames and changed regions to final answers. We develop LongEarth through supervised fine-tuning with explicit sequence identifiers and structured chain-of-thought supervision. Building on LongEarth, LongEarth-R1 applies group relative policy optimization with format, temporal, and spatial rewards. LongEarth-R1 achieves the best results on all 12 long-sequence tasks while remaining competitive on standard remote sensing benchmarks.

Related papers

• ChronoEarth-492K: A Large Scale and Long Horizon Spatiotemporal Hyperspectral Earth Observation Dataset and Benchmark [29.791943499456426]
  ChronoEarth-492K is the first large-scale, temporally calibrated hyperspectral archive built upon NASA's EO SSL-1 mission.<n>It comprises 492,354 harmonized patches across 185,398 global locations over 17 years, with 28,786 sites containing multi-temporal sequences.<n>Building on this foundation, we establish the ChronoEarth-Benchmark, a unified evaluation suite spanning static, short-horizon, and longhorizon- temporal tasks.
  arXiv  Detail & Related papers  (2026-05-15T06:44:23Z)
• TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation [71.39112735716172]
  We introduce TerraScope, a unified vision-language model that delivers pixel-grounded geospatial reasoning.<n>We also propose TerraScope-Bench, the first benchmark for pixel-grounded geospatial reasoning.
  arXiv  Detail & Related papers  (2026-03-19T15:38:02Z)
• CrossEarth-SAR: A SAR-Centric and Billion-Scale Geospatial Foundation Model for Domain Generalizable Semantic Segmentation [51.94680303125737]
  We present CrossEarth-SAR, the first billion-scale SAR vision foundation model built upon a novel physics-guided sparse mixture-of-experts (MoE) architecture.<n>To facilitate large-scale pre-training, we develop CrossEarth-SAR-200K, a weakly and fully supervised dataset that unifies public and private SAR imagery.<n>We also introduce a benchmark suite comprising 22 sub-benchmarks across 8 distinct domain gaps, establishing the first unified standard for domain generalization semantic segmentation on SAR imagery.
  arXiv  Detail & Related papers  (2026-03-12T14:52:35Z)
• Where on Earth? A Vision-Language Benchmark for Probing Model Geolocation Skills Across Scales [61.03549470159347]
  Vision-language models (VLMs) have advanced rapidly, yet their capacity for image-grounded geolocation in open-world conditions has not been comprehensively evaluated.<n>We present EarthWhere, a comprehensive benchmark for VLM image geolocation that evaluates visual recognition, step-by-step reasoning, and evidence use.
  arXiv  Detail & Related papers  (2025-10-13T01:12:21Z)
• GeoChain: Multimodal Chain-of-Thought for Geographic Reasoning [1.0799568216202955]
  GeoChain is a benchmark for evaluating step-by-step geographic reasoning in multimodal large language models (MLLMs)<n>It pairs each image with a 21-step chain-of-thought (CoT) question sequence (over 30 million Q&A pairs)<n>These sequences guide models from coarse attributes to fine-grained localization across four reasoning categories.
  arXiv  Detail & Related papers  (2025-06-01T02:24:46Z)
• TiMo: Spatiotemporal Foundation Model for Satellite Image Time Series [39.22426645737932]
  TiMo is a novel hierarchical vision transformer foundation model tailored for SITS analysis.<n>At its core, we introduce atemporal attention mechanism that dynamically captures multiscale patterns across both time and space.<n>Extensive experiments across multipletemporal tasks-including deforestation monitoring-demonstrate TiMo's superiority over state-of-theart methods.
  arXiv  Detail & Related papers  (2025-05-13T16:35:11Z)
• AerialVG: A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations [51.44608822712786]
  Visual grounding aims to localize target objects in an image based on natural language descriptions.<n>AerialVG poses new challenges, emphe.g., appearance-based grounding is insufficient to distinguish among multiple visually similar objects.<n>We introduce the first AerialVG dataset, consisting of 5K real-world aerial images, 50K manually annotated descriptions, and 103K objects.
  arXiv  Detail & Related papers  (2025-04-10T15:13:00Z)
• SpectralEarth: Training Hyperspectral Foundation Models at Scale [47.93167977587301]
  We introduce SpectralEarth, a large-scale multitemporal dataset designed to pretrain hyperspectral foundation models.<n>We pretrain a series of foundation models on SpectralEarth, integrating a spectral adapter into classical vision backbones.<n>In tandem, we construct nine downstream datasets for land-cover, crop-type mapping, and tree-species classification.
  arXiv  Detail & Related papers  (2024-08-15T22:55:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.