Research map / Vision-language and multimodal / Multimodal LLMs
GeoReason: Bridging Logical Reasoning and Spatial Fidelity in Remote Sensing Segmentation
Cheng Fan, Han Wu, Tingzhang Luo, Minjing Dong, Yehui Tang, Jianyuan Guo
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GeoReason, an MLLM-based approach for instruction-driven reasoning segmentation in remote sensing imagery, aimed at fixing spatial fidelity degradation and target misidentification in the Embedding-as-Mask paradigm.
Problem
Reasoning segmentation of remote sensing imagery with MLLMs under cluttered geographic contexts and extreme scale variation, which causes spatial fidelity degradation and target misidentification.
Approach
Extends the Embedding-as-Mask paradigm of MLLM-based segmentation to remote sensing reasoning segmentation; the abstract does not specify the mechanism beyond targeting spatial fidelity and logical reasoning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Anticipatory Prior (prompt-side [SEG] token) (ablation) | GeoReason · EarthReason test | cIoU: 70.45 -> 78.87 | Improved |
| add Latent PTA loss (ablation) | GeoReason · EarthReason test | cIoU: 70.45 -> 73.01 | Improved |
| add Location-aware CoT (BBox) alone, without AP (ablation) | GeoReason · EarthReason test | gIoU / cIoU: gIoU 75.32 -> 74.54; cIoU 70.45 -> 70.59 | No clear effect |
| add BBox alone (ablation) | GeoReason · RRSIS-D test | gIoU: 66.42 -> 65.29 | Hurt |
| add PTA loss (ablation) | GeoReason · RRSIS-D test | gIoU: 66.42 -> 66.99 | Improved |
| add BBox and PTA on top of AP (full model) (ablation) | GeoReason · EarthReason test | cIoU: 78.87 -> 80.60 | Improved |
| place extra [SEG] in the answer stream instead of prompt (ablation) | GeoReason · EarthReason | Avg of gIoU and cIoU: 78.61 -> 74.14 | Hurt |
| remove CoT reasoning (Zero-Reasoning) from baseline (ablation) | Baseline · EarthReason | Avg of gIoU and cIoU: 72.79 -> 74.63 | Improved |
| remove reasoning from GeoReason (ablation) | GeoReason · EarthReason | Avg of gIoU and cIoU: 78.61 -> 74.63 | Hurt |
| add CoT to baseline answer-side [SEG] (ablation) | Baseline · EarthReason test | image attention mass: 0.246 -> 0.212 | Hurt |
| add Anticipatory Prior to baseline (ablation) | GeoReason · EarthReason test | image attention mass: 0.212 -> 0.245 (answer-side), 0.270 prompt-side | Improved |
| rewrite queries as direct referring expressions (oracle) | Baseline · EarthReason | Avg of gIoU and cIoU: 72.79 -> 80.84 | Improved |
Later papers that compared with this method
1 later accepted paper reports GeoReason as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Evaluating and Enhancing Negation Comprehension in Remote Sensing MLLMs (ECCV 2026 · better on 3 of 3 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- GeoPixel — GeoPixel: Pixel Grounding Large Multimodal Model in Remote Sensing (ICML 2025) · better on 1 of 2 shared results
- Text4Seg++ — Text4Seg: Reimagining Image Segmentation as Text Generation (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Builds on: Embedding-as-Mask
Closest papers by meaning
- Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.92)
- LISAt: Language-Instructed Segmentation Assistant for Satellite Imagery (NeurIPS 2025 · similarity 0.91)
- RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.91)
- Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- GeoCoT: Towards Reliable Remote Sensing Reasoning with Manifold Perspective (CVPR 2026 · similarity 0.91)
- TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation (CVPR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
