REMOTE SENSING · EVIDENCE · REASONING

DisasterAgent

An Evidence-Driven Agent for
Verifiable Disaster Report Generation

From Remote Sensing Imagery

Zhijian Li, Jiawei Song, Chao Ren, Yidan Zhang, Peijin Wang, Xian Sun, Kun Fu

Aerospace Information Research Institute, Chinese Academy of Sciences University of Chinese Academy of Sciences Key Laboratory of Target Cognition and Application Technology

Paired optical remote sensing images before and after an event DisasterAgent processes the image pair
Generate a verifiable disaster report with traceable evidence.

01 / THE IDEA

Every claim,
backed by evidence.

DisasterAgent links report statements to supporting evidence through inline citations, so readers can trace each claim back to its source and verify it against the imagery.

Structured memory

STPG · Spatial relations & temporal changes

Multi-source evidence extraction

SCE · VGE · EEE

02 / SEE IT IN ACTION

Demo Show

Follow the evidence

A case study of the 26 August 2026 debris-flow event at Gyirong Port on the China–Nepal border.

DisasterAgent / Workflow demo01:01 · No narration
A guided walkthrough of skill routing and the PAG workflow.

03 / HOW IT WORKS

Framework

Two parts. One evidence-driven agent.

Full framework

The complete framework. Skill routing exposes the relevant tools. The PAG workflow turns their outputs into structured memory, verified evidence, and traceable reports.

04 / EXPERIMENTS

More accurate, more verifiable.

Evaluated on RSCC and DisasterM3.

A Qwen3.5-27B central VLM, without disaster-specific fine-tuning. Agent baselines use the same specialized vision toolkit.

RSCC · ST5-SCS67.18%+15.34 percentage points over the base VLM
Disaster captioning2.87/ 5vs. 1.88 for the base VLM
Restoration advice3.22/ 5vs. 2.86 for the base VLM

2,512 optical image pairs · disaster report generation

Disaster report generation on DisasterM3
MethodDisaster captioningScores / 5 ↑Restoration adviceScores / 5 ↑
AVGDAPDDRFCAVGRNAPPSC
Kimi K2.61.371.611.211.302.963.522.782.57
MiniMax-M32.032.311.801.972.793.372.582.42
Qwen3.5-27B1.882.161.651.812.863.472.672.44
Qwen3.5-9B1.712.021.471.652.623.232.432.19
InternVL3.5-8B1.161.430.971.092.553.262.262.14
GeoChat-7B0.780.970.341.031.722.181.601.39
TeoChat-7B1.291.560.641.651.622.111.471.29
EarthAgent2.362.722.032.343.153.762.962.74
OpenEarthAgent2.282.611.942.293.133.762.922.70
DisasterAgent2.873.202.522.883.223.773.052.84
Disaster captioning across eight disaster types.
Same base VLM. Stronger evidence support.

05 / BEYOND ONE EVENT

Different disasters. Traceable evidence.

Look closer
CASE 01

Explosion

From a misidentified event to an evidence-grounded explanation: DisasterAgent links the explosion assessment to numbered visual references.

Original qualitative example from the paper. Colors and numbered citations connect the generated claims to the supporting imagery.

06 / CITE THIS WORK

Citation

If you find DisasterAgent useful in your research, please consider citing our work.

DisasterAgent robot
BibTeX
@article{li2026disasteragent,
  title   = {DisasterAgent: An Evidence-Driven Agent for Verifiable Disaster Report Generation From Remote Sensing Imagery},
  author  = {Li, Zhijian and Song, Jiawei and Ren, Chao and Zhang, Yidan and Wang, Peijin and Sun, Xian and Fu, Kun},
  journal = {arXiv preprint arXiv:2611.13776},
  year    = {2026},
  eprint  = {2611.13776},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url     = {https://arxiv.org/abs/2611.13776}
}

Figure