🌍Any Disaster Mapping

Earth Observation for Disaster Mapping: Benchmarks, Methods, Challenges and Future Perspectives

Hongruixuan Chen^1,†, Jian Song^1,†, Weihao Xuan^2,1,†, Junjue Wang^2,†, Heli Qi¹, Zeqi Zhou³, Pengyu Dai^1,2
Olivier Dietrich⁴, Erika Gutierrez⁵, Lars Bromly⁵, Edoardo Nemni⁶, Yafei Ou¹, Jie Zhao⁷, Zhuo Zheng⁸, Yonghao Xu⁹
Ronny Hänsch¹⁰, Wenzhe Jiao¹¹, Marco Chini¹², Claudio Persello¹³, Junshi Xia¹, Shijian Lu¹⁴, Lixin Wang¹⁵, Zhe Zhu¹⁶
Evan Shelhamer¹⁷, Jocelyn Chanussot¹⁸, Konrad Schindler⁴, Naoto Yokoya^2,1

^†Equal contribution
¹ RIKEN AIP, ² The University of Tokyo, ³ Brown University, ⁴ ETH Zurich, ⁵ United Nations Satellite Centre
⁶ Barcelona School of Economics, ⁷ Technical University of Munich, ⁸ Stanford University, ⁹ Linköping University
¹⁰ German Aerospace Center (DLR), ¹¹ Texas A&M University, ¹² Luxembourg Institute of Science and Technology
¹³ University of Twente, ¹⁴ Nanyang Technological University, ¹⁵ Indiana University Indianapolis
¹⁶ The University of British Columbia, ¹⁷ University of Connecticut, ¹⁸ Université Grenoble Alpes

Paper | Installation | Dataset Preparation | Pretrained Weights | Quick Start | Repo Layout

🔭Overview

Any Disaster Mapping is the official repository for our review paper in EO-based disaster mapping.

One of our key motivations is that current disaster mapping research is highly fragmented: benchmarks, tasks, and model implementations are often inconsistent across papers, making fair evaluation, reproduction, and further development unnecessarily difficult.

This repo unifies widely used disaster mapping benchmarks and representative deep learning models across major research directions, and provides a consistent training and evaluation pipeline for:

• Infrastructure damage
• Flood mapping
• Landslide segmentation
• Wildfire analysis

It is designed to help researchers:

• Reproduce the results reported in our paper
• Evaluate models under a unified protocol
• Use strong baselines out of the box
• Build and test their own improvements with minimal engineering overhead

🛠️Installation

Base environment and optional model-specific extras.

# NOTE: --index-url should match the version of your local CUDA toolkit for compiling ChangeMamba kernels (cu126 is just an example)
pip install torch torchvision xformers --index-url https://download.pytorch.org/whl/cu126
pip install -e .

Some models require optional extras:

• ChangeMamba selective scan kernel:

  # run `conda install -c conda-forge gcc=13 gxx=13 -y` if you meet GCC issues
  cd src/models/ChangeMamba/kernels/selective_scan
  pip install . --no-build-isolation

• Local pretrained checkpoints under pretrained_weight/ for model families such as SegFormer, HRNet, SAM/SAM2, DINOv3, HyperSigma, SkySense, SpectralGPT, and ChangeMamba. See Pretrained Weights below.

🧪Dataset Preparation

Dataset preparation guides are organized by disaster domain:

• Infrastructure damage: scripts/data_prep/infra_damage/README.md
• Flood: scripts/data_prep/flood/README.md
• Landslide: scripts/data_prep/landslide/README.md
• Wildfire: scripts/data_prep/wildfire/README.md

📦Pretrained Weights

Recommended local checkpoint layout for the supported model zoo.

Create the local checkpoint directory first:

mkdir -p pretrained_weight

Direct Downloads

# pretrain-vit-base-e199.pth
wget -O pretrained_weight/pretrain-vit-base-e199.pth \
  https://zenodo.org/records/7338613/files/pretrain-vit-base-e199.pth

# SpectralGPT+.pth
wget -O "pretrained_weight/SpectralGPT+.pth" \
  "https://zenodo.org/records/8412455/files/SpectralGPT+.pth?download=1"

# spec-vit-base-ultra-checkpoint-1599.pth
wget -O pretrained_weight/spec-vit-base-ultra-checkpoint-1599.pth \
  https://huggingface.co/WHU-Sigma/HyperSIGMA/resolve/main/spec-vit-base-ultra-checkpoint-1599.pth

huggingface-cli download UTokyo-Yokoya-Lab/AnyDisaster-Pretrained_Weight \
  vssm_tiny_0230_ckpt_epoch_262.pth --local-dir pretrained_weight --local-dir-use-symlinks False

Additional Upstream Checkpoints

• DINOv3 Source: https://ai.meta.com/resources/models-and-libraries/dinov3-downloads/ Download and save as:

  • dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth (ViT-B/16, LVD-1689M)
  • dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth (ViT-L/16, LVD-1689M)
  • dinov3_vitl16_pretrain_sat493m-eadcf0ff.pth (ViT-L/16, SAT-493M)

• SAM v1 Source: https://github.com/facebookresearch/segment-anything Download and save as:

  • sam_vit_b_01ec64.pth (SAM ViT-B)
  • sam_vit_l_0b3195.pth (SAM ViT-L)

• SAM 2.1 Source: https://github.com/facebookresearch/sam2 Download and save as:

  • sam2.1_hiera_small.pt (SAM 2.1 Hiera-Small)
  • sam2.1_hiera_base_plus.pt (SAM 2.1 Hiera-Base+)

• SegFormer MiT encoders Source: https://github.com/NVlabs/SegFormer Download and save as:

  • mit_b0.pth
  • mit_b1.pth
  • mit_b2.pth
  • mit_b3.pth
  • mit_b4.pth
  • mit_b5.pth

• HyperSIGMA spatial backbone Source: https://huggingface.co/WHU-Sigma/HyperSIGMA Download the upstream file, rename, and save as:

  • HSI_spatial_checkpoint-1600.pth

• SkySense backbone Source:

  • https://github.com/Jack-bo1220/SkySense
  • https://www.notion.so/SkySense-Checkpoints-a7fcff6ce29a4647a08c7fe416910509 Select the hr (high-resolution RGB / RGBNIR) variant, not the s2 Sentinel-2 variant. Save as:
  • skysense_model_backbone_hr.pth For commercial use, contact the authors (yansheng.li@whu.edu.cn).

After completing all downloads, pretrained_weight/ should contain:

pretrained_weight/
├── dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth
├── dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth
├── dinov3_vitl16_pretrain_sat493m-eadcf0ff.pth
├── HSI_spatial_checkpoint-1600.pth
├── mit_b0.pth
├── mit_b1.pth
├── mit_b2.pth
├── mit_b3.pth
├── mit_b4.pth
├── mit_b5.pth
├── pretrain-vit-base-e199.pth
├── sam2.1_hiera_base_plus.pt
├── sam2.1_hiera_small.pt
├── sam_vit_b_01ec64.pth
├── sam_vit_l_0b3195.pth
├── skysense_model_backbone_hr.pth
├── spec-vit-base-ultra-checkpoint-1599.pth
├── SpectralGPT+.pth
└── vssm_tiny_0230_ckpt_epoch_262.pth

🚀 Quick Start

Minimal training and evaluation commands.

Train

Train with a YAML config:

python train.py --config configs/infra/xbd/unet.yaml

Evaluate

Evaluate an experiment directory:

python test.py --exp_path results/xbd/unet

🗂️Repository Layout

Main directories and what they are responsible for:

• src/core/: trainer, config loader, registry, augmentation, metrics
• src/tasks/: task handlers for segmentation, change detection, and semantic change detection
• src/datasets/: dataset adapters and runtime data contracts
• src/models/: model wrappers and vendored third-party implementations
• configs/: experiment configs grouped by domain and dataset
• scripts/data_prep/: dataset preparation guides and helper scripts
• docs/: architecture notes and extension guidance

🏗️Architecture And Extension

Internal runtime design and extension entry points.

• Architecture overview: docs/ARCHITECTURE.md
• Extension guide: docs/EXTENSION_GUIDE.md

📜Reference

If this code repo contributes to your research, please kindly consider citing our paper and give this repo ⭐️ :)

🙋Q & A

For any questions, please feel free to leave it in the issue section or send inquiry email to qschrx@gmail.com.