DSGR: Domain Shift across Geographic Regions

A large-scale Domain Generalisation (DG) benchmark for land-use classification in satellite imagery under spatial domain shift.

Official dataset of the paper: "Analysing Satellite Imagery Classification under Spatial Domain Shift across Geographic Regions", International Journal of Computer Vision (IJCV), 2025.

Sara A. Al-Emadi, Yin Yang, Ferda Ofli — Qatar Computing Research Institute (QCRI), Hamad Bin Khalifa University (HBKU), Qatar

📄 IJCV Paper · 🎬 AAAI'26 Talk · 💻 GitHub · 🌐 Project Page

---

Dataset Summary

DSGR is a novel Domain Generalisation benchmark designed to portray the effects of spatial domain shift (covariate shift) on satellite imagery classification at a global scale. Domains are defined as the six geographic regions established by the Population Division of the United Nations Department of Economic and Social Affairs: Asia (AS), Africa (AF), Oceania (OC), Europe (EU), Latin America and the Caribbean (LAC), and Northern America (NA).

The dataset comprises 362,582 samples across 41 land-use classes, built on top of the raw satellite imagery from fMoW (Functional Map of the World).

	AS	AF	OC	EU	LAC	NA
Training	51,266	20,254	8,610	97,625	37,603	65,065
Validation	7,243	2,865	1,420	14,853	5,072	9,438
Test	7,264	2,883	1,396	14,644	5,452	9,629
Countries	49	52	6	43	43	3

---

⚠️ Important: fMoW Raw Imagery Required

This repository contains the DSGR metadata only — the geographic region assignments, domain splits, and per-image annotations that define the benchmark. It does not contain the satellite images themselves.

To use DSGR, you must download the raw RGB imagery from the publicly available fMoW dataset:

1. Follow the official download instructions at https://github.com/fMoW/dataset (fMoW-rgb variant).
2. Download the DSGR metadata from this repository.
3. Join the two using the img_path column, which points to each image's relative location inside the fMoW directory structure (e.g. train/tunnel_opening/tunnel_opening_370/tunnel_opening_370_6_rgb.jpg).

from huggingface_hub import snapshot_download

# Download DSGR metadata
snapshot_download(repo_id="RWGAI/DSGR", repo_type="dataset", local_dir="/data/dsgr")

import pandas as pd
import os

FMOW_ROOT = "/data/fmow"   # your fMoW-rgb download
df = pd.read_csv("/data/dsgr/<metadata_file>.csv")

# Resolve full image paths
df["full_path"] = df["img_path"].apply(lambda p: os.path.join(FMOW_ROOT, p))

---

Data Fields

Field	Type	Description
split	string	Original fMoW split (train, val, test, or ignore for samples excluded from DSGR)
img_filename	string	Image file name
img_path	string	Relative path to the image inside the fMoW directory
spatial_reference	string	Coordinate reference system (GCS_WGS_1984)
epsg	int	EPSG code (4326)
category	string	Land-use class label
visible	bool	Whether the object is visible in the image
img_width / img_height	int	Image dimensions in pixels
country_code	string	ISO 3166-1 alpha-3 country code
cloud_cover	int	Cloud cover percentage
timestamp	string	Satellite image acquisition time (ISO 8601)
lat / lon	float	Coordinates of the imaged site
regions	string	DSGR geographic region (UN classification) — the domain label
subregions	string	UN geographic subregion
split_new	string	DSGR split assignment (train, val, test, or ignore)

---

Land-Use Classes (41)

Airport, Airport hangar, Airport terminal, Amusement park, Archaeological site, Barn, Burial site, Crop field, Dam, Educational institution, Electric substation, Factory or powerplant, Fire station, Flooded road, Gas station, Golf course, Ground transportation station, Helipad, Lighthouse, Office building, Oil or gas facility, Park, Parking lot or garage, Place of worship, Police station, Port, Prison, Railway bridge, Recreational facility, Road bridge, Shipyard, Shopping mall, Single-unit residential, Smokestack, Stadium, Storage tank, Surface mine, Swimming pool, Tower, Water treatment facility, Wind farm.

---

Benchmark Usage

DSGR supports the experimental protocols evaluated in the paper:

• Single-source DG — train on one region, evaluate on the remaining five (Section 5.3)
• Multi-source DG — leave-one-domain-out training (Section 5.4)
• Foundation model backbones — CLIP ViT-B/16 and ViT-L/14 (Section 5.5)
• OOD-aware validation — a second left-out domain used as OOD validation set (Section 5.6)

Training and evaluation code is available at github.com/RWGAI/DSGR.

---

Licensing

The DSGR metadata is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0). The underlying satellite imagery is distributed separately by fMoW under the Functional Map of the World Challenge Public License; users must comply with both licenses.

---

Citation

If you use DSGR in your research, please cite:

@article{alemadi2025dsgr,
  title     = {Analysing Satellite Imagery Classification under
               Spatial Domain Shift across Geographic Regions},
  author    = {Al-Emadi, Sara A. and Yang, Yin and Ofli, Ferda},
  journal   = {International Journal of Computer Vision},
  volume    = {133},
  number    = {11},
  pages     = {7672--7709},
  year      = {2025},
  publisher = {Springer Nature},
  doi       = {10.1007/s11263-025-02518-z}
}

---

Contact

For questions, please open a discussion on this repository or contact Sara A. Al-Emadi at salemadi@hbku.edu.qa.