The AU8 Benchmark Dataset

This is the code repository for our paper, “AU8: A Multimodal Benchmark Dataset from Eight Australian Cities for Urban Profiling and Analysis”. AU8 comprises eight major Australian cities including Greater Sydney, Greater Melbourne, Greater Brisbane, Greater Perth, Greater Adelaide, Greater Canberra, Greater Darwin, and Greater Hobart and contains 101,604 satellite image tiles, each paired with a declarative textual description and nine key urban indicators (e.g., population density, median income, housing price, land use). The dataset is available at https://huggingface.co/datasets/anonymous-for-review/AU8.

In AU8 we also provide comprehensive metadata for all Images, which are listed below.

Attribute	Description
image_name	Image file name with latitude and longtitude
SA2_Code	ABS Statistical Area Level 2 (SA2) code
latitude	Latitude of the image centre
longitude	Longitude of the image centre
Median house price	Median price of established house transfers, 2023 (AUD)
Population density	Population density, 2023 (persons / km²)
Median income	Median total income excluding government pensions and allowances, 2020 (AUD)
No. of businesses	Total number of businesses, 2023 (mean)
protected land	Total protected land area, 2022 (ha, mean)
No. of jobs	Number of jobs, 2020 (mean)
Persons employed	Total persons employed aged 15 years and over, 2021 (mean)
Agricultural land	Area of agricultural land, 2021 (ha, mean)
Rural residential	Rural residential and farm infrastructure area, 2016 (ha, mean)
Description	Textual description of the image generated by GPT-5

🛰️ Experiments

🔧 Setup

Before running, make sure you have installed all required dependencies:

pip install -r requirements.txt

---

🚀 Quick Start

Each method has a simple two-step (train → predict) or one-step pipeline.

---

1️⃣ UrbanCLIP

Step 1: Train UrbanCLIP

python UrbanCLIP/main.py

Step 2: Predict with pretrained model

python UrbanCLIP/UrbanCLIP_predict.py

---

2️⃣ GeoVit-HNM

Step 1: Train GeoVit-HNM

python GeoVit-HNM/GeoVit-HNM.py

Step 2: Predict with trained model

python GeoVit-HNM/GeoVit-HNM_predict.py

---

3️⃣ Tile2Vec

Step 1: Train Tile2Vec embeddings

python tile2vec/tile2vec_triplet.py

Step 2: Predict urban indicators

python tile2vec/tile2vec_predict.py

---

4️⃣ PCA

Run PCA + XGBoost directly:

python pca/PCA.py

---

5️⃣ ResNet18

Run ResNet18 + XGBoost directly:

python resnet-18/resnet-18.py

---

📌 Notes

• Training scripts automatically save model checkpoints.
• Prediction scripts output evaluation metrics and prediction results.
• Make sure your dataset files are placed in the correct data/ directory.