issai/DL-SLICER-dataset

🌍 DL-SLICER Cities Dataset

Dataset Summary

The DL-SLICER Cities Dataset contains satellite imagery from 45 cities worldwide.
It was developed and used in the study:

DL-SLICER: Deep Learning for Satellite-Based Identification of Cities with Enhanced Resemblance
Ulzhan Bissarinova, Aidana Tleuken, Sofiya Alimukhambetova, Huseyin Atakan Varol, Ferhat Karaca

The dataset supports deep learning–based city classification and explainable AI (XAI) analysis of urban features.
In the published work, a ResNet-based model trained on this dataset achieved an overall accuracy of 84% in city classification.

Figure 1: Cities represented by their International Air Transport Association (IATA) codes and corresponding raw sattelite views from one of the sites. Raw site images are part of the dataset and were used for city identification model training.

Supported Tasks and Benchmarks

• Image Classification: Identify the city from a satellite tile.
• Explainability: Apply Class Activation Maps (CAMs) to analyze visual features that distinguish cities.
• Urban Similarity Analysis: Group cities by visual patterns to study sustainability, smart cities, and policy implications.

Dataset Structure

• raw/ folder contains original satellite images of size 4800 by 4800, across 3-4 sites through each presented city.
• processed/ folder contains divided satellite patches of size 480 by 480, with overlap size of 240, across the same sites and cities.

Data Fields

• image: RGB satellite image tile.
• label: City name (from 45 classes).

Example

{
  "image": "Almaty (ALA)/ALA_S1_2018_5_31_0.jpg",
  "label": "ALA"
}

For example, an image of Almaty city from site 1, captured on 31 May 2022, was saved with the filename ‘ALA_S1_2018_5_31.jpg’.

Data Splits

Split	# Images
Train	370,386
Val	86,526
Test	86,526

---

Cities

The dataset covers 45 cities across diverse:

• Geographic regions
• Human Development Index (HDI) levels
• Population sizes

---

Dataset Creation

Curation Rationale

The dataset enables comparative urban analysis through deep learning, allowing researchers and policymakers to understand similarities among cities based on satellite imagery.

Source Data

• Satellite imagery collected between 2018–2022.
• Original raw satellite images of 4800 x 4800 pixel size.
• Processed into uniform 480 x 480 pixel size tiles for ResNet training.

Annotation

• Each image labeled by city (label of the city is the first 3 letters of the image file name).
• No manual bounding boxes or masks — classification labels only.

---

Licensing

This dataset is released under the CC-BY-4.0 license.

---

Citation

If you use this dataset, please cite the article:

@article{bissarinova2024dl,
  title={DL-SLICER: Deep Learning for Satellite-Based Identification of Cities with Enhanced Resemblance},
  author={Bissarinova, Ulzhan and Tleuken, Aidana and Alimukhambetova, Sofiya and Varol, Huseyin Atakan and Karaca, Ferhat},
  journal={Buildings},
  volume={14},
  number={2},
  pages={551},
  year={2024},
  publisher={MDPI}
}

---

How to Use

from datasets import load_dataset

ds = load_dataset("issai/dl-slicer-dataset")

sample = ds["train"][0]

---

Limitations

• Limited to 45 cities → may not generalize globally.
• Satellite tiles may not capture socioeconomic or functional urban features.

---

Ethical Considerations

• Intended for urban planning, sustainability studies, and smart city research.

---

Contact

Maintainer: Ulzhan Bissarinova ([email protected]) Institute of Smart Systems and Artificial Intelligence, Nazarbayev University