README.md

---
task_categories:
- object-detection
- image-classification
tags:
- satellite
- airplane
- airport
pretty_name: FineAir
size_categories:
- 1K<n<10K
extra_gated_prompt: >-
  You need to agree to share your information and agree to the [Maxar Group Project License Version B4-21-21](https://huggingface.co/datasets/UniBwM-Informatik-IV/FineAir/resolve/main/LICENSE)
extra_gated_description: Our team may take a few days to process your request
extra_gated_fields:
  Family Name: text
  First Name: text
  Affiliation if applicable: text
  Institution Email: text
  Street Address: text
  Postal Code: text
  Country: text
  I agree to the license: checkbox
license: other
license_name: maxar-group-project-license-b4-21-21
---
# Dataset Card FineAir

<!-- Provide a quick summary of the dataset. -->

FineAir is a high-resolution optical image dataset specifically designed for fine-grained airplane classification. The dataset leverages transponder data (TD) from FlightRadar24 to provide accurate labels for aircraft at various levels of granularity, including Finest-Grained Class (FtGC). FineAir is the first dataset to offer detailed airplane categorization in 30 cm spatial resolution optical images, addressing the limitations of existing datasets by enhancing annotation accuracy.


#### Key Features

- 5,520 airplanes annotated on various granularity
- 1,350+ airplanes labeled with FtGC
- 20 fine-grained airplane classes
- Non-overlapping train, validation, and test splits, ensuring balanced distributions
- High-resolution imagery (30 cm GSD) from multiple satellites (WorldView-2, WorldView-3, and GeoEye)
- 16-bit Multi-band imagery (RGB + Near-Infrared)

#### Annotations

FineAir is created by matching satellite imagery with airplane transponder from FlightRadar24 data collected around the image acquisition time. This allows for accurate airplane labeling, including FtGC, which provides detailed information such as aircraft model variations (e.g., Airbus A320-200). The dataset comprises images from 29 airports, including major hubs such as Denver International Airport (DEN), covering an area  50 km².
The images are sourced from different satellites with varying ground sample distances (GSD), which are normalized to 30 cm through upsampling and pansharpening. The images are processed using proprietary high-pass Laplacian filters to enhance edge details, ensuring superior object recognition and classification.

FineAir provides a hierarchical class structure:

- Finest-Grained Class (FtGC) (e.g., A320-200)
- Fine-Grained Class (FGC) (e.g., A320)
- Role (e.g., Airliner, Private Jet, Propeller)
- Coarse-Grained Class (CGC) (e.g., Airplane)

Annotations are obtained from two primary sources:

- Transponder Data (TD) from FR24: Airplane TD is matched with satellite image timestamps, allowing for precise FtGC labeling. However, due to transponder inactivity, only 15% of the airplanes in the dataset are labeled with FtGC.
- Mask Templates: When TD is unavailable, aircraft are classified based on 105 mask templates, which consider features such as wing span-to-length ratio, engine count, cockpit shape, and propeller presence.

Additionally, aircraft that cannot be labeled via TD or templates are assigned a general role (e.g., airliner, private jet, propeller). Military-related aircraft are tagged with the suffix "-Military."

#### FineAir Class (FAC)

To address imbalanced class distributions, FineAir introduces FineAir Class (FAC), a simplified categorization ensuring all airplanes have at least one class label.
Rare classes (fewer than 50 instances) are merged into broader roles, and military aircraft are merged into their parent roles.

#### Dataset Splits

FineAir is divided into train, validation, and test sets, ensuring:

- Consistent distributions across splits
- Balanced FineAir Class (FAC) ratios
- Non-overlapping airplane instances (airplanes in different splits do not share image parts)

A random sampling process is performed to minimize the class distribution differences between the splits, ensuring reliable model evaluation.


### **Curated by:** Murat Osswald<sup>1</sup>, Louis Niederlöhner<sup>1</sup>, Sascha Köjer<sup>1</sup>, Tobias Ziedorn<sup>1</sup>, Valerio Gulli<sup>2</sup>, Michael Mommert<sup>3</sup>, Helmut Mayer<sup>1</sup>

<sup>1</sup>*Institute for Applied Computer Science, University of the Bundeswehr Munich*

<sup>2</sup>*European Space Imaging, Germany*

<sup>3</sup>*Stuttgart University of Applied Sciences, Germany*

### Dataset Sources 


- **Repository:** [satellitepy](https://github.com/Iammuratc/satellitepy)
- **Paper:** [FineAir](https://openaccess.thecvf.com/content/WACV2025W/CV4EO/html/Osswald_FineAir_Finest-grained_Airplanes_in_High-resolution_Satellite_Images_WACVW_2025_paper.html)

## Dataset Structure

In progress...

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

#### Who are the source data producers?

European Space Imaging ([EUSI](https://www.euspaceimaging.com/))

## Dataset Card Contact

[email protected]