ATFMTraj: Aircraft Trajectory Classification Data for Air Traffic Management

Since the aircraft trajectory data in the field of air traffic management typically lacks labels, it limits the community's ability to explore classification models. Consequently, evaluations of clustering models often focus on the correctness of cluster assignment rather than merely the closeness within the cluster. To address this, we labeled the dataset for both classification and clustering tasks by referring to aeronautical publications. The process of obtaining the ATFM trajectory dataset consists of data sourcing, preprocessing, and annotation.

Airport Data Retrieval and Integration

Incheon International Airport (ICAO airport code: RKSI)

• The ADS-B recordings were sourced from the Opensky database [Opensky]. The data, covering the period from 2018 to 2023, were queried based on the flight identification numbers of the flights that departed from and arrived at the airport, as per the schedule on the Airportal website [Airportal]. To improve balance, we downsampled the trajectory data from the southbound and southeastbound flights. The datasets for arrivals and departures are denoted as RKSIa and RKSId, respectively.
• The data retrieval process for our study involves two main sources: AirPortal and OpenSky. Initially, flight data, including times, IATA and ICAO callsigns, and airport information, are collected for flights departing from and arriving at Incheon Airport, and stored in the Flights table. Subsequent database tables such as Airport, Callsign, Route, and Aircraft Carrier are created to organize and link the collected data effectively. For trajectory data, we match ICAO callsigns from the Flights table with OpenSky's dataset, adjusting timestamps to enhance data accuracy. This trajectory data, encompassing details like latitude, longitude, and altitude, is then integrated into the Trajectory table, linking each entry back to the corresponding flight record. This meticulous approach ensures a comprehensive and cohesive dataset that supports our analysis of flight trajectories and operational patterns.

Stockholm Arlanda Airport (ICAO airport code: ESSA)

• The data was sourced from the Swedish Civil Air Traffic Control (SCAT) dataset [Nilsson et al. 2023], which includes surveillance data, weather, flight plans, and airspace data. For this paper, we exclusively focused on the surveillance data for positional states and flight plan data to assist our manual labeling. Although departure data are available, they are found to be incomplete; consequently, we focus only on arrival data for this airport.

Zurich Airport (ICAO airport code: LSZH)

• Data was sourced from the reference dataset provided by [Olive et al. 2020]. The work focuses on analyzing arrival flight trajectories at Zurich Airport, and as such, we utilized this arrival data for our experiments."

Data Preprocessing Steps

To prepare the trajectory data for analysis, we follow several key steps:

1. Clean: Filter out incomplete trajectories to ensure data quality.
2. Transform: Convert geographic coordinates (Latitude, Longitude, Altitude) to Cartesian coordinates (x, y, z) in the East-North-Up (ENU) coordinate system, bounded by r_max.
3. Resample: Adjust the sampling rate to a 1-second interval without interpolating to a fixed length.
4. Smooth: Remove outliers and apply a Savitzky-Golay filter to smooth the data.
5. Scale: Normalize the coordinates by dividing by r_max, ensuring all data points scale within the range ([-1, 1]).
6. Pad: Use NaNs to pad trajectories of unequal lengths to T_max, the maximum data length.

Parameters:

• r_max values: 120 km for Incheon, 100 km for Stockholm Arlanda, and 40 nautical miles for Zurich.
• T_max is the maximum allowable data length across all datasets.

Annotation process

• RKSIa and LSZH (Arrival Trajectories at Incheon and Zurich Airports)

  • Arrival Corridors: Segmentation begins with K-means clustering on the set of first positional states (x,y) for each trajectory, followed by manual adjustments.
  • Runway Labeling: Individual runways are labeled using manually drawn linear classification lines on the set of last positional states (x,y) to accurately classify touchdown points.
  • Approach Procedures: Use a manual classification of IAF to identify different approach procedures.
  • Final Labeling: Classes are defined by runways, IAFs, and STARs combinations.

• RKSId (Departure Dataset at Incheon Airport)

  • Departure Corridors: Segmentation begins with K-means clustering on the set of last positional states (x,y) for each trajectory, followed by manual adjustments.
  • Runway Labeling: Dependent pairs of runways (same SID) are labeled using manually drawn linear classification lines on the set of first positional states (x,y) to accurately classify airborne points.
  • Final Labeling: Classes are defined by only SIDs.

• ESSA (Arrival Trajectory Data at Stockholm Arlanda Airport)

  • Integration with Flight Plan data: We used the STARs and Runways recorded in the flight plan data.
  • Corrections: Perform manual corrections on STARs and runway assignments.
  • Approach Procedures: Use a manual classification of IAF to identify different approach procedures.
  • Final Labeling: Classes are defined by runways, IAFs, and STARs combinations.

Please refer to the {dataset}_stats.csv for detailed class descriptions and statistics.

Example Usage

This Python function is designed to load ATFMTraj datasets from TSV files, manage missing values, and return both data and labels. The function assumes that the dataset is split into separate files based on a base name, mode, and coordinate variable (X, Y, Z).

import numpy as np
import pandas as pd

def load_ATFM(dset_name, mode, path):
    variables = ['X', 'Y', 'Z']
    data = []
    labels = None

    for var in variables:
        tsv_filename = f'{path}/{dset_name}_{mode}_{var}.tsv'
        df = pd.read_csv(tsv_filename, sep='\t', header=None, na_values='NaN')
        if labels is None:
            labels = df.values[:, 0]
        var_data = df.values[:, 1:]
        data.append(var_data)

    data = np.stack(data, axis=-1)

    return data, labels.astype(int)

# Example usage
train_data, train_labels = load_ATFM('LSZH', 'TRAIN', 'LSZH')
test_data, test_labels = load_ATFM('LSZH', 'TEST', 'LSZH')

Citation

Please cite our work if you use any of the datasets shared here (RKSIa, RKSId, ESSA, LSZH):

@dataset{ATFMTraj2024,
  title={ATFMTraj: Aircraft Trajectory Classification Data for Air Traffic Management},
  author={Phisannupawong, Thaweerath and Damanik, Joshua Julian and Choi, Han-Lim},
  year={2024},
  note={https://huggingface.co/datasets/petchthwr/ATFMTraj}
}

@misc{ATSCC2024,
  title={Aircraft Trajectory Segmentation-based Contrastive Coding: A Framework for Self-supervised Trajectory Representation}, 
  author={Phisannupawong, Thaweerath and Damanik, Joshua Julian and Choi, Han-Lim},
  year={2024},
  eprint={2407.20028},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2407.20028}, 
}

Additional Citations

• If you use ESSA Dataset, please cite the original data provider, [Nilsson et al. 2023].

@article{Nilsson2023,
  title = {Swedish civil air traffic control dataset},
  volume = {48},
  ISSN = {2352-3409},
  url = {http://dx.doi.org/10.1016/j.dib.2023.109240},
  DOI = {10.1016/j.dib.2023.109240},
  journal = {Data in Brief},
  publisher = {Elsevier BV},
  author = {Nilsson,  Jens and Unger,  Jonas},
  year = {2023},
  month = jun,
  pages = {109240}
}

• If you use LSZH Dataset, please cite the original data provider, [Olive et al. 2020].

@article{Olive2020,
  title = {Detection and identification of significant events in historical aircraft trajectory data},
  volume = {119},
  ISSN = {0968-090X},
  url = {http://dx.doi.org/10.1016/j.trc.2020.102737},
  DOI = {10.1016/j.trc.2020.102737},
  journal = {Transportation Research Part C: Emerging Technologies},
  publisher = {Elsevier BV},
  author = {Olive,  Xavier and Basora,  Luis},
  year = {2020},
  month = oct,
  pages = {102737}
}

• The citation for [Opensky] and [Airporal]

@inproceedings{Schafer2014,
  title = {Bringing up OpenSky: A large-scale ADS-B sensor network for research},
  url = {http://dx.doi.org/10.1109/IPSN.2014.6846743},
  DOI = {10.1109/ipsn.2014.6846743},
  booktitle = {IPSN-14 Proceedings of the 13th International Symposium on Information Processing in Sensor Networks},
  publisher = {IEEE},
  author = {Schafer,  Matthias and Strohmeier,  Martin and Lenders,  Vincent and Martinovic,  Ivan and Wilhelm,  Matthias},
  year = {2014},
  month = apr 
}

@misc{MLIT2024Airportal,
  author = {{Ministry of Land, Infrastructure and Transport, South Korea}},
  title = {Airportal},
  year = {2024},
  url = {https://www.airportal.go.kr/},
  note = {Accessed: April 7, 2023},
  howpublished = {Electronic resource}
}