Transportation Research Part C 124 (2021) 102892

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

Aircraft taxi time prediction: Feature importance and

their implications
Xinwei Wang a, Alexander E.I. Brownlee b, *, John R. Woodward a, Michal Weiszer c,
Mahdi Mahfouf d, Jun Chen c, *
a
School of Electronic Engineering and Computer Science, Queen Mary University of London, UK
b
Division of Computing Science and Mathematics, University of Stirling, UK
c
School of Engineering and Materials Science, Queen Mary University of London, UK
d
Department of Automatic Control and Systems Engineering, University of Sheffield, UK

A R T I C L E I N F O A B S T R A C T

Keywords: Taxiing remains a major bottleneck at many airports. Recently, several approaches to allocating
Air traffic management efficient routes for taxiing aircraft have been proposed. The routing algorithms underpinning
Feature importance these approaches rely on accurate prediction of the time taken to traverse each segment of the
Machine learning
taxiways. Many features impact on taxi time, including the route taken, aircraft category, oper­
Prediction
Taxi time
ational mode of the airport, traffic congestion information, and local weather conditions. Working
with real-world data for several international airports, we compare multiple prediction models
and investigate the impact of these features, drawing conclusions on the most important features
for accurately modelling taxi times. We show that high accuracy can be achieved with a small
subset of the features consisting of those generally important across all airports (departure/
arrival, distance, total turns, average speed and numbers of recent aircraft), and a small number
of features specific to particular target airports. Moving from all features to this small subset
results in less than a 1 percentage-point drop in movements correctly predicted within 1, 3 and 5
min.

1. Introduction

The number of flights made globally by the airline industry has increased steadily since the early 2000s and has reached 39 million
in 2019 (International Air Transport Association, 2019). Moreover, it is expected that passenger numbers could double to 8.2 billion
from 2017 to 2037, and importantly, there will be tighter environmental regulations (International Air Transport Association, 2020;
European Commission, 2020a). This trade-off between increasing demand of airport capacity and tighter environmental regulations
has been recognised as one of the grand transport challenges (European Commission, 2020b). To better address the challenge and
limited capacity offered by the existing airport infrastructure, there is an urgent need to develop more intelligent and automated
airport traffic management systems.
Developing airports’ digital asset twins is the way forward, since it provides a high-fidelity simulation platform to identify bot­
tlenecks and potential operational failures. Simply having digital data does not spontaneously move airports towards intelligence and

* Corresponding authors.
E-mail addresses: [email protected] (X. Wang), [email protected] (A.E.I. Brownlee), [email protected] (J.R. Woodward), m.
[email protected] (M. Weiszer), [email protected] (M. Mahfouf), [email protected] (J. Chen).

https://doi.org/10.1016/j.trc.2020.102892
Received 17 April 2020; Received in revised form 8 October 2020; Accepted 19 November 2020
Available online 19 December 2020
0968-090X/© 2020 Elsevier Ltd. All rights reserved.
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 1
Overview of the studies in taxi time prediction.
Authors Year Methodology Features considered in the model Data source Airport Movements

Pujet et al. (2000) 2000 Queuing model Airport operational information, e.g. ASQP and BOS /
scheduled and actual pushback time, PRAS
takeoff/landing time and gate arrival
times, and departure/arrival runways
and capacity
Idris et al. (2002) 2002 Queuing model Airport operational information, e.g. ASQP and BOS 26,302
scheduled and actual pushback time, PRAS
takeoff/landing time and gate arrival
times, and departure/arrival runways
and capacity, downstream restrictions
Balakrishna et al. (2008) 2008 Reinforcement learning 5 features including time spent in the ASPM JFK 254 days
runway queue, the number of
departure/arrival aircraft during
taxiing, average taxi time of the
previous half hour, the time of day
Balakrishna et al. (2010) 2010 Reinforcement learning 8 features including number of flights ASPM TPA ̃132,200
features, average taxi time features,
and time of day
Jordan et al. (2010) 2010 Linear regression 17 features including taxi distance, Runway DFW 4720
number of flights features, airline, Status
runway direction Lights
System
Simaiakis and Pyrgiotis 2010 Analytical queuing model Number of flights features, the ASPM BOS and 99,196
(2010) departure capacity, and time of day EWR
Chen et al. (2011) 2011 Fuzzy rule-based systems 14 features including taxi distance, taxi Airport ZRH 679
turning angle, departure/arrival,
number of flights features, operational
modes
Srivastava (2011) 2011 Linear regression model Taxi distance, number of flights ASDE-X and JFK 43 days
features, average taxi time of the SWAP
previous quarter, severe weather or not
Diana (2013) 2013 Survival and frailty analytical Airport operational information, e.g. ASQP JFK 1250
models block delay, departure/arrival delay
and percentage of capacity utilized,
good weather or not
Ravizza et al. (2013) 2013 Multiple linear regression 15 features including taxi distance, taxi Airports ARN 1340
turning angle, departure or arrival, and
number of flights features and some ZRH
less important factors
Ravizza et al. (2014) 2014 Multiple linear regression, least 16 features including taxi distance, taxi Airports ARN 7607
median squared linear regression, turning angle, departure or arrival, and
support vector regression, M5 number of flights features and some ZRH
model trees and fuzzy rule-base less important factors
systems
Lee et al. (2015) 2015 Linear optimized sequencing, Taxi distance, assigned gate, spot and SARDA CLT 332
linear regression, support vector runway, number of flights features
machines, k-nearest neighbors and
random forest
Lee et al. (2016) 2016 Linear regression, support vector 12 features including terminal SARDA and CLT 246,083
machines, k-nearest neighbors, concourse, gate, spot, runway, ASDE-X
random forest and neural networks departure fix, aircraft model, aircraft
weight, taxi distance, time of day,
number of flights features and
unimpeded taxi time
Lordan et al. (2016) 2016 Linear regression 15 features including gate, runway, Airport BCN 35,858
departure or arrival, number of flights
features
Chen et al. (2017) 2017 Multi-objective fuzzy rule-based same as Ravizza 2013 FR24 MAN 1413
systems
Diana (2018) 2018 Ensemble machine learning, 5 features including departure ASPM SEA 2760
ordinary least-squared and demand/throughput, percentage of
penalized algorithms airport capacity utilised, approach
conditions and runway configurations
Herrema et al. (2018) 2018 Neural networks, regression tree, 42 features including operational Airport CDG ̃1 million
reinforcement learning and information, congestion/capacity level,
multilayer perceptron unimpeded taxi time and number of
departures
Lian et al. (2018) 2018 Airport PEK 17 days
(continued on next page)

2
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 1 (continued )
Authors Year Methodology Features considered in the model Data source Airport Movements

Two improved support vector 6 features including taxi distance,

regression methods, generalized number of flights features, delay and
linear regression, softmax takeoff/pushback times
regression and artificial neural
network
Yin et al. (2018) 2018 Machine learning with a 21 features including surface Airport PVG /
macroscopic network topology instantaneous flow indices, cumulative
flow indices, aircraft queue indices and
slot resource demand indices
Mirmohammadsadeghi 2019 Statistic regression, percentiles 5 features including speed, waiting ASDE-X CLT /
et al. (2019) method and data surveillance time, departure runway and wheels- et al.
on/off

automation. To realise this, airport traffic management systems based on trajectory-based taxiing operations deserve further inves­
tigation (Chen et al., 2015, 2016). Aligned with the European advanced surface movement, guidance and control systems (Interna­
tional Civil Aviation Organization, 2020) and the US next generation air transport system programme (Joint Planning and
Development Office, 2020), airport operations based on trajectory-based taxiing will lead to more accurate and efficient aircraft
ground movements. Furthermore, they are closely interconnected with other airport optimization problems, e.g. runway scheduling,
stand allocation, and airport bus scheduling (Atkin et al., 2007; Weiszer et al., 2015).
In order to realise trajectory-based taxiing operations, accurate taxi time prediction has played an indispensable role. It is not only
important to create more robust schedules and identify choke points between gate and runway for practitioners, but also helps the
government analysts to estimate the optimal airport capacity and evaluate the regulation impacts (Ravizza et al., 2013). Various
modelling techniques have been utilised in the literature to estimate and predict the taxi time, e.g., queuing models (Pujet et al., 2000),
statistical regression approaches (Ravizza et al., 2013), fuzzy rule-based systems (Chen et al., 2011) and machine learning techniques
(Balakrishna et al., 2008). Meanwhile, existing research has adopted different numbers of features (from 5 up to 42) that may affect
taxi time from different data sources, including Airline Service Quality Performance (ASQP) and Preferential Runway Assignment
System (PRAS) (Pujet et al., 2000; Idris et al., 2002), Aviation System Performance Metrics (ASPM) (Balakrishna et al., 2008, 2010;
Diana, 2018), Airport Surface Detection Equipment, Model X (ASDE-X) and Severe Weather Avoidance Programs (SWAP) (Srivastava,
2011; Lee et al., 2016), Spot and Runway Departure Advisor (SARDA) (Lee et al., 2015, 2016), FlightRadar24 (FR24) (Chen et al.,
2017). This leads to a potential risk that certain features related to the taxi time have not been included. For instance, only a few
publications considered the impact of weather conditions on taxi time (Simaiakis and Pyrgiotis, 2010; Diana, 2013; Lee et al., 2015),
and these relevant studies simply considered the weather conditions as severe and fine weather. Meanwhile, some defined features are
only obtained when the taxi procedure is completed, which is not realistic for taxi time prediction in practise (Ravizza et al., 2013).
Moreover, more information can be obtained through digging into the raw data source, e.g., defining new metrics to directly reflect the
surface traffic congestion (Ravizza et al., 2014; Yin et al., 2018).
Another drawback we have observed falls in the feature importance identification process. Although as mentioned, various pre­
diction methodologies have been applied in this domain, only a few of existing research addressed the identification of feature
importance (Jordan et al., 2010; Herrema et al., 2018). Assuming there is no priority information for the features, Jordan et al. (2010)
applied a sub-optimal subset selection method (Pudil et al., 1999) to capture much of the variability in the data source, though with no
guarantee of ensuring high-quality prediction accuracy. Herrema et al. (2018) defined two metrics to rank the candidate features, and
selected the top ten of them as the most important features. Note a sufficient analysis of the selected features number was not provided.
Therefore, there is still a strong need to better understand which features and to what extent the features affect the taxi time. We aim to
develop a feature importance identification procedure with quantitative analysis to address this issue, and better underpin subsequent
applications of the taxi time prediction in uncertain airport runway and taxi scheduling problems.
Consequently, the overarching original contributions of this new study include the following: (1) To the best of our knowledge, a
complete feature group for the taxi time prediction has not been provided; for the first time we introduce a set of weather features,
aircraft speed and runway utilisation information that may affect the taxi time, and compare the prediction performance using
different models and feature groups. (2) A backward feature importance identification process with quantitative analysis is applied to
taxi time prediction. (3) The feature importance identification process integrating with prediction model is validated using real-world
freely-available data for three international airports. (4) We show that the taxi time prediction models only including aircraft de­
parture/arrival, distance, number of existing departure flight, average speed of recent aircraft and specific features for target airports
can provide a high level of accuracy.
This paper is organized as follows: Section 2 provides a comprehensive literature review in the taxi time prediction; Section 3
introduces three international airports and a complete set of features that could impact taxi time; the prediction models, performance
metrics as well as the developed feature importance identification process are described in Section 4, followed by the computational
results and discussions presented in Sections 5 and 6; finally, conclusions are drawn in Section 7, highlighting the important contri­
butions of our work and potential future directions. Appendices listing the abbreviations we use throughout the paper, and giving
standard deviation of the results are also included.

3
X. Wang et al. Transportation Research Part C 124 (2021) 102892

2. Literature review

To clearly present the state-of-the-art of taxi time prediction methodologies, we classify the reviewed literature in line with four
categories: queuing models, statistical regression approaches, Fuzzy Rule-Based Systems (FRBSs) and other machine learning tech­
niques. An overview of the studies in taxi time prediction is listed in Table 1.
The queuing model has been generally utilised in the early research of taxi time prediction. Pujet et al. (2000) and Idris et al. (2002)
analysed real-world data at Boston Logan International Airport (BOS) and identified the runway configuration, the airline/terminal,
the downstream restrictions and the takeoff queue size as the main features that affect the taxi time. Consequently, the queuing models
were developed to further improve the accuracy of taxi time prediction. Simaiakis and Pyrgiotis (2010) modelled the aircraft departure
process as a queuing system, aiming to predict taxi time through analytic taxiway and runway queues approximations. This developed
runway queuing model is stochastic and easily transferred to different airports; it has been validated against real data at BOS and
Newark Liberty International Airport (EWR).
Meanwhile, statistical regression approaches have been applied in the taxi time prediction. Jordan et al. (2010) presented a sta­
tistical linear regression (LR) approach to modelling aircraft taxi time at Dallas/Fort Worth Airport (DFW). Combining with a feature
selection method, the developed model achieved 98.3% prediction accuracy within 3 min absolute error. Notice only data on good
weather days were applied to train and test the model. Building on a historical traffic flow database, Srivastava (2011) established an
adaptive taxi time prediction model with LR analysis, where a set of explanatory variables including aircraft queue position, taxi
distance are included. Using actual data from John F. Kennedy International Airport (JFK), the prediction model has been demon­
strated with high accuracy.
Combining both airport layout and historic taxi time information, Ravizza et al. (2013) presented a taxi time prediction model with
a multiple linear regression (MLR) analysis. Data from Stockholm-Arlanda Airport (ARN) and Zurich Airport (ZRH) was utilized for the
experiments, and the taxi distances, the sum of turning angle, aircraft departures or arrivals and the amount of traffic when the aircraft
is taxiing were identified as the important features for taxi time prediction. Furthermore, various regression approaches including MLR
and least-medium-squared LR were testified in Ravizza et al. (2014). Inspired by Ravizza et al. (2013), Lordan et al. (2016) considered
route- and interaction-specific features and designed a log-LR model for taxi time prediction at Bacrelona-EI Airport (BCN). Experi­
mental results have verified the strong predictive validity of the proposed model, while a sample size covering an extensive airport
operational period is required. Recently, Mirmohammadsadeghi et al. (2019) introduced a regression method, which is being applied
by the Federal Aviation Administration for unimpeded taxi time estimation.
FRBSs (Mendel, 2017), which can offer more explanations of the underlying behavior, were introduced for taxi time prediction by
Chen et al. (2011). The results at ZRH indicate that FRBSs are a valuable alternative to existing statistical methods. Detailed com­
parisons including various regression approaches and FRBSs were conducted in Ravizza et al. (2014), demonstrating the FRBSs
outperformed other approaches in terms of prediction accuracy at ZRH. To address the prediction accuracy as well as associated
uncertainty, Chen et al. (2017) further developed a multi-objective FRBS based approach for taxi time prediction at Manchester Airport
(MAN), in which the structure of the FRBS was simplified and only one predominate rule accounted for one taxi scenario.
In addition to FRBSs, other machine learning techniques also contribute to the taxi time prediction literature. Based on one of the
busiest US airport JFK, Balakrishna et al. (2008) designed a probabilistic framework with reinforcement learning (RL) strategy to
predict the aircraft taxi time. The results indicated the RL estimator is capable to capture the dynamics at challenging airports such as
JFK, while the prediction accuracy on individual flight needs to be improved. Furthermore, Balakrishna et al. (2010) conducted a case
study at Tampa International Airport (TPA), realising 81% prediction accuracy with a standard error of 2 min.
Lee et al. (2015) applied the linear optimized sequencing approach to develop a discrete-event fast-time simulation tool for taxi
time prediction. A data-driven analytical method using four machine learning techniques is introduced for comparison as well. These
methods were evaluated with actual data at Charlotte Douglas International Airport (CLT), and experiments indicate that the
developed simulation tool is competitive. Lee et al. (2016) further considered weather conditions at CLT to improve the taxi time
prediction. However, simulation results indicate that the prediction accuracy has not been improved. The reason could be that the
weather conditions were simply divided into good weather and rainy days, and other weather properties such as wind, visibility and
temperature have not been investigated.
Assuming the taxi time is a function of several features which may not be expressed in existing models, Diana (2013) conducted a
survival and frailty analysis, and revealed the block delay and capacity utilisation percentage would impact the taxi time as well. A
comprehensive comparison of various prediction models was then presented at Seattle International Airport (SEA) (Diana, 2018). The
ensemble machine learning, ordinary least-squared and penalized approaches were testified and the results suggest that no algorithm
outperforms others in all cases, and one should strike a balance between the prediction bias and variance. Herrema et al. (2018)
focused on Neural Networks (NN), Regression Tree (RT), RL and multilayer perceptron (MLP) methods for the taxi time prediction at
Charles de Gaulle Airport (CDG). The top 10 out of 42 features, e.g., unimpeded taxi time, congestion level, and number of departures
in the last 20 min were chosen in the feature selection process, and RT turned out to be the most efficient method.
Comparing with traditional taxi time prediction methods, e.g., LR, softmax regression and NN, Lian et al. (2018) developed two
improved support vector regression (SVR) approaches in a case study of Beijing International Airport (PEK). Several features including
queue length, taxi distance and potential landing number were identified, and a high prediction accuracy up to 95% within 5 min was
achieved. Based on a macroscopic network topology, Yin et al. (2018) formulated the taxi time relevant features into four groups:
surface instantaneous flow, cumulative flow, aircraft queue length and slot resource demand. Three machine learning methods
including LR, SVR and Random Forest (RF) were then applied in the taxi time prediction. Using the historical data at Shanghai Pudong
International Airport (PVG), computational results demonstrated the effectiveness of the proposed features and machine learning

4
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Fig. 1. Sketches of the layout at three airports.

techniques.
Throughout the above literature review, we can identify the gaps in the current research for taxi time prediction as follows: due to
the incomplete data source, several features that may influence the taxi time have not been comprehensively investigated. For
instance, although existing literature noticed the presence of weather conditions, the weather features were simply classified as severe/
fine weather (Simaiakis and Pyrgiotis, 2010; Diana, 2013; Lee et al., 2015). Meanwhile, to ensure taxi time prediction accuracy and
underpin robust airport traffic management, e.g., airport ground movement routing/scheduling and runway scheduling, the feature
importance identification with quantitative analysis requires to be further addressed. We aim to fill these gaps in this research.

3. Data

This study utilises data from three international airports in Europe and Asia: Manchester Airport (MAN), the third biggest airport in
the UK; Zurich Airport (ZRH), the largest airport in Switzerland; Hong Kong International Airport (HKG), ranking among the global top
ten busiest airports. The layout of the three airports are illustrated in Fig. 1. The real-world aircraft movement information is taken
from freely-available data on the website FlightRadar24, following the techniques described in Brownlee et al. (2018, 2020)) (The
tools are available at https://github.com/gm-tools/gm-tools). FlightRadar24, which has also been used to gather airborne flight tracks
(Petersen et al., 2013; Turner et al., 2013; Ptak et al., 2014; Eele et al., 2015), collects automatic dependent surveillance-broadcast
(ADS/B) messages transmitted by many aircraft. These messages contain the latitude, longitude and altitude, usually every 5 to 10
s. The coordinates have a resolution of 10− 4 degrees: approximately 10 m for our target airports. While not all aircraft broadcast ADS/B
data, and of the broadcast data has calibration errors or corruptions needing cleaned before use, enough flight movements are present
to allow reliable taxi time estimation models to be derived. For this work, all tracks for aircraft with an altitude of zero within 5 km of
each airport’s centre were collected. The raw tracks were snapped to the actual taxiways by searching for all taxiways within 10 m of
each coordinate and deriving the most likely route taken, taking the shortest path between the coordinate points from ADS/B except
where those lead to sharp turns. Each movement contained the taxi route taken and the real time at the start and end. The corre­
sponding weather information is extracted from the METAR Weather Service (https://www.aviationweather.gov/metar). We accessed
148 72 movements data for MAN from 21st Jan 2017 to 13rd Apr 2017, 19 808 movements for ZRH from 21st Jan 2017 to 18th Mar
2017, and 42 397 movements for HKG from 15th Jan 2017 to 21st Feb 2017. After snapping these raw movements to the known
taxiways at each airport, the data contained 10 216, 11 271 and 33 095 tracks for MAN, ZRH and HKG respectively; the missing tracks
did not have enough points aligned with the taxiways for the taxi routes to be determined with certainty. These were further reduced to
10 210, 11 248 and 33 060 tracks after removing any taxi routes longer than 45 min (substantial outliers and deemed to be erroneous).
For the purposes of this study, we consider taxi-time to be the actual push-back time to the actual line-up time for departures, and
the actual time of leaving the runway to the actual on-block time for arrivals. The average taxiing times at MAN, ZRH, HKG are 9.6, 6.6
and 11.3 min respectively. Notice the movements without complete information (i.e., full path between runway and stand) have been
removed in advance.
In order to ensure taxi time prediction accuracy, one should comprehensively consider relevant features that may affect taxi time.
In this study, we use up to 33 features, aiming to provide a sufficient set of features for the taxi time prediction. These relevant features
are divided into four categories, including aircraft and airport operational factors, airport congestion, aircraft average speed and
weather information.

5
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 2
Features of aircraft and airport operational information.
Features Type Description

depArr Binary Departure(0) or arrival(1)

distance Numerical Sum of taxi distance in metres
distance_long Numerical Sum of taxi distance in metres on straights of more than 500 metres
angle_sum Numerical Sum of turning angle in degrees
angle_error Numerical Count of 180 degree turns (usually 0, or 1 corresponding to a pushback)
aircraft_weight Categorical Aircraft weight category: small, medium and large
runway_info Categorical Aircraft runway information category
budget_airline Binary Budget (0) or non-budget (1) airline

Table 3
Features of airport congestion information.
Features Type Description

NDepDep Numerical Number of other aircraft on the way to runway when current aircraft pushes back
NDepArr Numerical Number of other aircraft on the way to stand when current aircraft pushes back
NArrDep Numerical Number of other aircraft on the way to runway when current aircraft lands and starts taxiing
NArrArr Numerical Number of other aircraft on the way to stand when current aircraft lands and starts taxiing

Table 4
Features of airport historical congestion information.
Features Type Description

QDepDep Numerical Number of other aircraft that reach runway and depart while current aircraft is on the way to runway
QDepArr Numerical Number of other aircraft that arrive at stand while current aircraft is on the way to runway
QArrDep Numerical Number of other aircraft that reach runway and depart while current aircraft is on the way to stand
QArrArr Numerical Number of other aircraft that arrive at stand while current aircraft is on the way to stand

3.1. Aircraft and airport operational features

Eight aircraft and airport operational features which can be easily extracted from the raw data are listed in Table 2. The binary
feature depArr indicates whether the movement is a departure or arrival flight. Distance is the entire taxi distance from the gate to the
runway; distance_long is the sum of straight taxiway lengths exceeding 500 metres. Two features related to aircraft turning angles that
may affect the taxi speed are included in the prediction model as well. Besides, aircraft weight has been considered as a potential
candidate factor for taxi time prediction (Ravizza et al., 2013; Herrema et al., 2018). In line with the aircraft wake vortex (Hallock and
Holzäpfel, 2018), we introduced feature aircraft_weight to categorise the aircraft as small, medium and large.
The runway_info feature reflects which runway as well as which direction is used for aircraft landing or take-off. For instance, as
shown in Fig. 1(a), MAN has two runways for aircraft departure and arrival. Typically the two runways are both used from 6 am to 10
pm of the day, while only one runway is utilised for the rest of the day. During the dual runway operation, the aircraft normally use
runway 23L/R for arrival and departure, while it might use runway 05L/R when the wind direction changes. Therefore we define 2 ×
4 = 8 runway information categories for MAN, where 2 indicates two runway operation conditions, and 4 the two directions of the two
runways. This identifies the specific runway used by each flight. We use this in contrast to the more course-grained ‘operating mode’ (i.
e. which runways were in use at the time of the flight) as it also serves to indicate which area of the airport in which the aircraft either
began or completed taxiing. feature is introduced to classify the runway utilisation conditions. To distinguish the runway operational
conditions for MAN, we adopt 2 × 4 = 8 operational modes for the feature operation_mode (2 for dual or single operational mode, 4 for
four directions of the two runways). Similarly, ZRH has 6 runway_info types since there are two directions of three runways. For HKG, 4
runway_info modes are categorised given two directions of two runways.
The last feature of aircraft information is budget_airline, aiming to investigate whether budget airline flight would impact taxi time
for economic reasons. In this research, Ryanair and easyJet are categorised as budget airlines, while others as non-budget airlines.

3.2. Airport congestion features

The airport congestion features are first introduced to the taxi time prediction in Ravizza et al. (2013, 2014) inspired by the queuing
model (Idris et al., 2002). These features shown in Tables 3 and 4 have been demonstrated as effective factors to improve taxi time
prediction performance. Eight features are designed to represent the airport congestion conditions via counting the number of arrivals
and departures during the time the current aircraft is taxiing for departure or arrival. Note the features in Table 3 indicate the
congestion information when the current aircraft starts taxiing, while the ones in Table 4 denote the congestion conditions until
current aircraft has completed the movement.
The information of the features in Table 4 can only be obtained when the aircraft has completed its taxiing process. Therefore, these

6
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 5
Features of aircraft average speed information.
Features Type Description

AvgSpdLast5Dep Numerical Average speed of latest departing 5 aircraft

AvgSpdLast5Arr Numerical Average speed of latest 5 arriving aircraft
AvgSpdLast5 Numerical Average speed of latest 5 aircraft
AvgSpdLast10Dep Numerical Average speed of latest departing 10 aircraft
AvgSpdLast10Arr Numerical Average speed of latest 10 arriving aircraft
AvgSpdLast10 Numerical Average speed of latest 10 aircraft

Table 6
Features of weather conditions.
Features Type Description

Pressure Numerical Air pressure in Hg

Temperature Numerical Temperature in Celsius
WindSpeed Numerical Wind speed in metres per second
Visibility Numerical Visibility in metres
isRain Binary Whether it is raining
isSnow Binary Whether it is snowing
isDrizzle Binary Whether it is drizzling
isFog Binary Whether it is fogging
isMist Binary Whether it is misting
isHaze Binary Whether it is hazing
isHail Binary Whether it is hailing

features actually cannot be utilised for practical taxi time prediction. In light of this, we only choose the congestion features listed in
Table 3, excluding the historical congestion information in Table 4.
Given the aircraft queuing length could have more impact on the taxi time of departure flight (Pujet et al., 2000; Ravizza et al.,
2014), we expect features QDepDep and QDepArr could have stronger contributions to the models compared to QArrDep and QArrArr.

3.3. Aircraft average speed features

The taxi speed of other moving aircraft can reasonably be expected to be related to the actual taxi time, acting as a proxy for many
other factors influencing the taxi speed. The idea is that average speed of other aircraft is easy to calculate and might capture con­
founding factors that are difficult to explicitly measure. Thus, it is somewhat surprising that, until now, the features relevant to the
speed of other aircraft have not been studied in the literature. We are the first to introduce the speed features into taxi time prediction,
through defining six average speed features as shown in Table 5, where the units are metre per minute. Notice we collect the latest
average speed for departure, arrival or both of them. This is because that the previous research indicated the airport congestion
conditions of departure aircraft have larger impact on the taxi time (Ravizza et al., 2014). In case the average speed features also have
different impacts on departure and arrival aircraft, we define these speed features for departure and arrival aircraft respectively, better
capturing the influence of speed features on the taxi time. Precisely how many aircraft should be counted for these features is
somewhat arbitrary, so we explored two numbers (5 and 10) to determine whether the number made much difference. Similar to the
airport congestion features, the average speed features related to departure flight, e.g., AvgSpdLast5Dep and AvgSpdLast10Dep, are
expected to have more contributions to the taxi time predictions compared to arrival related features.

3.4. Weather information features

Although existing research did not indicate close correlation between the weather conditions and taxi time (Simaiakis and Pyr­
giotis, 2010; Diana, 2013; Lee et al., 2015), it is expected that the airport local weather conditions should impact the taxi time to some
extent (Ravizza et al., 2013; Herrema et al., 2018; Diana, 2018). To better address and analyse possible influence of the weather
conditions on aircraft taxi time, a list of eleven weather features rather than simply defining bad/fine weather are presented in Table 6.
The abundant weather information is promising to reveal potential relations between the weather conditions and taxi time. However,
as the flight could be postponed or cancelled under extreme weather conditions, e.g., blizzard or heavy rains, the collected aircraft
taxiing data probably does not include terribly bad weather. This could limit the impact of the weather conditions on taxi time
prediction.

4. Prediction models and feature importance

In this section, five prediction models are introduced for predicting the taxi time, and seven performance metrics are adopted to
address and compare the model performance. We also develop a feature importance identification procedure, aiming to identify
important features and provide high prediction accuracy with a narrow subset of features.

7
X. Wang et al. Transportation Research Part C 124 (2021) 102892

4.1. Models

The selection of prediction models could impact the prediction performance. In this study, we introduce five models for taxi time
prediction, specifically MLP, LR, Polynomial Regression (PR), Gradient Boosted Regression Trees (GBRT) and RF.

4.1.1. Multilayer perceptron

NNs are brain inspired models which allow a machine to learn from available data (Haykin, 1994). As a class of feedforward
artificial NN, MLP consists of at least three layers of nodes: an input layer, a number of hidden layers and an output layer. The nodes are
connected between each layer with different weight. Except for the input nodes, each node is a neuron that uses a nonlinear activation
function. For instance, the commonly used activation function is constructed as
y(v) = max(0, v) (1)

where v is the input and y(v) denotes the output.

MLP trains the network by using back propagation (Gardner and Dorling, 1998), which is widely utilised for a general NN.

4.1.2. Linear regression

The concept of LR is that there is a relationship between an independent feature and a dependent one. If the two variables move in
the same direction, then there is a positive relationship between the two variables. On the other hand, if the independent variable
increases and the dependent variable reduces (and vice versa), there is a negative relationship between the two variables.
The general linear regression formula is the following:
̂y = αx + β (2)

where ̂
y is the prediction output, x is the vector of input features, α and β are coefficients vectors.
When fitting this model, we aim to find α and β that minimize the defined cost function, e.g. the ordinary least squares as

1∑ n
min y i )2
(yi − ̂ (3)
2 i=1

where yi and ̂
y i are the ith actual and predictive output values respectively, and n is the number of data samples.

4.1.3. Polynomial regression

PR is a technique of regression analysis, where the relationship between the dependent feature and the independent feature is
described by certain polynomial degrees in the dependent variable. LR is a special case of PR, in which the polynomial order equals
one. This method is beneficial for describing curvilinear relationships, while it may easily be over-fitted. Therefore one should
carefully select the polynomial orders for PR. The model with nth order polynomial regression is constructed as

̂y = β0 + β1 x + β2 x2 + … + βm xm (4)

y is the prediction value, xz = {xz1 , xz2 , …, xzf }(z = 1, 2, …, m) are the zth polynomial input features, f is the number of input
where ̂
features and βi (i = 0, 1, …, m) are regression coefficients.
Similar to LR, the coefficients of the PR model are determined by minimising the cost function, e.g. the ordinary least squares.

4.1.4. Gradient boosting regression trees

GBRT is a flexible non-parametric statistical learning technique for regression. It builds the model in a stage-wise fashion, and
allows optimization of an arbitrary differentiable cost function (Friedman, 2002).
Like other boosting methods, gradient boosting combines weak learners into a single strong learner in an iterative manner. At each
stage s of the gradient boosting, there exists a current prediction model Fs that can be further improved. Through adding an estimator h,
a more accurate prediction model is obtained by constructing Fs+1 (x) = Fs (x) + h(x). This implies that the selection of estimator h can
be expressed as
h(x) = y − Fs (x) (5)

where y is the actual output value.

Specifically, the gradient boosted regression tree, i.e., GBRT, has been widely used due to its efficiency, accuracy and interpret­
ability (Ke et al., 2017), making it a promising approach to accurately predict the aircraft taxi time. GBRT builds a series of regression
trees, statistical models generated for supervised prediction problems. They make their predictions by a series of decisions represented
in a tree structure, in which each node is a split in the possible values for one feature. The benefit of decision tree regression is that it is
easy to interpret and visualise, and can possibly reveal patterns which may be difficult in being identified through traditional
regression methods.

8
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Fig. 2. Feature importance identification with backward elimination.

4.1.5. Random forest

RF is a prediction model made up of many decision trees (Dietterich, 1995; Breiman, 2001). RF can usually generate good results
even without tuning the hyperparameters (Bernard et al., 2009). Each tree in a RF learns from a set of randomly sampled data during
the training process, a technique known as bagging. Notice the samples would be repeatedly applied in a single tree, so that the entire
forest will have lower variance without increasing the bias. The outputs of RF are obtained through averaging predictions of each
decision tree in RF.
Whilst decision trees search for a split on each feature in each node, RF investigates for a split on only one feature in a node. First, a
small subgroup of explanatory features is randomly selected. Next the node is split with the best feature among the small number of
randomly selected features. After splitting, a new list of eligible features is chosen arbitrarily. This process continues until the tree is
completely grown. Ideally, in every terminal node there will be only one observation. As the number of features increases, the eligible
feature set will be quite different from node to node. Nevertheless, significant features finally appear in the tree and their respective
success in prediction will lead to more reliability.

4.2. Performance metrics

In this section, we provide seven prediction performance metrics to evaluate and compare the introduced prediction models.

4.2.1. Accuracy
The accuracy of prediction models measures the relative percentage difference between the predictive and actual values. It is
defined as

100 ∑
n
yi − ̂yi
accuracy = (100 − | |)% (6)
n i=1 yi

where n is the number of data samples.

4.2.2. R2
R2, namely the coefficient of determination, denotes the variation in the dependent variable explained by the independent vari­
ables. The value of R2 is no more than 1, which is the best possible value. Notice it could be negative when a model tries to fit nonlinear
functions to sampled data (Diana, 2018). The mathematical definition of R2 is expressed as
∑n
(yi − ̂ y i )2
R2 = 1 − ∑n 1 ∑ n 2
(7)
1
1 (yi − n 1 yi )

4.2.3. Mean absolute error

Mean Absolute Error (MAE) is defined to measure the average absolute deviations between the predictive and actual values. It is
formulated as

9
X. Wang et al. Transportation Research Part C 124 (2021) 102892

1∑ n
MAE = yi|
|yi − ̂ (8)
n i=1

4.2.4. Root mean squared error

Similar to MAE, Root Mean Squared Error (RMSE) is also an important metric to evaluate the model performance. By squaring the
errors it gives a greater weighting to data points with a larger error. It is characterised as
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑n
RMSE = (yi − ̂ y i )2 (9)
n i=1

4.2.5. Prediction accuracy within 1,3 and 5 min

In practical aircraft taxi time prediction, small prediction errors are tolerable; we are more interested in the prediction accuracy
within the predefined threshold (Ravizza et al., 2014; Lee et al., 2016; Herrema et al., 2018; Yin et al., 2018). In this study, we set the
threshold as 1, 3 and 5 min respectively, and provide the prediction accuracy within the corresponding thresholds.

4.3. Feature importance identification procedure

As discussed in Section 1, previous research mainly focused on the performance of developed prediction models, while little work
has been done on the feature importance identification with quantitative analysis. To better address this issue, we use an effective
backward feature elimination procedure (Bursac et al., 2008) to identify the important features that affect the taxi time. A flowchart of
the proposed feature importance identification process is illustrated in Fig. 2. The idea of backward feature elimination is to select one
with the least importance value at one time (Note the importance value is calculated via the prediction model itself, and the value is
dynamically changed when features are removed), and check whether the termination conditions have been met when removing the
selected feature. If not, the selected feature would be removed and step into choosing another feature; otherwise the procedure ter­
minates and the selected feature with remaining ones would be defined as important.
The core of the procedure is to set proper termination conditions with quantitative analysis, aiming to provide strong confidence in
the feature importance identification. Two termination conditions are provided as follows: (1) check whether current feature
importance is greater than a certain value, where the feature importance value is calculated by the prediction model itself; or (2) check
whether the prediction performance goes down by a certain value, in which the prediction accuracy within 5 min is selected as the
performance metric.

5. Results: model configuration and tuning

In this section, we first present the experiment setup, and then report the performance of different models for taxi time prediction at
three international airports. The goal of this section is to settle on a modelling approach that produces accurate results for the target
application before we focus on our investigation of feature importance.

5.1. Experimental setup

The computational experiments are conducted on a laptop equipped with Intel Core i5 CPU at 2.5 GHz and 8 GB of RAM on a
Windows 10 64-bit OS. The prediction models are all implemented in Python using the Scikit-Learn library. The polynomial order of PR
is selected as 2 according to preliminary experiments. Detailed parameters for MLP, GBRT and RF are set as follows: the activation
function, size of hidden layer, learning rate and maximum iteration number of MLP are Relu, 100, 0.001 and 10,000 respectively. The
loss function, learning rate, number of boosting stages and evaluation criterion of GBRT are least squares regression, 0.1, 1000 and
mean squared error. For RF, the number of trees, evaluation criterion, maximum depth of the tree and minimum number of samples are
150, mean squared error, expanding until all leaves are pure, and 10 respectively. The information of the utilised data is referred to
Section 3, in which 33 features are collected from three international airports, i.e., MAN, ZRH and HKG. 70% of the entire data is used
for training, and the remaining 30% is used for testing. The training data was then split using 10-fold cross validation. To fairly
compare different prediction models, the same seeds were used to randomly generate the subsets for cross-validation. The performance
metrics marked Training presented in the tables below are the average values over the ten folds, and those marked Testing are on the
30% unseen test data.
For convenient performance comparisons on different features, we further divide the introduced features into five groups as fol­
lows. These feature groups will be applied and compared in the following experiments.

• Group A (8 features): Include general aircraft and airport operational information denoted in Table 2.
• Group B (12 features): Add airport congestion information (Table 3) to Group A.
• Group C (18 features): Add aircraft average speed information (Table 5) to Group B.
• Group D (29 features): Add airport local weather information (Table 6) to Group C.
• Group E (33 features): Add airport historical congestion information (Table 4) to Group D.

10
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 7
Results of base line prediction methods.
MAN ZRH HKG

EURO SELF EURO SELF EURO SELF

Accuracy (%) − 24.93 41.28 − 262.33 − 32.30 − 185.83 55.56

R2 − 1.19 0.32 − 1.68 0.38 − 3.16 0.63
MAE 6.59 3.16 7.54 2.74 14.57 3.40
RMSE 7.84 4.37 8.79 4.21 15.82 4.69
< 1 min (%) 8.05 23.41 6.40 30.17 1.07 19.66
< 3 min (%) 24.62 59.27 19.91 71.13 3.65 60.90
< 5 min (%) 39.82 81.01 34.66 84.82 8.14 77.73

Table 8
Prediction models performance comparisons for Manchester Airport. Bold numbers indicate the best performance in terms of the corresponding
metrics.
Model MLP LR PR GBRT RF

Training Accuracy (%) 50.06 63.53 63.72 68.18 69.89

R2 0.30 0.55 0.49 0.58 0.60
MAE 3.30 2.48 2.52 2.38 2.29
RMSE 4.44 3.58 3.79 3.43 3.35
< 1 min (%) 21.10 31.13 30.90 31.97 33.84
< 3 min (%) 56.72 70.77 70.75 72.63 74.38
< 5 min (%) 78.63 87.82 87.89 89.20 89.75

Testing Accuracy (%) 57.29 63.31 64.08 67.04 70.34

R2 0.43 0.56 0.56 0.59 0.61
MAE 2.81 2.46 2.46 2.35 2.25
RMSE 3.98 3.52 3.49 3.40 3.31
< 1 min (%) 26.04 30.90 30.34 31.75 33.80
< 3 min (%) 66.00 71.91 70.93 72.92 75.40
< 5 min (%) 84.31 87.93 88.03 89.85 89.85

Table 9
Prediction models performance comparisons for Zurich Airport. Bold numbers indicate the best performance in terms of the corresponding metrics.
Model MLP LR PR GBRT RF

Training Accuracy (%) 8.25 16.31 27.29 43.29 47.24

R2 0.57 0.60 0.12 0.67 0.67
MAE 2.29 2.07 2.05 1.84 1.76
RMSE 3.49 3.34 4.81 3.06 3.06
< 1 min (%) 34.35 41.73 43.99 48.17 52.15
< 3 min (%) 75.79 78.84 79.74 81.58 83.03
< 5 min (%) 90.00 90.86 91.34 92.64 92.74

Testing Accuracy (%) − 22.15 21.88 29.13 44.65 46.23

R2 0.09 0.61 0.62 0.67 0.67
MAE 4.16 2.12 2.06 1.87 1.78
RMSE 5.24 3.44 3.38 3.15 3.10
< 1 min (%) 8.19 42.55 43.94 47.58 51.74
< 3 min (%) 36.93 78.33 79.30 81.73 82.85
< 5 min (%) 73.71 90.21 90.57 92.58 92.53

5.2. Base line

Before we proceed to our modelling results, as a base line to our experiments, Table 7 presents prediction results using the average
taxi-in and taxi-out time from EUROCONTROL (EURO) and the collected data itself (SELF).The average taxi-in and taxi-out times at
MAN, ZRH and HKG are 6.7/14.3, 5.4/13.5 and 6.6/23.7 min from EUROCONTROL report. The corresponding times from the data
itself are 5.0/11.6, 2.8/9.6 and 4.8/17.2 min.

11
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 10
Prediction models performance comparisons for Hong Kong Airport. Bold numbers indicate the best performance in terms of the corresponding
metrics.
Model MLP LR PR GBRT RF

Training Accuracy (%) 47.65 75.40 79.02 79.63 80.42

R2 0.58 0.82 0.82 0.85 0.84
MAE 3.79 2.19 2.02 1.96 1.96
RMSE 4.99 3.30 3.27 3.03 3.10
< 1 min (%) 18.22 39.54 44.36 46.10 47.87
< 3 min (%) 50.60 75.24 77.61 78.29 77.90
< 5 min (%) 71.67 89.01 90.69 90.73 90.63

Testing Accuracy (%) 45.11 76.43 79.55 80.31 80.80

R2 0.54 0.81 0.83 0.83 0.84
MAE 4.16 2.28 2.08 2.03 1.96
RMSE 5.32 3.45 3.27 3.23 3.10
< 1 min (%) 10.62 38.50 44.18 46.55 48.09
< 3 min (%) 40.52 73.74 76.68 77.36 77.79
< 5 min (%) 73.29 88.40 89.90 90.19 90.53

Table 11
Prediction performance comparisons on different feature groups for Manchester Airport. Bold numbers indicate the best performance in terms of the
corresponding metrics excluding Group E.
Feature Group A B C D E

Training Accuracy (%) 70.07 70.29 70.13 70.09 73.32

R2 0.57 0.59 0.61 0.61 0.70
MAE 2.34 2.29 2.27 2.28 2.04
RMSE 3.48 3.40 3.34 3.34 2.92
< 1 min (%) 34.32 34.39 34.13 34.28 36.84
< 3 min (%) 73.47 74.30 74.54 74.56 77.88
< 5 min (%) 89.28 89.72 89.79 89.76 92.51

Testing Accuracy (%) 70.67 70.89 70.66 70.47 73.69

R2 0.57 0.59 0.60 0.60 0.69
MAE 2.35 2.30 2.30 2.29 2.05
RMSE 3.50 3.43 3.39 3.37 2.96
< 1 min (%) 34.06 34.30 34.25 34.13 37.13
< 3 min (%) 73.75 74.04 73.93 74.60 77.66
< 5 min (%) 89.19 89.67 89.50 89.80 92.33

5.3. Performance of different prediction models

Since Group D contains all available features that are practical for predicting aircraft taxi time, it is utilised for model comparisons
at three international airports. The results are provided in Tables 8–10, in which the best value for each performance metric is marked
in bold.
Overall, RF outperforms other models across the three airports, closely followed by GBRT. RF also has a very similar performance
on the training and testing sets. PR performs better than LR in all metrics for ZRH and HKG, while LR has close MAE and RMSE
compared to PR. This is logical – LR can be viewed as a special PR model with only 1st linear order, and the addition of 2nd order in PR
may sacrifice certain absolute errors to improve the prediction accuracy within time intervals. MLP always has poorest results, and the
performance differences of the training and testing sets are relatively sinificant, specially for ZRH. We note that accuracy for ZRH is
lower than for the other airports, whereas the other metrics vary much less. The results for the baseline suggest that simple averages
yield poorer results at ZRH than the other two airports; in turn suggests a greater variation in times. This makes sense as the runway
entry/exit points are much more spread out than at the other airports. With runway crossings too, it is a considerably more complex
picture so unsurprising that it is harder to model. The taxi times at Zurich are on average a little shorter, so the metric accuracy can
drop while the metrics related to the 1, 3 and 5 min thresholds remain much the same.
We also observe that all prediction models have consistent performance across the three airports, e.g., the variation of prediction
accuracy within 5 min for different airports is less than 5% for each model. In conclusion, RF has the best prediction performance
among the introduced five models, and we apply RF as the prediction model for the following experiments using different feature
groups.

12
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 12
Prediction performance comparisons on different feature groups for Zurich Airport. Bold numbers indicate the best performance in terms of the
corresponding metrics excluding Group E.
Feature Group A B C D E

Training Accuracy (%) 48.92 48.95 47.37 47.24 64.16

R2 0.63 0.66 0.66 0.67 0.84
MAE 1.93 1.78 1.79 1.76 1.33
RMSE 3.24 3.11 3.13 3.06 2.10
< 1 min (%) 49.14 52.28 52.00 52.15 57.13
< 3 min (%) 79.79 82.58 82.70 83.03 89.30
< 5 min (%) 91.54 92.51 92.44 92.74 96.63

Testing Accuracy (%) 48.67 48.01 46.14 46.23 64.37

R2 0.64 0.66 0.65 0.67 0.85
MAE 1.94 1.81 1.82 1.78 1.33
RMSE 3.27 3.16 3.18 3.10 2.15
< 1 min (%) 48.74 52.26 51.66 51.74 58.01
< 3 min (%) 79.42 82.17 82.17 82.85 89.50
< 5 min (%) 91.41 92.29 92.27 92.53 96.69

Table 13
Prediction performance comparisons on different feature groups for Hong Kong Airport. Bold numbers indicate the best performance in terms of the
corresponding metrics excluding Group E.
Feature Group A B C D E

Training Accuracy (%) 76.36 79.53 80.44 80.42 85.86

R2 0.72 0.82 0.84 0.84 0.93
MAE 2.61 2.07 1.96 1.96 1.34
RMSE 4.06 3.27 3.10 3.10 1.99
< 1 min (%) 43.95 47.15 47.94 47.87 55.72
< 3 min (%) 68.13 76.30 77.80 77.90 88.90
< 5 min (%) 81.83 89.26 90.57 90.63 97.31

Testing Accuracy (%) 76.80 79.93 80.79 80.80 86.52

R2 0.73 0.82 0.84 0.84 0.93
MAE 2.60 2.07 1.97 1.96 1.36
RMSE 4.06 3.28 3.11 3.10 2.07
< 1 min (%) 44.21 47.41 48.17 48.09 55.49
< 3 min (%) 68.49 76.03 77.70 77.79 88.32
< 5 min (%) 82.10 89.01 90.42 90.53 97.08

6. Results: feature importance

In this section we investigate the prediction performance with respect to the different features. The goal is to determine which
features are most important for constructing accurate models of taxi time. We begin by using different feature groups. We then
investigate and discuss the taxi time distributions and importance for individual features.

6.1. Performance of different feature groups

We introduced a set of new features in Section 3 that could improve the taxi time prediction performance. To investigate the impact
of different features, we test the RF model on Groups A, B, C, D and E, respectively. Note Group E contains the historical congestion
features which cannot be obtained in practice, and here we use the results on Group E to evaluate and compare to other feature groups.
The prediction performance of RF using different feature groups are reported in Tables 11–13, where the best performance value for
each metric among Groups A, B, C, D is in bold. The prediction results of the training and testing sets using different feature groups are
consistent for each airport. Besides, as shown in Tables 18–20 in Appendix B, the standard deviations of the prediction results have
very small variations across feature groups. There is no apparent trend that the standard deviations increase with adding more features,
indicating the RF model with random parameters has stable performance when considering more features.
It is evident Groups C and D with additional aircraft speed and weather condition features have overall better performance
comparing to Groups A and B. Moreover, the historical congestion features in Group E significantly improved the prediction results in
terms of all performance metrics, indicating the congestion conditions indeed have huge impact on the taxi time. A possible expla­
nation for this is simply that, even with constant traffic levels, the longer the taxi time, the more aircraft will stop or start moving.
When we closely look at the prediction results on Groups B and C, similar performances can be observed. Compared to Group B, the

13
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Fig. 3. Feature importance ranking for the three airports.

prediction accuracy on Group C increases slightly for HKG, while it has almost the same results (or even slightly worse in some metrics)
for MAN and ZRH. Only when aircraft speed and weather condition features are considered together in Group D, does the model have
better performance comparing to Groups A and B.
It demonstrates that the introduced new features, at least some of them, contribute to more accurate taxi time prediction. However,
which and to what extent the features are important to the taxi time remains unclear. Therefore, we further analyse and identify the
feature importance in the next section.

6.2. Feature importance identification

As discussed in Section 4.3, the feature importance value extracted from the prediction model is necessary for the identification
procedure. The RF model is a combination of tree predictors which can provide useful internal estimates, e.g.error, correlation and
feature importance (Breiman, 2001). Specifically, the importance of a feature in RF is computed as the normalized total reduction of
the criterion brought by that feature. It is also known as the Gini importance. Detailed explanations of Gini importance and feature
importance calculation are referred to(Nembrini et al. (2018). We therefore can directly use the feature importance metric from RF,
enabling the proposed feature importance identification procedure.
The feature importance rankings from RF on Group D (29 features) are illustrated in Fig. 3, where the features with importance
value less than 0.01 are omitted. Note the sum of importance value for all features is one. To better understand and evaluate the feature
importance values, some representatives of taxi time distributions with different features are illustrated in Figs. 4–6.

14
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Fig. 4. Taxi time distributions with different features for Manchester Airport.

Clearly depArr, denoting whether the aircraft is departure or arrival, is always the most important. In particular, the importance
value of depArr is up to 0.7 for HKG, indicating whether the flight is departure or arrival has dominant influence on the taxi time. The
same conclusions can be made when we look at the taxi time distributions with depArr, in which the arrival aircraft indeed have an
average shorter taxi time than that of departure flight.
Following depArr, distance, NDepDep (the number of other aircraft on the way to runway when the current aircraft starts departing),
angle_sum, distance_long, and average speed features also have large importance values across the three airports. In line with the
distribution plots with these features, the taxi time indeed varies against corresponding feature values. For example, as the values of

15
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Fig. 5. Taxi time distributions with different features for Zurich Airport.

distance and NDepDep increase, the aircraft mean taxi time also has an increasing trend. Meanwhile, we observe that the importance
values of other congestion related features, e.g.,NDepArr, NArrArr and NArrDep, are less than 0.01 for three airports. These findings
demonstrate that only the congestion information, in particular the departure queuing length, heavily impacts the departure aircraft.
The above findings are in line with the queuing model (Idris et al., 2002) applied to taxi time prediction. Similarly, the average speed
features, especially the ones for departure flights AvgSpdLast5Dep and AvgSpdLast10Dep, have evident impacts on the taxi time. This
observation also reveals that the departure aircraft are easily influenced by the airport operational conditions.
The budget_airline and runway_info features are considered in the prediction model following discussions with airport practitioners.
However, budget_airline only has a small importance value. The distributions of taxi time with budget_airline in Fig. 4(h) confirm this
finding; there is no evident differences of the taxi time distributions between the budget and non-budget airlines. It is interesting to see
that runway_info only has a large importance value for ZRH, which is consistent with its taxi time distributions in Fig. 5(e). This could
be due to the crossing runways at ZRH. As seen in Fig. 1(b), ZRH has three runways and two of them cross each other. Fig. 5 shows that
the taxi times are very different for the modes 32/34 and 14/16. The crossing has as well a small influence for cases where runway 28 is
in use as arrival runway but most flights try to leave the runway before the intersection. Therefore, the runway operational conditions,
namely runway_info, probably have a huge impact on the taxi time for ZRH. The importance value of aircraft_weight ranks 5th for MAN.
Its importance can be verified with the distributions in Fig. 5(g), where the heavy aircraft typically have longer taxi time. This can be
further attributed to the fact that MAN has the largest gradient variation of its taxiways (MAN has a glide slope of 3 degrees equal to
varying 52.3 metres per kilometer, while ZRH and HKG typically have flat taxiways). Heavy aircraft in general require more time to
accelerate in such a case.
As for the weather condition features, although 11 features are introduced in the prediction model, only Temperature has relatively
large importance value for MAN and ZRH, and Pressure has importance value greater than 0.01 for ZRH. These results are also

16
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Fig. 6. Taxi time distributions with different features for Hong Kong Airport.

Table 14
Feature importance identification with importance value (Group F). Bold features are common to all 3 airports, underlined common to
any 2.
No. MAN ZRH HGK

1 depArr depArr depArr

2 distance distance NDepDep
3 angle_sum NDepDep distance
4 AvgSpdLast5Dep Temperature AvgSpdLast5Dep
5 NDepDep runway_info distance_long
6 aircraft_weight angle_sum AvgSpdLast10Dep
7 AvgSpdLast10Dep distance_long angle_sum
8 AvgSpdLast10 AvgSpdLast10Dep
9 distance_long Pressure
10 AvgSpdLast5 AvgSpdLast10Arr
11 AvgSpdLast10Arr AvgSpdLast5
12 AvgSpdLast5Arr AvgSpdLast10
13 Pressure AvgSpdLast5Arr
14 Temperature AvgSpdLast5Dep
15 angle_error

17
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 15
Feature importance identification with prediction performance (Group G). Bold features are common to all 3 airports, underlined
features common to any 2 airports.
No. MAN ZRH HGK

1 depArr depArr depArr

2 distance distance NDepDep
3 angle_sum NDepDep distance
4 AvgSpdLast5Dep angle_sum distance_long
5 AvgSpdLast10Dep Temperature AvgSpdLast5Dep
6 AvgSpdLast10 AvgSpdLast10Dep
7 NDepDep
8 aircraft_weight

Table 16
Prediction performance of RF using selected features.
Airport MAN ZRH HKG

Feature Group D F G D F G D F G

Training Accuracy 70.09 69.90 69.06 47.24 46.57 40.34 80.42 80.14 79.92
R2 0.61 0.60 0.59 0.67 0.67 0.63 0.84 0.84 0.83
MAE 2.28 2.28 2.33 1.76 1.77 1.87 1.96 1.99 2.01
RMSE 3.34 3.34 3.41 3.06 3.08 3.24 3.10 3.14 3.16
< 1 min 34.28 34.18 33.82 52.15 52.21 51.04 47.87 47.60 47.29
< 3 min 74.56 74.54 73.57 83.03 82.83 81.22 77.90 77.49 77.28
< 5 min 89.76 89.72 88.96 92.74 92.59 91.55 90.63 90.11 89.92

Testing Accuracy 70.47 70.31 69.47 46.23 45.54 39.01 80.80 80.48 80.26
R2 0.60 0.60 0.58 0.67 0.67 0.63 0.84 0.84 0.83
MAE 2.29 2.30 2.35 1.78 1.79 1.90 1.96 2.00 2.01
RMSE 3.37 3.37 3.44 3.10 3.12 3.29 3.10 3.15 3.17
< 1 min 34.13 34.22 33.93 51.74 51.81 50.61 48.09 47.65 47.40
< 3 min 74.60 74.18 73.58 82.85 82.65 80.86 77.79 77.35 77.18
< 5 min 89.80 89.74 88.85 92.53 92.34 91.62 90.53 89.95 89.76

consistent with taxi time distribution plots. For instance, as indicated in Fig. 5(d) for ZRH, the temperature range above 0 degree is
prone to a higher density for shorter taxi time. In contrast, when the temperature is below 0 degree, which normally indicates not good
weather, longer taxi time may be required. Notice weather conditions seem not impacting taxi time prediction for HKG. This is
probably due to its dominant depArr feature with a 0.7 importance value; the importance values of other features are more difficult to
be greater than 0.01 for HKG comparing to that for MAN and ZRH.
Next, the feature importance identification procedure with backward elimination manner is applied. The feature importance value
less than 0.01 and prediction accuracy within 5 min are used as termination conditions, respectively. Note the importance value for the
same feature dynamically changes during the backward elimination process.
The identification results with > 0.01 feature importance and < 1% prediction accuracy are shown in Tables 14 and 15 respec­
tively, where features identified as important across the three airports are marked in bold, and the ones identified at two airports are in
underline. As highlighted in Table 14, depArr, distance, NDepDep, angle_sum, distance_long and departure related speed features have
high importance values across the three airports. These features can be identified as generally important features. Besides, several
features are important only for specific airports, e.g., aircraft_weight, Temperature and Pressure for MAN; Temperature, runway_info,
Pressure for ZRH. We regard these features as airport specific ones.
As the termination condition switches to the prediction performance, the number of selected features is reduced as shown in
Table 15. Notice that all features in Table 15 are presented in Table 14. Moreover, depArr, distance and NDepDep are identified as the
most three important features across the three airports.
As for the accuracy provided by the selected features, the prediction results are presented in Table 16. Clearly the performance with
more selected features whose importance values are greater than 0.01 (denoted as Feature Group F) is better than that with fewer
features identified with < 1% accuracy threshold (denoted as Feature Group G). Moreover, it should be noted that the performance of
the RF model using Group F are very close to that using Group D, and even slightly better for certain performance metrics, e.g.,
prediction accuracy within 1 min for MAN and ZRH.
To further investigate which subsets of the available features can provide high prediction accuracy, we remove the termination
conditions in the feature importance identification procedure, iteratively reducing features until only containing the most important
three features, i.e., depArr, distance and NDepDep. The results are presented in Fig. 7, where the red dot denotes the corresponding
prediction accuracy within 5 min with the finally selected features subset. The prediction accuracy remains quite stable during the
first-stage feature elimination process; therefore we aim to select fewer features which can maintain the same accuracy level compared
to Group D. We observe that for MAN, the first dramatic decrease is from 89.64% to 89.17% when removing feature angle_error. The

18
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Fig. 7. Prediction performance with iteratively reducing features until only containing the most important three ones. The Red dot denotes cor­
responding prediction accuracy within 5 min with the finally selected features subset.

feature group identified to provide high prediction accuracy happens to be Group F (15 features) for MAN. When we iteratively reduce
the features for ZRH, the prediction accuracy has very limited variations, until its value decreases to 92.34% when feature angle_error is
removed. Therefore, we need to select the subset including Group F with extra features angle_error and WindSpeed (16 features in total)
for ZRH, so that the high prediction performance can be realised. Similarly, the critical point is found when removing feature Visibility
for HKG, and its corresponding subset consists of Group F and runway_info, NDepArr, Temperature, Pressure, WindSpeed,

19
X. Wang et al. Transportation Research Part C 124 (2021) 102892

AvgSpdLast10Arr, angle_error, AvgSpdLast10, AvgSpdLast5Arr and AvgSpdLast5 (17 features).

In conclusion, we have demonstrated that high prediction accuracy of taxi time can be achieved with a narrow subset of the features
consisting of generally important ones across all airports and specific ones for target airports. These findings have provided quanti­
tative insights for aircraft taxi time prediction.

7. Conclusions

In this article, we have investigated aircraft taxi time prediction models and introduced a set of features that may affect the taxi
time, among which the runway operational mode, budget/non-budget airlines and aircraft speed features are considered for the first
time. We aim to select a prediction model that works best for our target airports, i.e., Manchester Airport, Zurich Airport and Hong
Kong Airport. Five regression models for taxi time prediction are compared and extensive experiments demonstrate that the Random
Forest model outperforms other prediction models for the three airports.
The major focus of our study was a feature importance identification procedure with backward elimination, where the feature
importance value is extracted from the Random Forest model. Quantitative analysis shows that the depArr (aircraft departure or
arrival), distance (sum of taxi distance) and NDepDep (number of other aircraft on the way to runway when current aircraft pushes
back) features are identified as the most important features across the three airports, followed by angle_sum (sum of turning angle),
distance_long (sum of taxi distance on straights of more than 500 metres) and departure related speed features. Moreover, the proposed
feature identification method can define specifically important features for target airports, e.g., the aircraft_weight (aircraft weight
category), Temperature and Pressure for MAN, and runway_info (airport operation mode in line with runway utilisation), Temperature
and Pressure for ZRH. An important result is that high accuracy can be achieved with a narrow subset of the features consisting of
generally important ones across all airports and specific ones for target airports.
Future research for aircraft taxi time prediction can be oriented in several directions. (1) The slope of the airport could have a huge
impact on taxi time for specific airports, e.g. Manchester Airport. How to model the slope factor and define corresponding feature value
remains unexplored. Currently, our hypothesis is that aircraft_weight somehow acts as a good proxy for such information. However, a
more dedicated feature that can capture the slope information may further improve the prediction accuracy for certain airports. (2)
The taxi time prediction underpins practical aircraft routing and scheduling system. To address online decision making requirements,
current taxi time prediction model should be revised to support adaptive prediction with respect to dynamically varying features
information. (3) Improving the feature importance identification procedure is also an avenue for future research. Alternatively we
could apply model compression techniques (Frosst and Hinton, 2017), in which the generated single decision tree with high inter­
pretability is promising to provide better feature importance metrics, while maintaining high prediction accuracy. (4) The range of
machine learning approaches is constantly growing, and it would also be interesting to test alternative models, such as Lasso
Regression, Elastic Net Regression, Regression Trees, Support Vector Regression, k-Nearest Neighbors and Extra Trees Regression. (5)
It would, of course, also be interesting to further extend the features covered. We could explicitly include whether or not an aircraft
crosses a runway during its manoeuvre, or uses single or bidirectional taxiways. Wind direction and humidity could be useful addi­
tional indicators. The present study only covered Winter and Spring months at each airport: expanding the study to an entire year
would also provide further insight into the impact of changing weather. There is certainly room for improving the accuracy of the
models over our results and these directions may offer a means to achieving that.

Author contribution

Xinwei Wang: Methodology, Software, Validation, Investigation, Formal analysis, Writing - original draft, Visualization. Alex­
ander E.I. Brownlee: Data curation, Conceptualization, Methodology, Software, Validation, Investigation, Funding acquisition. John
R. Woodward: Writing - review & editing, Funding acquisition. Michal Weiszer: Data curation, Software, Writing - original draft,
Visualization, Funding acquisition. Mahdi Mahfouf: Methodology, Writing - review & editing, Funding acquisition. Jun Chen:
Methodology, Writing - review & editing, Funding acquisition.

Acknowledgements

This work was funded by the UK Engineering and Physical Sciences Research Council [grants EP/N029577/2, EP/N029496/2, and
EP/N029356/1]. Grateful thanks are also extended to MSc students Nikolaos Prousalis and Anoosid Kiatkamolvong for their assistance
in the experimental studies, and to the anonymous reviewers for their helpful comments and feedback.

Appendix A. Abbreviations

Table 17.

20
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 17
Abbreviations list.
ARN Stockholm-Arlanda Airport

ASDE-X Airport Surface Detection Equipment, Model X

ASPM Aviation System Perfor-mance Metrics
ASQP Airline Service Quality Performance
BCN Bacrelona-EI Airport
BOS Boston Logan International Airport
CDG Charles de Gaulle Airport
CLT Charlotte Douglas International Airpor
DFW Dallas/Fort Worth Airport
EWR Newark Liberty International Airport
FR24 FlightRadar24
FRBSs Fuzzy Rule-Based Systems
GBRT Gradient Boosted Regression Trees
HKG Hong Kong International Airport
JFK John F. Kennedy International Airport
LR Linear Regression
MAE Mean Absolute Error
MAN Manchester Airport
MLP Multilayer Perceptron
MLR Multiple Linear Regression
NN Neural Networks
PEK Beijing International Airport
PR Polynomial Regression
PRAS Preferential Runway Assignment System
PVG Shanghai Pudong International Airport
RF Random Forest
RL Reinforcement Learning
RMSE Root Mean Squared Error
RT Regression Tree
SARDA Spot and Run-way Departure Advisor
SEA Seattle International Airport
SVR Support Vector Regression
SWAP Severe Weather Avoidance Programs
TPA Tampa International Airport
ZRH Zurich Airport

Appendix B. Standard deviation of RF for three airports

Tables 18–20.

Table 18
Standard deviation of RF for Manchester Airport.
Feature Group A B C D

Training Accuracy 0.315 0.332 0.258 0.300

R2 0.009 0.009 0.010 0.010
MAE 0.023 0.026 0.025 0.025
RMSE 0.039 0.037 0.038 0.035
< 1 min 0.662 0.579 0.503 0.506
< 3 min 0.243 0.396 0.413 0.428
< 5 min 0.441 0.244 0.239 0.218

Testing Accuracy 0.413 0.401 0.319 0.259

R2 0.015 0.016 0.019 0.019
MAE 0.039 0.046 0.043 0.044
RMSE 0.071 0.072 0.073 0.076
< 1 min 0.746 0.846 0.758 0.674
< 3 min 0.596 0.752 0.859 0.939
< 5 min 0.401 0.446 0.436 0.416

21
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Table 19
Standard deviation of RF for Zurich Airport.
Feature Group A B C D

Training Accuracy 0.780 0.698 0.768 0.681

R2 0.005 0.008 0.005 0.006
MAE 0.014 0.017 0.015 0.014
RMSE 0.037 0.049 0.046 0.042
< 1 min 0.373 0.434 0.341 0.289
< 3 min 0.373 0.236 0.263 0.225
< 5 min 0.192 0.181 0.147 0.181

Testing Accuracy 1.832 1.856 2.070 1.835

R2 0.009 0.014 0.014 0.017
MAE 0.028 0.021 0.023 0.026
RMSE 0.092 0.095 0.103 0.110
< 1 min 0.915 0.606 0.684 0.687
< 3 min 0.477 0.609 0.350 0.416
< 5 min 0.398 0.427 0.392 0.422

Table 20
Standard deviation of RF for Hong Kong Airport.
Feature Group A B C D

Training Accuracy 0.148 0.241 0.175 0.174

R2 0.005 0.000 0.000 0.000
MAE 0.015 0.008 0.011 0.009
RMSE 0.021 0.014 0.018 0.017
< 1 min 0.180 0.156 0.227 0.167
< 3 min 0.249 0.245 0.155 0.153
< 5 min 0.186 0.130 0.130 0.092

Testing Accuracy 0.330 0.579 0.425 0.431

R2 0.005 0.003 0.003 0.003
MAE 0.021 0.017 0.020 0.019
RMSE 0.038 0.037 0.041 0.039
< 1 min 0.524 0.345 0.399 0.326
< 3 min 0.270 0.360 0.378 0.304
< 5 min 0.330 0.189 0.193 0.201

References

Atkin, J.A., Burke, E.K., Greenwood, J.S., Reeson, D., 2007. Hybrid metaheuristics to aid runway scheduling at london heathrow airport. Transp. Sci. 41 (1), 90–106.
Balakrishna, P., Ganesan, R., Sherry, L., Levy, B.S., 2008. Estimating taxi-out times with a reinforcement learning algorithm. In: 2008 IEEE/AIAA 27th Digital Avionics
Systems Conference, pp. 3.D.3–1–3.D.3–12.
Balakrishna, P., Ganesan, R., Sherry, L., 2010. Accuracy of reinforcement learning algorithms for predicting aircraft taxi-out times: a case-study of Tampa Bay
departures. Transp. Res. Part C: Emerg. Technol. 18 (6), 950–962. Available: http://www.sciencedirect.com/science/article/pii/S0968090X1000029X.
Bernard, S., Heutte, L., Adam, S., 2009. Influence of hyperparameters on random forest accuracy. In: International Workshop on Multiple Classifier Systems. Springer,
pp. 171–180.
Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5–32.
Brownlee, A.E., Weiszer, M., Chen, J., Ravizza, S., Woodward, J.R., Burke, E.K., 2018. A fuzzy approach to addressing uncertainty in airport ground movement
optimisation. Transp. Res. Part C: Emerg. Technol. 92, 150–175.
Brownlee, A., Atkin, J., Woodward, J., Burke, E., 2020. Methods and sources for underpinning airport ground movement decision support systems, University of
Stirling, Tech. Rep. Available: http://hdl.handle.net/1893/30962.
Bursac, Z., Gauss, C.H., Williams, D.K., Hosmer, D.W., 2008. Purposeful selection of variables in logistic regression. Source Code Biol. Med. 3 (1), 17.
Chen, J., Ravizza, S., Atkin, J.A.D., Stewart, P., 2011. On the utilisation of fuzzy rule-based systems for taxi time estimations at airports. In: 11th Workshop on
Algorithmic Approaches for Transportation Modelling, Optimization, and Systems, pp. 134–145. Available: http://drops.dagstuhl.de/opus/volltexte/2011/3273/
.
Chen, J., Weiszer, M., Stewart, P., Shabani, M., 2015. Toward a more realistic, cost-effective, and greener ground movement through active routing–part I: Optimal
speed profile generation. IEEE Trans. Intell. Transp. Syst. 17 (5), 1196–1209.
Chen, J., Weiszer, M., Locatelli, G., Ravizza, S., Atkin, J.A., Stewart, P., Burke, E.K., 2016. Toward a more realistic, cost-effective, and greener ground movement
through active routing: a multiobjective shortest path approach. IEEE Trans. Intell. Transp. Syst. 17 (12), 3524–3540.
Chen, J., Weiszer, M., Zareian, E., Mahfouf, M., Obajemu, O., 2017. Multi-objective fuzzy rule-based prediction and uncertainty quantification of aircraft taxi time. In:
20th International Conference on Intelligent Transportation Systems (ITSC). Yokohama: IEEE, pp. 1–5. Available: http://ieeexplore.ieee.org/document/
8317826/.
Diana, T., 2013. An application of survival and frailty analysis to the study of taxi-out time: a case of New York Kennedy Airport. J. Air Transp. Manage. 26, 40–43.
Available: http://www.sciencedirect.com/science/article/pii/S0969699712001287.
Diana, T., 2018. Can machines learn how to forecast taxi-out time? A comparison of predictive models applied to the case of Seattle/Tacoma International Airport.
Transp. Res. Part E: Logist. Transp. Rev. 119, 149–164. Available: https://linkinghub.elsevier.com/retrieve/pii/S136655451830543X.
Dietterich, T., 1995. Overfitting and undercomputing in machine learning. ACM Comput. Surv. 27 (3), 326–327.

22
X. Wang et al. Transportation Research Part C 124 (2021) 102892

Eele, A.J., Maciejowski, J.M., 2015. Sequential Monte Carlo Optimisation for Air Traffic Management, Cambridge University Engineering Department, Tech. Rep.
CUED/F-INFENG/TR.693.
European Commission, 2020a. Airport capacity and quality. https://ec.europa.eu/transport/modes/air/airports/airport_capacity_and_quality_en (accessed February,
2020).
European Commission, 2020b. Roadmap to a single european transport area. https://www.eea.europa.eu/policy-documents/roadmap-to-a-single-european (accessed
February, 2020).
Friedman, J.H., 2002. Stochastic gradient boosting. Comput. Stat. Data Anal. 38 (4), 367–378.
Frosst, N., Hinton, G., 2017. Distilling a neural network into a soft decision tree, arXiv preprint arXiv:1711.09784.
Gardner, M.W., Dorling, S., 1998. Artificial neural networks (the multilayer perceptron)–a review of applications in the atmospheric sciences. Atmos. Environ. 32
(14–15), 2627–2636.
Hallock, J.N., Holzäpfel, F., 2018. A review of recent wake vortex research for increasing airport capacity. Prog. Aerosp. Sci. 98, 27–36.
Haykin, S., 1994. Neural Networks: A Comprehensive Foundation. Prentice Hall PTR.
Herrema, F., Curran, R., Visser, H., Huet, D., Lacote, R., 2018. Taxi-out time prediction model at charles de gaulle airport. J. Aerosp. Inf. Syst. 15 (3), 120–130.
Available: https://arc.aiaa.org/doi/10.2514/1.I010502.
Idris, H., Clarke, J.-P., Bhuva, R., Kang, L., 2002. Queuing model for taxi-out time estimation. Air Traffic Control Quart. 10 (1), 1–22. Available: https://arc.aiaa.org/
doi/10.2514/atcq.10.1.1.
International Air Transport Association, 2020. International air transport association annual review 2019. https://www.iata.org/en/publications/annual-review/
(accessed February, 2020).
International Air Transport Association, 2020. 20 year passenger forecast. https://www.iata.org/en/publications/store/20-year-passenger-forecast/ (accessed
February, 2020).
International Civil Aviation Organization, 2020. Advanced surface movement guidance and control systems manual. https://www.icao.int/Meetings/anconf12/
Document%20Archive/9830_cons_en%5B1%5D.pdf (accessed February, 2020).
Joint Planning and Development Office, 2020. Concept of operations for the next generation air transport system. https://www.hsdl.org/?abstract&did=747519
(accessed February, 2020).
Jordan, R., Ishutkina, M.A., Reynolds, T.G., 2010. A statistical learning approach to the modeling of aircraft taxi time. In: 29th Digital Avionics Systems Conference,
pp. 1.B.1–1–1.B.1–10.
Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., Liu, T.-Y., 2017. Lightgbm: A highly efficient gradient boosting decision tree. Adv. Neural Inf. Process.
Syst. 3146–3154.
Lee, H., Malik, W., Zhang, B., Nagarajan, B., Jung, Y.C., 2015. Taxi time prediction at charlotte airport using fast-time simulation and machine learning techniques. In:
15th AIAA Aviation Technology, Integration, and Operations Conference. Dallas, TX: American Institute of Aeronautics and Astronautics. Available: http://arc.
aiaa.org/doi/10.2514/6.2015-2272.
Lee, H., Malik, W., Jung, Y.C., 2016. Taxi-out time prediction for departures at charlotte airport using machine learning techniques. In: 16th Aviation Technology,
Integration, and Operations Conference. Washington, D.C.: AIAA, pp. 1–11. Available: http://arc.aiaa.org/doi/10.2514/6.2016-3910.
Lian, G., Zhang, Y., Desai, J., Xing, Z., Luo, X., 2018. Predicting taxi-out time at congested airports with optimization-based support vector regression methods. Math.
Probl. Eng. 2018, 1–11. Available: https://www.hindawi.com/journals/mpe/2018/7509508/.
Lordan, O., Sallan, J.M., Valenzuela-Arroyo, M., 2016. Forecasting of taxi times: the case of Barcelona-El Prat airport. J. Air Transp. Manage. 56, 118–122. Available:
https://linkinghub.elsevier.com/retrieve/pii/S0969699716301570.
Mendel, J.M., 2017. Uncertain rule-based fuzzy systems. In: Introduction and New Directions. Springer, p. 684.
Mirmohammadsadeghi, N., Hotle, S., Trani, A., Gulding, J., 2019. Taxi event extraction from surveillance for surface performance evaluation. J. Air Transp. 1–8.
Nembrini, S., König, I.R., Wright, M.N., 2018. The revival of the gini importance? Bioinformatics 34 (21), 3711–3718.
Petersen, C., Mühleisen, M., Timm-Giel, A., 2013. Evaluation of the aircraft distribution in satellite spotbeams. In: Bauschert, T. (Ed.), Adv. in Comm. Networking,
ser., LNCS, vol. 8115 Springer, pp. 46–53.
Ptak, P., Hartikka, J., Ritola, M., Kauranne, T., 2014. Long-distance multistatic aircraft tracking with VHF frequency doppler effect. IEEE T Aero Elec. Sys. 50 (3),
2242–2252.
Pudil, P., Fuka, K., Beranek, K., Dvorak, P., 1999. Potential of artificial intelligence based feature selection methods in regression models. In: 3rd International
Conference on Computational Intelligence and Multimedia Application. IEEE, pp. 159–163.
Pujet, N., Delcaire, B., Feron, E., 2000. Input-output modeling and control of the departure process of busy airports. Air Traffic Control Quart. 8 (1), 1–32. Available:
https://arc.aiaa.org/doi/10.2514/atcq.8.1.1.
Ravizza, S., Atkin, J.A.D., Maathuis, M.H., Burke, E.K., 2013. A combined statistical approach and ground movement model for improving taxi time estimations at
airports. J. Oper. Res. Soc. 64 (9), 1347–1360. Available: https://www.tandfonline.com/doi/full/10.1057/jors.2012.123.
Ravizza, S., Chen, J., Atkin, J.A., Stewart, P., Burke, E.K., 2014. Aircraft taxi time prediction: comparisons and insights. Appl. Soft Comput. 14, 397–406. Available:
https://linkinghub.elsevier.com/retrieve/pii/S1568494613003384.
Simaiakis, I., Pyrgiotis, N., 2010. An analytical queuing model of airport departure processes for taxi out time prediction. In: 10th Aviation Technology, Integration,
and Operations (ATIO) Conference. Fort Worth, Texas: AIAA. Available: http://arc.aiaa.org/doi/10.2514/6.2010-9148.
Srivastava, A., 2011. Improving departure taxi time predictions using ASDE-X surveillance data. 2011 IEEE/AIAA 30th Digital Avionics Systems Conference, pp.
2B5–1–2B5–14.
Turner, R., Bottone, S., Stanek, C., 2013. Online variational approximations to non-exponential family change point models: with application to radar tracking. In:
Proc. of NIPS, 26, pp. 306–314.
Weiszer, M., Chen, J., Locatelli, G., 2015. An integrated optimisation approach to airport ground operations to foster sustainability in the aviation sector. Appl. Energy
157, 567–582.
Yin, J., Hu, Y., Ma, Y., Xu, Y., Han, K., Chen, D., 2018. Machine learning techniques for taxi-out time prediction with a macroscopic network topology. In: 2018 IEEE/
AIAA 37th Digital Avionics Systems Conference (DASC), London, pp. 1–8. Available: https://ieeexplore.ieee.org/document/8569664/.

23