GraphDAC: A Graph-Analytic Approach to

Dynamic Airspace Configuration

Ke Feng1a , Dahai Liu2b , Yongxin Liu2b , Hong Liu2b , Houbing Song1a
1 University of Maryland, Baltimore County, MD 21250 USA
2 Embry-Riddle Aeronautical University, FL 32114 USA
a {kfeng1,songh}@umbc.edu, b {liu89b, LIUY11, liuho}@erau.edu

Abstract—The current National Airspace System (NAS) is CellGeoSect [12], a cell clustering method, visualizes the
arXiv:2307.15876v1 [math.OC] 29 Jul 2023

reaching capacity due to increased air traffic, and is based airspace as hexagonal cells, maximizes flow connectivity, and
on outdated pre-tactical planning. This study proposes a more balances flight counts between clusters. It then modifies the
dynamic airspace configuration (DAC) approach that could
increase throughput and accommodate fluctuating traffic, ideal design to avoid significant flow’s geometric constraints. How-
for emergencies. The proposed approach constructs the airspace ever, these methods can decrease efficiency as controllers may
as a constraints-embedded graph, compresses its dimensions, and be unfamiliar with newly assigned airports and consequently
applies a spectral clustering-enabled adaptive algorithm to gener- bring safety concerns.
ate collaborative airport groups and evenly distribute workloads
among them. Under various traffic conditions, our experiments This study introduces a practical, adaptive algorithm for
demonstrate a 50% reduction in workload imbalances. This airspace configuration, which modifies existing configurations
research could ultimately form the basis for a recommendation with minimal changes, rather than designing entirely new
system for optimized airspace configuration. Code available at ones. This method entails a three-stage graph-based clustering
https://github.com/KeFenge2022/GraphDAC.git.
method. Firstly, we built simulated airspace from open-source
I. I NTRODUCTION data. We then convert the airspace into a relation graph that
embeds the operational constraints. This is done by only
The National Airspace System (NAS) comprises a complex connecting the geographically adjacent airports and setting
interplay of airports and facilities, ensuring safe and smooth air their edge weights negatively related to their gross workloads.
travel. Air traffic control (ATC), primarily handled by human Secondly, considering that airports within the relation graph
controllers, issues directives to pilots to avoid collisions and are sparsely connected, we then increase the computation
other dangers [1]. In ATC, controllers’ workload is crucial to efficiency by mapping each airport in the relation graph into a
safety [2] and is mitigated by manually dividing and merging low-dimension space, in which Singular value decomposition
airspace into groups with dedicated personnel. This practice (SVD) [13], [14] and Autoencoder [15] are compared. Finally,
is termed Airspace Configuration (AC). we perform a spectral clustering-enabled adaptive algorithm on
Airspace is typically pre-configured according to historical the low-dimensional space to get the new configuration with
plans, with minor alterations by Air Traffic Control (ATC) the traffic-center pattern surrounded by the non-busy airports.
managers [1]–[3]. However, with growing use of advanced The contributions of our work are as follows:
technologies like Unmanned Aircraft Systems (UAS), this
reliance on human expertise may prove insufficient for the • We propose a graph clustering-enabled algorithm for
increasing complexity of airspace dynamics [4]. Furthermore, DAC: we allow configuration change around adjacent
conventional ATC struggles to react swiftly to emergencies, airports to minimize collaboration costs while balancing
causing traffic congestion and delays [5], [6]. As a result, the ATC controllers’ workload.
Dynamic Airspace Configuration (DAC), a real-time, data- • We propose a three-stage graph clustering method: we
driven approach, has gained attention. Unlike static historical construct a graph from airspace with embedded spatial-
models, DAC adjusts to traffic demand while accommodating temporal constraints, then reduce the graph dimension
constraints such as weather, fleet diversity, congestion, and and perform an adaptive clustering to get the final sector
sector complexity [7], [8]. configuration, in which each busy sector is regarded as
Several methods have been proposed for Dynamic Airspace the center and is surrounded by non-busy airports. Our
Configuration (DAC), but their real-world effectiveness is experiments show that our method can reduce sector
debatable. For example, Dynamic Airspace Sectorization [9] workload unbalanced level by over 50% in different
overlooks controllers’ coordination workload and reconfigura- traffic conditions.
tion cost, creating entirely new configurations for each day’s • We investigated the efficacy of reducing the dimensions
segment. SectorFlow [10] groups flight trajectories to mini- of a graph using linear and non-linear methods: SVD
mize airspace complexity, assigning airspace to each cluster. (Singular Value Decomposition) and Autoencoder. Our
Its improved version refines initial partition using gradient findings demonstrate that SVD effectively reduces col-
search and keeps flow intersections off sector boundaries [11]. laboration workload when transitioning to new config-
 urations. On the other hand, the autoencoder excels at re-configuration into the cost function, this approach ensures
minimizing workload imbalances for new configurations. the stability of the configuration by making the busiest airspace
The remainder of this paper is organized as follows: A modules (SBBs) a fixed central component of each controlled
literature review of related work is presented in Section II. airspace sector. Only the generic SAMs change from one
We present the methodology in Section III. Evaluation and configuration to the next. The approach works best when the
discussion are presented in Section IV and conclusions in airspace is divided into small SAMs and SBBs.
Section V. Spectral clustering is applied in [16] and [9] to balance
air traffic controllers’ workload. These works first transform
II. R ELATED W ORK the airspace configuration problem into a graph partitioning
The leading solution for Dynamic Airspace Configuration problem and address it with spectral clustering. The main idea
(DAC) currently involves tactical, dynamic adjustment of of spectral clustering is first to perform eigendecomposition on
airspace to minimize demand and capacity imbalances [3]. the adjacency matrix to extract essential components, then use
This section will summarize commonly employed solutions. the k-means clustering algorithm to divide the airspace. The
First, dynamically creating entirely new controlled airspace. graph uses vertices and links to model airports, waypoints, and
Controlled airspace sector boundaries are newly created every air routes, then project real flight trajectories onto the graph as
time without relying on pre-existing structures. This approach edge weight. The key idea of spectral clustering is to reduce
is also referred to as dynamic sectorization. It is important to the graph dimension to help the clustering algorithm focus on
note that air traffic controllers need the training to work on the most critical feature. However, eigendecomposition is a
a specific airspace set [3]. So that previous work mentioned linear operation that may not have the flexibility to capture
[10], [12] following this path is not desirable operationally the main component of complex air traffic patterns [17].
since ATC personnel cannot familiarize themselves with the In general, the previous works contains the following draw-
newly assigned airports [9]. backs: (a) Omiting the consideration that only close airport
The second approach is to stick to existing building blocks, should collaborate. (b) Involving complicated hyperparameters
e.g., airspace modules, that can be dynamically combined to that are not self-adjusted. (c) The time complexity is high. This
form a controlled airspace sector [9]. Currently, this approach is the primary motivation for our research.
is more desirable operationally. A controlled airspace sector III. M ETHODOLOGY
is operated by a small team of controllers and comprised
of one or more airspace modules. In both the U.S. and A. Problem Definition & Datasets
European airspace, controlled airspace can be combined with We try to find an optimal plan to allocate non-busy airports’
others or split into smaller controlled regions to balance the air traffic control resources to assist busy airports with possibly
workload equally across available ATC resources and airports more delays. Our goal is to balance the workload of different
[3]. Several examples incorporating various constraints are as airports during emergency evacuations or other busy scenarios.
follows: We use a metric called the Regional Unbalanced Level (RUL)
In [1], a promising framework is proposed to realize DAC. to quantify the workload of handling regular or delayed flights.
First, features for evaluating air traffic controllers’ workload We first calculated the average number of non-delayed and
are extracted from the flight radar track, and sector operation delayed flights handled by airports within each cluster, where
history. These feature vectors are fed into a neural network m is the total number of flights in the cluster i:
Pm
to provide a workload indication for the ATC in terms of fk
high, normal, or low. Their algorithm generates different Fi = k=1 (1)
Pmm
new configurations by splitting airports into several smaller dk
airspace modules when the workload is high or merging with Di = k=1 (2)
m
other airports when the workload is low. Next, the tree search We then calculated the variance of the Fi and Di over all
methods explore all possible partitions while restricting them clusters, noted as S. This value helps to quantify the workload
to be operationally valid. This ensures the algorithm builds unbalance within the whole airspace in scope. n is the number
an optimal airspace partition where the workload is balanced of clusters in the configuration, D and F are the mean workload
across the airports and uses the restrictions to lower the of delayed and on-time flights:
reconfiguration cost. Pn
Sergeeva et al. [9] proposed to model dynamic airspace (Di − D)2
SD = i=1 (3)
configuration as a graph partitioning problem that can be Pn n − 1
optimized with a genetic algorithm. They define two different (Fi − F )2
SF = i=1 (4)
types of airspace modules. Those airspace modules that “are n−1
permanently busy areas with a high traffic load” are designated Where high SD or SF indicate the workload of handling
to be “Sector Building Blocks” (SBBs). Less busy and more delayed or regular flights are highly different across different
generic airspace modules are called “Sharable Airspace Mod- clusters. We assume that airports within the same cluster are
ules” (SAMs). A controlled airspace sector should consist of collaborating with each other to handle emerging workloads.
at least one SBB and multiple SAMs. Instead of incorporating Therefore, smaller SD or SF are preferred.
B. Dataset radial-based kernel is chosen to encode the workload into edge
Flight Delays and Cancellations were published by The U.S. weight, defined as below:
Bureau of Transportation in 2015 [18]. This dataset includes di + dj
loadij = (5)
statistics tracking the on-time performance of domestic flights fi + fj
operated by large air carriers. The original data has a total wi,j = B ((1−λ)(100loadij )+λdij −shif t) (6)
of 30 attributes, however, not every attribute was recorded for
each flight, thus columns with more than 25% missing values where i and j are two connected nodes, d is the number of
are removed. For the rest of the data, only the related attributes delayed flights in this time window, f is the total number of
are kept, including the scheduled date, airline, origin and flights in this time window. Thus, loadij is the percentage
destination airport, departure time, and delay time. Canceled of delayed flights of node i and j and normalized between
or diverted flights are removed. To reduce computation load, 0 and 1. We only use the percentage of delayed flights as
departure delay is transformed into the binary label, 0 is a an indicator of the ATC workload, a more comprehensive
non-delay flight and 1 is a delayed flight. We focused on the workload can be explored in the future and substituted here.
21 airports in Florida. dij is the geographical distance between the two airports.
λ is the geographical weight factor in balancing between
C. Hybrid Graph Modeling for Airspace considering geographical distance and gross delay ratio, this
factor is automatically adjusted in our program. B ∈ (0, 1)
Modeling air traffic system in the form of a graph can
is the base, such a base is to satisfy the negative relation
effectively preserve the spatial and temporal information in the
between workload and the edge weight. The sift is set as 300
system [19]. We used a novel data structure Hybrid Airport
as it is numerically more stable. Mathematically, the larger
Adjacency Graph (HAG) to model the airspace of incorporat-
the edge weight between two nodes (airports), the more likely
ing geographical adjacency and workload-based mergeability.
they should be connected to collaborate and form a cluster.
The procedures are as follows:
Step 1: Generating Initial Airport Adjacency Graph (IAG): D. Adaptive Clustering for Pre-Allocation
We identify if two nodes are connectible based on the We developed an adaptive clustering algorithm for par-
geographic location. Mathematically, we define that if two titioning the fine-tuning of the collaboration airports, with
airports, V 1 and V 2 are connected if V 2’s is the closest Spectral Clustering [20] as a key component. Mathematically,
neighbor geographical neighbor of V 1 at the same azimuth. spectral clustering first performs Eigen Decomposition of the
Figure 1 is an example taken from central Florida. When adjacency matrix of the Hybrid Airport Adjacent Graph to
setting the azimuth to 120 degrees, node MCO has three project data from a higher dimension to a lower dimension to
connected nodes which are SFB, TPA, and MLB. Whereas, remove redundancy and noise, then, clustering is done on the
PIE, DAB are not connected to MCO because they are not the low-dimension representation of data [20]. Our procedures are
closest node inside the 120 azimuths. The size of the azimuth as follows:
is a tuning parameter, where a smaller azimuth resulting a Step 1: Construct Hybrid Airport Adjacency Graph with
more dynamic graph, but may result in connecting to a non- λ = 0, in this way, the initial clustering will not consider the
realistic node that is too far away. geographical location of airports.
Step 2: Compress the HAG’s adjacency matrix, we explore
the following methods:
• SVD: we calculate the degree matrix of the graph; The
degree matrix is a diagonal matrix where the value at
entry (i, i) is the degree of the node i. Then calculate the
eigenvalues and eigenvectors of the degree matrix; then
we sort them based on the eigenvalues.
• Autoencoder: we use the encoder to compress each air-
port in the graph (each row) into lower dimensions. This
autoencoder contains two Dense layers for its encoder
Fig. 1. Illustration of Initial Airport Adjacency Graph (IAG). The example
airports are from central Florida. and decoder respectively.
Step 3: We perform the k-means clustering algorithm with
Step 2: Creating Hybrid Airport Adjacency Graph (HAG): an initial k value equal to half of the airports on the low-
We assigned edge weights for each connection of airports. dimension graph to get the initial clustering result.
In HAG, airports that are less busy should have stronger Step 4: Scan each cluster in the initial clustering result, if
connections to each other, and the busy airports should be any cluster contains more than three airports or with a diame-
far away from each other to avoid being clustered together. In ter greater than 100 nautical miles (the typical transmission
other words, the edge weight is negatively related to the total range that aircraft can communicate directly). We increase
workload quantified by the estimated delay flights and delay the number of clusters by 1 and simultaneously, increase the
ratio in the future two hours between the nodes. A modified geographical weight λ by 0.1, but λ can not exceed 0.5.
 Step 5: We repeat the clustering procedure as described in
Step (1) until all clusters satisfy the criteria defined in Step 4.
Steps 4 and 5 make the clustering process adaptive and this
algorithm can gradually evolve to use geographical constraints
to create clusters with reasonable spatial size. Therefore, we
do not require select dedicate k and λ values for different
scenarios.
After spectral clustering on HAG, the airports that are
geographically close and with relatively low workloads are
combined as a new cluster. Simultaneously, the busy airports
with more delayed flights will be picked up and isolated.
E. Fine-Tuning for Dynamic Workload Balancing
In this stage, we aim to merge different airports’ governing
regions to rebalance the workload of the area when there is
an abrupt increase in travel demand or flight delays. For this (a)
purpose, we have the following assumptions:
Assumption I: All flight plans are known at least two hours
in advance from flight plans or predictions.
Assumption II: The abrupt increase in travel demand under
emergency situations could cause significant delays in flights.
In our experiment, if an airport’s delayed flights within the
predicted time window are greater than 2 delayed flights per
hour (with a regional airport) or with a delay percentage
within this time window being 30% (medium or large airports),
we then mark this airport as a busy airport and needs external
assistance.
Assumption III: The nearby airport that used to assist
a busy airport should: (a) have fewer delayed flights if the
category of the airports is identical, or a lower percentage of
delay if the category of the airport is different. (b)
Based on assumptions II and III, we develop the fine-tuning Fig. 2. Airspace configuration generated at different times on 12/24 (a) 7:00-
algorithm for each busy airport as follows: 9:00, (b) 12:00-14:00.
• Step 1: we created a ranked list of busy airports based
on (a) a user-defined priority level with a default value
of zero, (b) the number of delayed flights, (c) delay ratio, scenarios. Additionally, we compare the computational effi-
and (d) number of scheduled flights within the predicted ciency of Autoencoder and SVD for graph compression.
time window. Here, the user-defined priority level can be A. DAC with different scenarios
filled when there’s an emergent situation.
• Step 2: we scanned all airports within 100 nautical
1) Different times on the same day: We tested the DAC
miles of the busy airports in step one and determine if algorithm on December 24th, 2015, a day with heavy flight
a specific airport can be merged to assist an adjacent traffic due to the holiday season. Three distinct 2-hour time
busy airport based on these criteria: (a) distance, (b) windows are selected: 7:00-9:00 for low traffic, 12:00-14:00
less number of predicted delayed flights, (c) lower delay for high traffic, and 19:00-21:00 for medium traffic. The
ratio. Specifically, we created a ranked list of non-busy experiment results show that the algorithm changes the config-
candidates and picked the closest one. uration based on the different traffic conditions successfully,
• Step 3: if any two airports are selected as a collaborative
and the workload unbalanced level of the new configuration
pair, we created a new cluster containing only the two in terms of handling regular and delayed flights has been
airports, to prevent airspace conflict, we also ensure that significantly reduced at all different traffic conditions, as seen
there’s no other busy airport within the combined airspace in Table I.
before establishing the collaboration relationship. As in Figure 2, for low traffic conditions between 7:00-
9:00, the airspace configuration algorithm combines adjacent
In general, the algorithm allows reallocating more resources
sectors to balance the workload, such as MCO with SFB,
from less busy regions in the airspace.
FXE with FLL, and PNS with VPS; In the meantime, several
IV. E VALUATION & D ISCUSSION airports are isolated independently because there is no non-
This section examines the efficiency of the graph clustering- busy airport within a reasonable range and without airspace
based dynamic airspace configuration method under different overlap with busy airports, such as MIA and EYW. When the
 TABLE I
R EDUCTION ON UNBALANCE LEVEL AFTER RECONFIGURATION AT
DIFFERENT TIMES ON THE SAME DAY

Handling regular Handling delayed

flights flights
7:00-9:00 (Low traffic) 42.85% 10.8%
12:00-14:00 (High traffic) 56.9% 61.04%
19:00-21:00 (Medium traffic) 42.86% 60.1%

traffic load is high during 12:00-14:00, the algorithm changes

the configuration to balance the workload among sectors to (a)
adapt to the increasing traffic load.

(b)
Fig. 4. Summary of airspace reconfiguration actions observed on 12:00 PM
(a) to 14:00 PM on different days of 2015: (a) airspaces that are merged. (b)
airspaces that need further separation.

TABLE II
R EDUCTION ON UNBALANCE LEVEL AFTER RECONFIGURATION
AT THE SAME TIMES ON DIFFERENT DAYS

High traffic dates Handling regular Handling delayed

flights flights
7/3 56.77% 56.37%
(b) 11/25 59.1% 62.5%
Fig. 3. Summary of airspace reconfiguration actions within Dec 24th 2015 12/31 56.9% 59.9%
(a) Merged airspace. (b) Airspaces that need further separation.

We also noticed some common patterns. First, the airspace Low traffic dates Handling regular Handling delayed
merging actions in the three time periods are summarized in flights flights
Figure 3(a). The three most frequent merges are FLL-FXE, 2/17 64.1% 66.7%
MCO-SFB, and TPA-PIE. These airports are selected to merge 11/25 60% 67.1%
into collaborative pairs because a) executive airports are not 12/31 53% 44.5%
usually as busy as large international airports even if they
are closely located, so they can always assist busy airports.
Secondly, Figure 3(b) shows that MIA and its nearby airports There are also some common merging strategies employed
are extremely busy all day round, which makes it impossible for both high and low traffic conditions. As shown in figure
to find a collaborative airport that is less busy. 4, the actions taken most frequently are the merging between
2) Same time for different days: We also evaluate the FLL-FXE, MCO-SFB, PGD-RSW and PIE-TPA, showing that
algorithm on different dates, three high-traffic volume dates are large international airports are always busy all year round and
selected, which are 7/3, 11/25 (one day before Thanksgiving), need assistance from other regional or executive airports. GNV
and 12/31 (one day before the new year) in 2015. we also is surrounded by medium and busy airports and is dynamically
compare the airspace configuration results on the busy hours assigned as a flexible collaborator. Our algorithm also indicates
(12:00 PM to 14:00 PM) on these low-traffic days: 2/17, 6/9, that some busy airports in southern Florida are surrounded by
and 9/8, 2015. Our algorithm significantly decreases unbalance airports that are busy simultaneously. Consequently, they can
level of the ATC’s workload in both busy and non-busy not find collaborative peers and further divisions are needed
scenarios as shown in table II. in their internal airspace.
B. Comparison of pre-clustering graph data compression AEs can leverage past training experiences, unlike SVD which
methods starts computation from scratch each time.
This is because the computation of trained AEs can easily
SVD and Autoencoder (AE) are employed to reduce the get accelerated by hardware. Also, SVD has to start compu-
dimension of the HAG’s adjacency matrix. In general, both tation from scratch every time while AEs can leverage past
techniques yield low-dimension graphs that preserve a sig- experiences from training.
nificant amount of essential information, but each has its
own advantages. Specifically, SVD excels in preserving the
TABLE III
embedded constraints within the graph, whereas the AE is C OMPARISON OF TIME REQUIRED IN REAL - TIME APPLICATION BETWEEN
particularly adept at achieving balanced clustering. In Figure SVD AND AUTOENCODER
6, the clustering results of the SVD-compressed graph adhere
HAG’s dimension SVD prediction Autoencoder Autoencoder
to the adjacency constraint that we only want to merge (10−5 second) prediction (10−5 Training time
geographically close airports. Comparably, an AE allows the second) (second)
algorithm to deviate slightly. For instance, when using the 7*7 43.9 1.16 27.78
AE depicted in (a) for compression, certain clusters, such as 10*10 44.8 1.70 27.04
15*15 44.5 4.06 69.47
clusters 0 and 7 contain airports that are not directly connected. 18*18 48.1 3.30 64.67
To compare the impact of the dimension of the latent space 21*21 373.8 3.54 65.98

to the strictness of AE, multiple AEs were trained with varying

latent dimensions (2, 5, 10 and 15). We found that there is a
certain pattern when AEs do not follow the restrictions. As in V. C ONCLUSION
figure 5, AEs always mistakenly combine ECP, TLH and DAB We’ve introduced an innovative method for dynamic
together or VRB, PBI, MIA, and EYW together, but this kind airspace reconfiguration, using a graph model to balance air
of error can easily be fixed by a post-processing algorithm. traffic controllers’ workload. This model incorporates geo-
Additionally, AEs are much faster than SVD after training, graphical adjacency and ATC workloads, utilizing a spectral
making them more suitable for use in real time. As in table clustering-enabled adaptive algorithm to generate new con-
III, as the dimension of the graph continues to increase, SVD figurations based on predicted delay and flight plans. The
is much slower than AEs. This is because the computation algorithm groups high-workload airports as centers surrounded
of trained AEs can easily get accelerated by hardware. Also, by lesser-engaged airports, redistributing ATC resources for

7-9 12-14 19-21

2 dim

10 dim

Fig. 5. Error patterns of autoencoder in pre-clustering graph data compression. The result of the first row and second row are generated by AE when the
latent space is 2 dimensions and 10 dimensions respectively.
 R EFERENCES
[1] D. Gianazza, “Forecasting workload and airspace configuration with
neural networks and tree search methods,” Artificial intelligence, vol.
174, no. 7-8, pp. 530–549, 2010.
[2] P. Lee, J. Mercer, B. Gore, N. Smith, K. Lee, and R. Hoffman,
“Examining airspace structural components and configuration practices
for dynamic airspace configuration,” in AIAA Guidance, Navigation and
Control Conference and Exhibit, 2008, p. 7228.
[3] S. Zelinski and C. F. Lai, “Comparing methods for dynamic airspace
configuration,” in 2011 IEEE/AIAA 30th Digital Avionics Systems Con-
ference. IEEE, 2011, pp. 3A1–1.
[4] J. Wang, Y. Liu, and H. Song, “Counter-unmanned aircraft system (s)(c-
uas): State of the art, challenges, and future trends,” IEEE Aerospace
and Electronic Systems Magazine, vol. 36, no. 3, pp. 4–29, 2021.
[5] Y. Yang, J. Yu, D. Liu, S.-A. Lee, S. Namilae, S. Islam, H. Gou, H. Park,
and H. Song, “Multiagent collaboration for emergency evacuation using
reinforcement learning for transportation systems,” IEEE Journal on
Miniaturization for Air and Space Systems, vol. 3, no. 4, pp. 232–241,
(a) 2022.
[6] Y. Yang, K. Zhang, H. Song, and D. Liu, “Machine learning-enabled
adaptive air traffic recommendation system for disaster evacuation,” in
2021 IEEE/AIAA 40th Digital Avionics Systems Conference (DASC).
IEEE, 2021, pp. 1–8.
[7] P. Kopardekar, K. Bilimoria, and B. Sridhar, “Initial concepts for
dynamic airspace configuration,” in 7th AIAA ATIO Conf, 2nd CEIAT
Int’l Conf on Innov and Integr in Aero Sciences, 17th LTA Systems Tech
Conf; followed by 2nd TEOS Forum, 2007, p. 7763.
[8] P. Kopardekar, K. D. Bilimoria, and B. Sridhar, “Airspace configuration
concepts for the next generation air transportation system,” Air Traffic
Control Quarterly, vol. 16, no. 4, pp. 313–336, 2008.
[9] M. Sergeeva, D. Delahaye, C. Mancel, and A. Vidosavljevic, “Dynamic
airspace configuration by genetic algorithm,” Journal of traffic and
transportation engineering (English edition), vol. 4, no. 3, pp. 300–314,
2017.
[10] C. R. Brinton and S. Pledgie, “Airspace partitioning using flight clus-
tering and computational geometry,” in 2008 IEEE/AIAA 27th Digital
Avionics Systems Conference. IEEE, 2008, pp. 3–B.
[11] C. Brinton, J. Hinkey, and K. Leiden, “Airspace sectorization by dynamic
density,” in 9th AIAA Aviation Technology, Integration, and Operations
(b)
Conference (ATIO) and Aircraft Noise and Emissions Reduction Sym-
Fig. 6. Comparison of pre-clustering compression method on the final result posium (ANERS), 2009, p. 7102.
of DAC (a) Nonlinear: Autoencoder and (b) Linear: SVD. [12] G. Sabhnani, A. Yousefi, and J. S. Mitchell, “Flow conforming op-
erational airspace sector design,” in 10th AIAA Aviation Technology,
Integration, and Operations (ATIO) Conference, 2010, p. 9377.
[13] U. Von Luxburg, “A tutorial on spectral clustering,” Statistics and
workload equilibrium. The model outperforms static airspace computing, vol. 17, pp. 395–416, 2007.
configurations, reducing workload imbalance by over 55% [14] L. Galluccio, O. Michel, P. Comon, and A. O. Hero III, “Graph based
under high traffic volume, as confirmed through evaluations k-means clustering,” Signal Processing, vol. 92, no. 9, pp. 1970–1984,
2012.
during various time windows and traffic conditions. Our real- [15] Y. Wang, H. Yao, and S. Zhao, “Auto-encoder based dimensionality
data simulations indicated that Miami and Sarasota airspace reduction,” Neurocomputing, vol. 184, pp. 232–242, 2016.
requires further partitioning to improve performance, as our [16] J. Li, T. Wang, I. Hwang, and I. Hwang, “A spectral clustering based
algorithm for dynamic airspace configuration,” in 9th AIAA Aviation
DAC algorithm struggled to locate less busy airports within Technology, Integration, and Operations Conference (ATIO) and Aircraft
60 miles for collaboration. Furthermore, Key West airport’s Noise and Emissions Reduction Symposium (ANERS), 2009, p. 7056.
remote location hindered collaboration with busy airports to [17] L. Gondara, “Medical image denoising using convolutional denoising
autoencoders,” in 2016 IEEE 16th international conference on data
share emergent workloads. mining workshops (ICDMW). IEEE, 2016, pp. 241–246.
Our future direction includes improving airspace ATC work- [18] D. o. Transportation, “2015 flight delays and cancellations,” Feb 2017.
load assessment by considering metrics beyond delayed and [Online]. Available: https://www.kaggle.com/datasets/usdot/flight-delays
[19] Y. Jiang, S. Niu, K. Zhang, B. Chen, C. Xu, D. Liu, and H. Song,
total flights. Additionally, we plan to explore neural networks’ “Spatial–temporal graph data mining for iot-enabled air mobility predic-
potential in generating comprehensive airspace configuration tion,” IEEE Internet of Things Journal, vol. 9, no. 12, pp. 9232–9240,
plans. 2021.
[20] A. Ng, M. Jordan, and Y. Weiss, “On spectral clustering: Analysis and an
algorithm,” Advances in neural information processing systems, vol. 14,
ACKNOWLEDGMENT 2001.
This research was supported by the Center for Ad-
vanced Transportation Mobility (CATM), USDOT Grant No.
69A3551747125, 270128BB(AWD00237), and by the U.S.
National Science Foundation under Grant No.2231629, Grant
No.2142154, No.2142514 and Grant No.2309760.