Astronomy & Astrophysics manuscript no.

main ©ESO 2024

November 6, 2024

Tuning into spatial frequency space

Satellite and space debris detection in the ZTF alert stream.
J. P. Carvajal1, 2,⋆ , F. E. Bauer1, 2, 3 , I. Reyes-Jainaga4 , F. Förster2, 5, 6 , A. M. Muñoz Arancibia2, 5 , M.
Catelan1, 2, 7 , P. Sánchez-Sáez8, 2 , C. Ricci9, 10 , and A. Bayo8

1
Instituto de Astrofísica, Pontificia Universidad Católica de Chile, Av. Vicuña Mackenna 4860, 7820436 Macul, Santiago, Chile
2
Millennium Institute of Astrophysics (MAS), Nuncio Monsenor Sótero Sanz 100, Providencia, Santiago, Chile
3
Space Science Institute, 4750 Walnut Street, Suite 205, Boulder, Colorado 80301, USA
4
Data Observatory, Av. Eliodoro Yáñez 2990, oficina A5, Providencia, Chile
5
arXiv:2411.03258v1 [astro-ph.IM] 5 Nov 2024

Center for Mathematical Modeling (CMM), University of Chile, AFB170001, Santiago, Chile
6
Data & Artificial Intelligence Initiative (ID&IA), University of Chile, Santiago, Chile
7
Centro de Astro-Ingeniería, Pontificia Universidad Católica de Chile, Av. Vicuña Mackenna 4860, 7820436 Macul, Santiago, Chile
8
European Southern Observatory, Karl-Schwarzschild-Strasse 2, 85748 Garching bei München, Germany
9
Instituto de Estudios Astrofísicos, Facultad de Ingeniería y Ciencias, Universidad Diego Portales, Av. Ejército Libertador 441,
Santiago, Chile
10
Kavli Institute for Astronomy and Astrophysics, Peking University, Beijing 100871, China
Received September 15, 1996; accepted March 16, 1997

ABSTRACT

Context. A significant challenge in the study of transient astrophysical phenomena is the identification of bogus events, among which
human-made Earth-orbiting satellites and debris remain a key contaminant. Existing pipelines effectively identify satellite trails but
can miss more complex signatures, such as collections of dots known as satellite glints. In the Rubin Observatory era, the scale of the
operations will increase tenfold with respect to its precursor, the Zwicky Transient Facility (ZTF), requiring crucial improvements in
classification purity, data compression for informative alerts, pipeline speed and other areas.
Aims. We explore the use of the 2D Fast Fourier Transform (FFT) on difference images as a tool to improve satellite detection machine
learning algorithms.
Methods. Adopting the single-stamp classification model from the Automatic Learning for the Rapid Classification of Events
(ALeRCE) broker as a baseline, we adapt its architecture to receive a cutout of the FFT of the difference image, in addition to
the three (science, reference, difference) ZTF image cutouts (hereafter stamps). We study different stamp sizes and resolutions for
these four channels, aiming to assess the benefit of including the FFT image, particularly in scenarios where data compression re-
mains essential due to alert size limitations and pipeline processing speed requirements (e.g., for surveys like the Legacy Survey of
Space and Time).
Results. The inclusion of the FFT improved satellite detection accuracy, with the most notable increase observed in the model with
the smallest field of view (16′′ ), where accuracy rose from 66.9% to 79.7% (a statistically significant improvement of ≈13% with a
95% confidence interval of 7.8% to 17.8%). This result demonstrates the effectiveness of FFT in compressing relevant information
and extracting features that characterize satellite signatures in larger difference images.
Conclusions. We show how FFTs can be leveraged to cull satellite and space debris signatures from alert streams.
Key words. Techniques: image processing – Techniques: photometric – Methods: data analysis – Surveys.

1. Introduction of Space and Time (LSST, LSST Science Collaboration et al.

2009; Ivezić et al. 2019) from the Vera C. Rubin Observatory
Large étendue telescopes engaging in time-domain surveys, such will be an order of magnitude larger, with a single-visit depth
as the Zwicky Transient Facility (ZTF, Bellm et al. 2018; Gra- almost 4 magnitudes deeper in the r band compared to ZTF.
ham et al. 2019), have quickly become a cornerstone of obser- A handful of so-called community alert brokers have dedicated
vational astrophysics. ZTF has revolutionized our ability to de- themselves to the distribution, filtering, and annotation of these
tect and characterize a wide variety of transients, active galax- massive alert volumes. These brokers ingest the alert streams
ies, and variable star phenomena (e.g., Chen et al. 2020; Per- and provide processed science products and services that enable
ley et al. 2020; Carrasco-Davis et al. 2021; Sánchez-Sáez et al. the community to pursue their science goals without overloading
2021, 2023), through both its 5-σ difference-image alert stream the observatories’ infrastructure. Current full-stream community
and frequent data releases (Masci et al. 2019). Overall, ZTF dis- brokers for ZTF and future LSST include ALeRCE (Förster
tributes up to 106 alerts each night (with a median closer to et al. 2021), AMPEL (Nordin et al. 2019), ANTARES (Math-
∼200 000) which equates to tens of GBs of data per night (Pat- eson et al. 2021), Babamul, Fink (Möller et al. 2020), Lasair
terson et al. 2018). The expected rates for the Legacy Survey (Smith et al. 2019), and Pitt-Google1 .
⋆ 1
Corresponding author: [email protected] https://pitt-broker.readthedocs.io/en/latest/

Article number, page 1 of 8

 A&A proofs: manuscript no. main

Science Image Difference Image FFT Image Science Image Difference Image FFT Image
ZTF19abcqppk

ZTF21abcjklk
(e) Satellite
(a) VS
ZTF20acogywb

ZTF21acaigul
(f) Satellite
(b) SNe

ZTF21abkkbsc
ZTF18achatrx

(g) Satellite
(c) Asteroid
ZTF18acgkbtq

ZTF21acjcosq
(h) Satellite
(d) Bogus

Fig. 1: Comparison of science image, difference image, and its FFT for various transient sources and contaminants in the ZTF alert
stream. The full 63′′ × 63′′ science and difference stamps are shown, with the central 16′′ × 16′′ denoted in red. The FFT image
stamp is cropped into the central 16 × 16 pixels. Panels (a)-(c) display astrophysical events: (a) variable star ZTF19abcqppk; (b)
supernova ZTF20acogywb; and (c) asteroid ZTF18achatrx. Panel (d) shows a typical bogus event (ZTF18acgkbtq), characteristic
of a bad subtraction. Panels (e)-(h) highlight different types of satellite signatures that reached the alert stream: (e) and (f), from
alerts ZTF21abcjklk and ZTF21acaigul, show the typical regular glint signature with different separation among glints; (g) shows a
satellite with irregular glints and an overall asymmetric signature; (h) shows a satellite with both a continuous signature and glints.
This small portion of Fourier space effectively captures the distinct extended signals of satellites.

One of the persistent challenges is the identification and is slated to go from 63′′ × 63′′ at a 1′′ /pixel resolution for ZTF
removal of contaminants (often referred to as “bogus” detec- (e.g., Fig. 1) to 6′′ × 6′′ at a 0.2′′ /pixel resolution for LSST, due
tions) from the observations and the transient event alert stream. to bandwidth limitations associated with increased image res-
Mitigation of bogus events is particularly important for robust olution. An upside of the increased resolution is highlighted in
early transient discovery and classification, such as identifying recent work by Tyson et al. (2024), which shows how glints from
flash-ionization episodes in young supernovae (SNe) or short- tumbling low-Earth orbit debris may become identifiable due to
lived kilonovae; this was a prime motivation for the development distinct defocus and shape characteristics. This, however, does
of the real-time first-stamp classifier by ALeRCE (Carrasco- not completely alleviate the concerns, as glint-producing satel-
Davis et al. 2021, hereafter, the “Stamp Classifier”). Notably, lites occupy a wide range of orbits (Karpov & Peloton 2023).
human-made satellites and space debris orbiting Earth can be As discussed in Reyes-Jainaga et al. (2023, hereafter RJ23), a
particularly complicated contaminants to identify in real-time promising approach to mitigate the FoV reduction in LSST is to
for existing algorithms and pipelines. Karpov & Peloton (2022, employ multiscale image stamps. These stamps capture a larger
2023) quantified the effect of satellite glints (see Fig. 1) that by- FoV with decreasing spatial resolution as the distance from the
passed both the streak masking algorithm (Laher et al. 2014) and source center increases, effectively compressing data by preserv-
real-bogus classifier (Duev et al. 2019a) components of ZTF’s ing high-detail information in the central region while also re-
pipeline (Masci et al. 2019), finding at least ≈3000 glint-related taining relevant contextual information, such as flux from satel-
alerts per month. For reference, this is ≈4.4 times the monthly lite trails, around the alert.
rate of new transients detected by ZTF and reported to the Tran- Motivated by the potential for image compression and by
sient Name Server (TNS)2 . the unique glint signatures of satellites and space debris (here-
This challenge is already a significant in the ZTF real-time after, simply "satellites"), we explore the spatial frequency do-
alert stream and stands to become even more problematic in the main of the difference images via the Fast Fourier Transform
LSST era due to the shift in stamp field of view (FoV), which (FFT) algorithm (Cooley & Tukey 1965). Satellite glints, pre-
sumably caused by rotating objects that reflect sunlight, exhibit
2
https://www.wis-tns.org/ a range of patterns from simple (e.g., Fig. 1 (f)) to complex (e.g.,
Article number, page 2 of 8
 J. P. Carvajal et al.: Tuning into spatial frequency space

Fig. 1 (g)). Additionally, multiple glints from a single satellite the source to lie in the center of the stamp complicates such
may not be captured within the FoV of a single stamp (e.g., modifications. An attempt to “repair” these cases has yet to
see Sect. 4.2). These glints generally consist of a periodic sig- be implemented.
nal, dominated by the satellite’s rotation, aligned along an ap- 2. In each of the stamps, a boolean integrity mask is con-
proximately straight line within the locality of the image. The structed, where infinite and NaN (Not a Number) values in
signal is convolved with the Point Spread Function (PSF), and the stamp are associated with a True (1), and finite values
under this locality approximation, we can assume a relatively with a False (0).
constant illumination geometry. Therefore, inspecting the spa- 3. The non-finite values in the masked stamp are then set to
tial frequencies of these satellite glint variations is a natural path zero (0).
to explore. Central, point-like emission maps to very extended 4. For computational optimization, the normalization is per-
scales in Fourier space, while extended emission maps to more formed in batches for each type of stamp. The maximum
compact scales. Thus, in contrast with the multiscale approach of for the normalization is the 99th percentile of the absolute
RJ23, the central portions of the Fourier map naturally tap into values. Then, the values are clipped to 2.0.
the extended emission, while imaging cropping effectively limits 5. The normalized stamps are then concatenated with the in-
the minimum scales probed, rather than the maximum ones. tegrity mask, ending up with a 6-channel image. When em-
Building upon the approach in RJ23, we analyze the effect of ploying the Multiscale model, this process is repeated across
including FFT information within the adapted architecture of the four different scales, resulting in a 24-channel image.
ALeRCE Stamp Classifier. The incorporation of the FFT slightly 6. Ultimately, the stamps are cropped or resized, depending on
increases the model’s input size3 , but we aim to assess its poten- the model they will go into.
tial for capturing extended spatial features in a compact form.
We present the methods and results in Sects. 2 and 3, respec- Beyond the preprocessing steps used in RJ23, we introduce
tively. In Sect. 4 we analyze the results and discuss the future new steps (highlighted in the red dashed box in Fig. 2) to incor-
prospects for the FFT, and how it can efficiently expand opportu- porate the FFT calculation. Specifically, these steps are:
nities for improving streak detection algorithms (e.g., Duev et al.
7. Starting from the complete (63 × 63 pixels) normalized dif-
2019b) to perform better on satellite glints while keeping inputs
ference image from step 5, we pad it to 128 pixels, filling
small, which will be fundamental in the LSST era both for inter-
with the median. The size of 128 pixels is chosen because
nal pipelines and downstream filtering.
the FFT algorithm is most efficient for dimension sizes that
are powers of two.
2. Enhancing the ALeRCE Stamp Classifier with 8. We then apply the two-dimensional FFT and take the norm
FFT to obtain the amplitude.
9. Finally, we crop the central portion of the FFT. In our exper-
The ALeRCE Stamp Classifier was first introduced in Carrasco- iments, we kept the central 16 × 16 px region in all cases.
Davis et al. (2021), adopting a Convolutional Neural Network
(CNN) model that classifies individual ZTF alerts based on
three-channel image cutouts (science, reference, and difference) Apply FFT Crop
and additional metadata (e.g., coordinates, flux and magni-
Addition

tude measurements, and seeing) into four distinct astrophysical

sources and bogus. The effectiveness of this machine learning
(ML) model is demonstrated by ALeRCE’s position as one of
the top three reporting groups of transients to TNS, with ≲1% Cropped FFT
contamination among spectroscopically classified transients first Pad difference image
reported by Förster et al. (2021).
In this work, we explore the effect of incorporating FFTs in
the Stamp Classifier, building upon previous works. Our modi-
fications start from the model architecture introduced in RJ23,
where the Satellite class was first introduced. The classifier cat-
Alert stamps

egorizes alerts into six distinct classes: supernovae (SNe), vari- Normalize,
mask & crop
able stars, asteroids, active galactic nuclei (AGN), satellites, and
bogus detections.
Cropped stamp cube
2.1. Preprocessing
After the ingestion of each alert, a preprocessing step takes place Fig. 2: Schematic of the preprocessing for our FFT-enhanced
before the model run. A schematic is shown in Fig. 2. The de- Stamp Classifier. The red dashed box encloses the new steps for
tailed steps to go from the base ZTF stamps to the cropped stamp the inclusion of the FFT with respect to previous versions of the
cube (blue box in the schematic), that serves as input for the classifier (e.g., RJ23).
model, are the following:

1. In the rare cases where the stamps do not have the original There are two reasons to apply the padding, and both are
dimensions of 63 × 63 pixels (e.g., the object is by the edge associated with the nature of the FFT algorithm. First, since
of the CCD), the alert is dropped. The current necessity for the transform maintains the sampling, padding allows increasing
the spatial frequency resolution. Second, the FFT algorithm is
3
For example, for a central 16 × 16 px portion, this results in 128 floats constructed under the assumption that the sequence to be trans-
(16 × 16 ÷ 2, due to the complex conjugate symmetry of the FFT). formed is infinitely periodic. Since this is not the case for these
Article number, page 3 of 8
 A&A proofs: manuscript no. main

images, artifacts that might appear due to this inconsistency are Table 1: Performance metrics for each model
mitigated by the padding. Similar measures are taken in other
fields that use the FFT, such as computational Fourier optics, to Label Model F1 score Precision Recall
deal with such issues (e.g., Voelz 2011). Although the FFT is a (%) (%) (%)
complex field, we focus on the amplitude due to the potential for 86.9 84.9 90.3
A Full
data compression. Alternatives to this are discussed in Sect. 4.2. (0.6) (0.8) (0.3)
87.3 86.7 88.1
B Multiscale
(0.3) (0.4) (0.5)
2.2. Models
85.8 85.3 86.6
C Cropped-16
Fig. 3 illustrates the architecture of the models evaluated in this (0.5) (0.9) (0.7)
study. The key components include the cropped stamp cube, the 87.1 86.2 88.1
D Full + FFT
FFT cutout from the preprocessing step, and additional metadata (0.5) (0.8) (0.5)
(primarily the coordinates of the alert). These inputs are pro- 87.6 86.6 89.0
E Multiscale + FFT
cessed through distinct neural network blocks before being com- (0.2) (0.4) (0.2)
bined in a final fully-connected layer that produces the model 87.6 86.5 89.0
F Cropped-16 + FFT
predictions. (0.4) (0.7) (0.6)
In the architecture, the cropped stamp cube is passed through
a convolutional block (referred to as the Stamp convolutional Notes. Mean values and standard deviations (in parentheses) are based
block) that applies rotational and flip augmentations. After the on five training runs for each model.
convolutional layers, the images are flattened, averaged over the
augmentation steps, and passed through a dropout layer before larger FoVs in these models. The inclusion of the FFT block in
being concatenated with the other features. While this block mir- these cases added complexity to the models, making them slower
rors the structural design from RJ23, for the models that feature in both training and running.
the FFT as an input (see the red short-dashed box in Fig. 3), the
hyperparameters were re-optimized separately.
Parallel to this, the FFT convolutional block processes the 2.3. Training
FFT cutout. Similar to the Stamp convolutional block, it includes
rotation and flip augmentation followed by convolutional and To ensure a fair comparison between models, we used the same
pooling layers, and finally, the output is flattened and concate- dataset and adhered closely to the training procedures outlined in
nated with the other features. RJ23, including maintaining the same train-test-validation split.
To assess the value of incorporating the FFT block, we com- The training process involved a hyperparameter (HP) search us-
pared models that included it with those that did not: ing Ray Tune (Liaw et al. 2018), optimizing parameters such as
the number of filters in the convolutional layers, learning rate,
A. Full: The full 63′′ × 63′′ (63 × 63 px) FoV alert at 1′′ reso- dropout rate, the size of the first convolutional kernel, and the
lution is used. The input consists of three images (science, size of the last dense layer.
reference, and difference) and their corresponding integrity For models A, B, and C (without the FFT block), we used
masks, creating a six-channel image. the hyperparameters from RJ23. For models D, E, and F (with
B. Multiscale: The stamp cube includes 8′′ , 16′′ , 32′′ , and 63′′ the FFT block), we performed additional optimization focusing
FoV stamps at corresponding resolutions of 1′′ , 2′′ , 4′′ , and on the FFT-specific convolutional layers. The search space for
8′′ . This produces a 24-channel image with four channels inherited parameters was narrowed to one-third of the original
per image type. range, centered around the optimal values from models A, B,
C. Cropped-16: The alert is cropped to a 16′′ FoV with the and C, while the FFT-specific parameters were searched within
full 1′′ resolution. a broad range of values.
D. Full + FFT: Combines the full 63′′ FoV with the FFT This approach ensured efficient HP optimization for the FFT
block. models, with search times comparable to the non-FFT models.
E. Multiscale + FFT: Combines the multiscale stamp cube After identifying the optimal hyperparameters, each model was
with the FFT block. trained five times using the entire training set to compute the F1
score, precision, and recall (see Section 5.7.2. in Murphy 2012,
F. Cropped-16 + FFT: Combines the cropped 16′′ FoV with
for definitions), as well as to estimate their respective standard
the FFT block.
deviations and assess the significance of changes. The results are
Model F (Cropped-16 + FFT) is particularly interesting due summarized in Table 1 for the test dataset.
to its scalability. The amount of information in an image scales
as the square of the number of pixels (N 2 ). However, the FFT al- 3. Results
gorithm’s computational cost scales as N ∗ log(N), making it sig-
nificantly more efficient, especially considering the subsequent Table 1 presents the performance metrics for each model. As in
cropping of the FFT to retain only the most relevant informa- RJ23, the Multiscale model outperforms the other non-FFT mod-
tion. This efficiency allows model F to capture better large-scale els. The inclusion of the FFT block in the Full model (A) did
features of transient phenomena, such as satellite glints, while not affect the F1 score, indicating that the FFT, while increas-
minimizing the inclusion of non-informative data (e.g., sky pix- ing the complexity of the model, did not compensate with ad-
els in a science image) that would slow the pipelines. ditional useful information. In the case of the Multiscale + FFT
For completeness, models D and E were also evaluated, al- model (E), there were marginal improvements, primarily in dis-
though they did not show statistically significant improvements tinguishing satellite glints from other types of bogus events that
compared to models A and B. This is because the FFT block the large-scale, low-resolution channels of the Multiscale model
provided redundant information that was already captured by the could not resolve. In contrast, the Cropped-16 + FFT model (F)
Article number, page 4 of 8
 J. P. Carvajal et al.: Tuning into spatial frequency space

Input Layer
Concatenate
Features

Metadata
x8 0°, 90°, 180°, 270° & flips ... + Model

...
output
Flatten
Convolutional Fully-connected
Layers Block
Cropped
Stamp cube Rotation & Flip Augmentation
+
Augmentation
x8 0°, 90°, 180°, 270° & flips ... averaging

...
FFT Flatten
Convolutional
Layers for FFT

Addition

Fig. 3: A schematic view of the architecture of the ML models used in this work. The input includes the cropped stamp cube, the
FFT cutout, and additional metadata. Both the stamp cube and FFT cutout undergo rotation and flip augmentation before being
processed through separate convolutional blocks (red long-dashed boxes). The outputs are then concatenated with the metadata and
passed through a fully-connected block for prediction. The red short-dashed box highlights the addition of the FFT block to the
architecture, which was included in models D, E, and F (see Sect. 2.2).

showed significant improvements over Cropped-16 (C) across all mental details, such as the position relative to nearby galaxies.
metrics, indicating that inclusion of the FFT effectively enhances The scale of this context information can vary; typical galaxy
classification in this scenario. sizes range from a few arcseconds to several tens of arcmin-
Fig.4 shows the confusion matrices for the Cropped-16 (C) utes for extreme cases like the local Pinwheel Galaxy (M101).
and Cropped-16 + FFT (F) models. The addition of the FFT An important example is SN 2023ixf, discovered by amateur
notably reduced misclassifications of satellites, bringing perfor- astronomer Koichi Itagaki (Itagaki 2023), within M101. Such
mance in line with models A and B as noted in Fig.5. The de- a large host galaxy can complicate early classification efforts,
tailed analysis of these results follows in the next section. as methods relying on context information, such as ALeRCE’s
Stamp Classifier (Carrasco-Davis et al. 2021), might miss SNe
in such contexts, as happened in this case, where it was misla-
4. Discussion beled as bogus. This underscores the importance of a multiscale
approach for capturing crucial contextual data. Satellite informa-
We evaluate here the impact of incorporating the FFT in satel- tion, however, is distinct due to its large linear spread and tran-
lite detection and explore potential applications for internal sur- sient nature, making the FFT particularly effective in isolating
vey pipelines responsible for generating transient alerts. By en- it.
hancing the early detection and classification process, FFT-based
methods may help streamline the flow of high-quality alerts in Our experiment showed that the Cropped-16 model (C) per-
large-scale surveys. We first analyze the experimental results be- formed worse overall in the classification task compared to the
fore discussing broader implications. Full (A) or the Multiscale (B) models (see Table 1). Since it con-
tains less context information (smaller FoV), this is to be ex-
pected. It was particularly bad at identifying satellites, frequently
4.1. The impact of the spatial frequency space confusing them with asteroids, SNe, and bogus detections (see
below for more details). By including the FFT Block in the
In Fourier space, the distinct signatures of satellites, character-
Cropped-16 + FFT model (F), these issues were largely resolved,
ized by transient and extended features, become clear. Satellite
with satellite identification improving enough to match the best
glints, challenging to detect in the spatial domain, are efficiently
performance metrics of the analyzed models. The confusion be-
isolated in the FFT space, enhancing detection accuracy and pro-
tween satellites (true label) and asteroids, other bogus, and SNe
viding a compact data representation for ML models.
was greatly reduced down to the level of the best-performing
To evaluate the benefits of FFT, we categorize image infor-
models. One issue that persisted was the confusion between SNe
mation into three types (source, context, and satellite), which
and asteroids, which was not improved by the FFT and did not
correspond to different spatial scales and play a critical role
match the performance of other models (see Fig. 5).
in classification tasks. Source information captures small-scale,
transient, or variable objects, such as the point-like SNe, AGN, These results suggest that the FFT of the difference image
or variable stars, or high proper motion asteroids such as near is highly effective for identifying satellites and capturing the pe-
earth objects, which may look extended and trail across small an- riodic nature of glints, which are challenging to detect in the
gular distances, typically spanning only a few arcseconds in size. spatial domain. However, compared to the Multiscale model, the
Context information provides larger-scale, unchanging environ- FFT fails to provide the necessary context information for clas-
Article number, page 5 of 8
 A&A proofs: manuscript no. main

Cropped-16 model (C) Confusion Matrix Full model (A) Confusion Matrix
90.1 0.0 0.6 0.0 2.0 7.3 89.8 0.0 0.8 0.0 2.9 6.4
SNe Satellite Bogus Asteroid AGN

SNe Satellite Bogus Asteroid AGN

(1.2) (0.0) (0.3) (0.0) (1.6) (1.0) (2.1) (0.0) (0.1) (0.0) (0.8) (1.7)
80 80
0.0 97.9 0.7 0.2 1.1 0.1 0.0 97.4 1.0 0.2 1.3 0.1
(0.0) (0.6) (0.1) (0.2) (0.5) (0.0) (0.0) (0.5) (0.4) (0.1) (0.4) (0.0)

0.7 1.8 93.4 1.5 1.8 0.9 60 0.6 1.1 93.4 2.4 1.7 0.8 60
True Label

True Label
(0.2) (0.2) (0.7) (0.3) (0.4) (0.2) (0.1) (0.5) (0.7) (0.2) (0.3) (0.2)

0.2 10.0 14.3 66.9 7.4 1.2 0.0 1.4 9.8 87.4 0.7 0.7
(0.3) (4.1) (4.6) (5.1) (2.8) (0.4) 40 (0.0) (1.0) (1.9) (1.8) (1.0) (0.3) 40

7.2 10.7 2.5 0.5 78.9 0.1 6.8 8.2 1.9 0.1 82.7 0.3
(2.2) (1.2) (0.8) (0.2) (3.4) (0.1) (0.5) (0.8) (0.5) (0.1) (0.9) (0.3) 20
20
6.5 0.2 0.6 0.0 0.3 92.3 7.4 0.2 0.8 0.0 0.3 91.2
(1.4) (0.0) (0.1) (0.0) (0.1) (1.6) (1.6) (0.1) (0.1) (0.0) (0.2) (1.9)
VS

VS
0 0
AGN Asteroid Bogus Satellite SNe VS AGN Asteroid Bogus Satellite SNe VS
Predicted Label Predicted Label
(a) (a)

Cropped-16 + FFT model (F) Confusion Matrix Multiscale model (B) Confusion Matrix
89.6 0.0 0.6 0.0 2.4 7.3 90.0 0.0 0.7 0.0 2.0 7.3
SNe Satellite Bogus Asteroid AGN

SNe Satellite Bogus Asteroid AGN

(0.8) (0.0) (0.1) (0.0) (0.5) (0.6) (1.1) (0.0) (0.2) (0.0) (0.6) (1.1)
80 80
0.0 97.7 0.7 0.2 1.3 0.1 0.0 97.8 0.8 0.2 1.1 0.1
(0.0) (0.2) (0.1) (0.2) (0.2) (0.1) (0.0) (0.7) (0.4) (0.1) (0.4) (0.0)

0.6 1.3 94.0 1.4 1.7 0.9 60 0.6 1.2 94.7 1.4 1.3 0.8 60
True Label

True Label

(0.1) (0.1) (0.3) (0.2) (0.2) (0.2) (0.1) (0.1) (0.4) (0.1) (0.2) (0.1)

0.0 2.1 14.5 79.7 2.1 1.7 0.0 2.6 21.4 71.9 2.8 1.4
(0.0) (0.4) (2.5) (2.6) (0.7) (0.0) 40 (0.0) (1.8) (1.8) (2.9) (1.0) (0.7) 40

6.4 10.4 2.4 0.0 80.6 0.3 6.9 8.8 2.5 0.1 81.5 0.3
(0.8) (0.8) (0.6) (0.0) (1.3) (0.1) 20 (1.0) (1.0) (0.7) (0.1) (1.5) (0.3)
20
6.7 0.2 0.5 0.0 0.4 92.2 6.3 0.1 0.6 0.0 0.2 92.8
(0.5) (0.0) (0.3) (0.0) (0.1) (0.7) (0.9) (0.0) (0.2) (0.0) (0.1) (0.9)
VS

VS

0 0
AGN Asteroid Bogus Satellite SNe VS AGN Asteroid Bogus Satellite SNe VS
Predicted Label Predicted Label
(b) (b)

Fig. 4: Confusion matrices for the Cropped and Cropped + FFT Fig. 5: Same as Fig. 4 for the Full and Multiscale models.
models (C & F). The values are normalized by the True labels
(rows). In parentheses are the standard deviations over 5 runs.
F to model A (Fig. 5a), we find that Model F improves Bogus-
Satellite confusion by 1.0% ± 0.3%, although Model A still out-
sifying some types of objects effectively, most noticeably in the performs Model F in Satellite-Bogus confusion by 4.7% ± 3.1%.
confusion of SNe and asteroids. Given the substantial class imbalance —with 27.5 times more
It is important to note that the dataset is highly imbalanced, bogus objects than satellites— these percentages should be in-
with significantly more bogus objects than satellites. To mitigate terpreted cautiously. Overall, the 16′′ FoV Model F, enhanced
this, we employed balanced training batches and class-balanced by the inclusion of FFT, performs at least as well as, if not better
cross-entropy loss for validation (see RJ23 for details). Despite than, the other models in satellite identification.
these efforts, some class imbalance effects may remain. In this
context, while the Bogus-Satellite confusion rate is the same 4.2. Future Prospects
for both models B (Fig. 5b) and F (Fig. 4b), Model F reduces
Satellite-Bogus confusion by 6.9% ± 3.1% compared to Model In this study, we focused on the amplitude of the FFT, using it
B. This highlights that Model F is distinctly better at distin- as a single additional channel for data compression. However,
guishing satellites from other bogus objects. Comparing model since the FFT provides a complex field, the phase information,
Article number, page 6 of 8
 J. P. Carvajal et al.: Tuning into spatial frequency space

Science img Difference img FFT (8x8 binning)

0 0 0

10
100 100

20
200 200
Satellite

30

300 300
40

400 400 50

60
500 500
0 100 200 300 400 500 0 100 200 300 400 500 0 10 20 30 40 50 60

Fig. 6: 512 × 512 pixel cutouts from a ZTF full-chip image containing the alert-triggering object ZTF20aapfawc at the center.
The satellite trail, highlighted with a red box, spans across the cutouts. Additional alerts raised by this satellite streak correspond
to objects ZTF20aapfavz and ZTF20aapfavx, located 60′′ to the bottom-left and 250′′ to the top-right of the center, respectively,
along the highlighted trail. The left and center panels display science and difference images, respectively, while the right panel shows
the FFT of the difference image with an 8-pixel binning. In spatial frequency space, the satellite exhibits a distinctive fringe-like
pattern, similar to the smaller stamps shown in Fig. 1.

though not considered here, may also prove valuable in distin- proaches for detecting trails, mentioned in Sect. 1, involve ana-
guishing certain features. Future work could investigate alterna- lyzing images to search for elongated trail-like shapes (e.g., us-
tive approaches, such as incorporating 1-channel real-part FFT ing the CREATETRACKIMAGE task from Laher et al. 2014) or em-
or 2-channel real- and imaginary-part FFT components, to cap- ploying CNNs (Duev et al. 2019b), similar to this work. Uncon-
ture a wider range of spatial frequency features. nected blobs from satellite glints pose a challenge for the former
Beyond the technical aspects of the FFT, its practical appli- method. Regarding the latter, a promising area for future devel-
cations in large-scale surveys like LSST are worth exploring. opment might be to incorporate FFTs into it, given the relative
Integrating FFT into internal pipelines could significantly im- ease to include FFT in ALeRCE’s Stamp Classifier.
prove the identification of satellite contaminants before the data However, it is important to acknowledge that no single
reaches downstream (alert) processing. However, it is important method is likely to completely address the problem of satellite
to note that for the FFT to be most effective, the satellite trail contamination. While we have highlighted the strengths of the
should be well-centered in the image. In our tests, an offset of 5 FFT approach, reducing the misclassification of satellite events
pixels already significantly distorted the signal, and larger offsets will likely require a combination of techniques, each address-
quickly led to the signal disappearance. Fig. 6 demonstrates the ing different aspects of the contamination problem. Combining
application of FFT over a larger cutout (512 px), beyond the typ- methods with complementary strengths, such as tracklet recon-
ical 63 px of ZTF alerts. The central portion of the image focuses struction, trail detection, CNN-based models, and FFT-enhanced
on a satellite trail, initially mislabeled as bogus but correctly classification, should offer more robust mitigation of satellite
identified as a satellite by our FFT-enhanced model, later con- contamination in large-scale surveys like LSST.
firmed by human inspection using the ALeRCE Explorer4 . Bin-
Another promising avenue involves trail-fitting techniques
ning the FFT (right panel of Fig. 6) smooths out high-frequency
like those used by Vereš et al. (2012) for asteroids, which could
details, allowing us to focus on large-scale variations that are
potentially be adapted to include light-curve estimation. This
more characteristic of satellite signatures. This approach demon-
could allow for the subtraction of satellite signatures from im-
strates how FFTs could be implemented in internal pipelines,
ages. Such trail-modeling could be crucial for distinguishing be-
leveraging the information from very extended trails that pro-
tween human-made and natural fast-moving objects, as well as
duce apparently independent blobs at smaller scales, which may
broader space pollution characterization.
otherwise even fall below the detection threshold. Identifying
and intelligently masking out these extended tracks could greatly As of this writing, the public plans for the Rubin Observa-
reduce the number of complex satellite stamps which currently tory’s data management facilities do not explicitly mention the
pass into the alert stream. use of graphics processing units (GPUs). However, given their
When comparing our method to other approaches for satel- widespread adoption in accelerating ML pipelines, GPUs would
lite detection, several distinctions arise. For instance, Karpov likely be a valuable addition to the survey’s infrastructure, either
& Peloton (2022, 2023) used tracklet reconstruction to identify from the outset or as a future upgrade. GPU-optimized FFT algo-
satellites by correlating non-repeating transient sources geomet- rithms, known for their scalability and efficiency (e.g., see More-
rically. While this is effective, and helped to reveal the impact land & Angel 2003; NVIDIA Corporation 2023; Advanced Mi-
satellites have on surveys, our method offers the advantage of cro Devices, Inc. 2023), offer an effective solution for process-
considering both the bright and low-brightness features of satel- ing the massive volumes of high-resolution data that LSST will
lite trails, which could become increasingly important as the generate. Their capacity for rapid FFT processing is essential to
LSST’s deeper exposures reveal fainter satellite glints. Other ap- ensure that FFT-based feature extraction keeps up with the data
flow in large-scale surveys like LSST. Incorporating GPUs into
4
https://alerce.online/ LSST’s on-site computation facilities would significantly boost
Article number, page 7 of 8
 A&A proofs: manuscript no. main

processing efficiency and scalability, though adapting existing data processing pipelines could offer a substantial improvement
pipelines to new hardware and software may pose challenges. in the purity and efficiency of alert systems, providing a more
robust framework for upcoming large-scale surveys.
In conclusion, the FFT offers a powerful tool for both en-
5. Conclusions hancing contaminant detection and improving data compression
The study of transient astrophysical events in modern surveys in astronomical surveys. By incorporating these techniques into
such as ZTF and the LSST comes with several challenges, in- future data processing pipelines, we can better prepare for the
cluding managing the vast volumes of data and mitigating the challenges of the LSST era and beyond, ultimately enabling
presence of contaminants like CCD artifacts, template subtrac- more efficient and accurate astronomical research.
tion artifacts, and satellite trails. In the context of the burgeon- Acknowledgements. We acknowledge support from the National Agency for
ing space industry, the latter is a rapidly growing problem that Research and Development (ANID) grants: Millennium Science Initiative
has yet to be adequately addressed by time-domain surveys (see, ICN12_009 (FEB, AMMA, IRJ, MC) and AIM23-0001 (FEB, FF, MC), BASAL
Center of Mathematical Modeling Grant FB210005 (FF, AMMA), BASAL
e.g., Catelan 2023, and references therein for a recent review). projects ACE210002 (AB, MC) and FB210003 (JPC, FEB, AB, MC), FONDE-
The current LSST bandwidth constraints on alert data volume CYT Regular 1241005 (FEB), FONDECYT Regular 1231637 (MC), Beca de
may strongly limit the ability of downstream brokers and users Doctorado Nacional (JPC). We also acknowledge the use of the Kultrún comput-
to optimally filter out various contaminants. Thus, investigating ing cluster at Universidad de Concepción, funded by Conicyt Quimal #170001,
efficient methods to improve the identification and filtering of Anillo ACT172033, Fondecyt regular 1180291, Fondecyt Iniciacion 11170268,
Basal AFB-170002, and Núcleo Milenio Titans NCN19-058.
contaminants early in the data stream, or alternatively to com-
press the data information more effectively into alerts, will hope-
fully ensure that the most scientifically valuable information is
retained and leveraged by end users. References
In this study, we explored the utility of the FFT in enhanc- Advanced Micro Devices, Inc. 2023, AMD rocFFT Library, version 1.0.29
ing the detection of satellite glints within astronomical survey Bellm, E. C., Kulkarni, S. R., Graham, M. J., et al. 2018, Publications of the
data streams and its potential as an image compression tool. We Astronomical Society of the Pacific, 131, 018002
Carrasco-Davis, R., Reyes, E., Valenzuela, C., et al. 2021, The Astronomical
specifically utilized ZTF data, while also anticipating the chal- Journal, 162, 231
lenges of the upcoming LSST era. Our results demonstrate that Catelan, M. 2023, in Memorie della Societa Astronomica Italiana, Vol. 94, 56
the FFT effectively isolates the distinct signatures of satellite Chen, X., Wang, S., Deng, L., et al. 2020, ApJS, 249, 18
Cooley, J. W. & Tukey, J. W. 1965, Math. Comput., 19, 297
trails and glints, concentrating these features in a compact region Duev, D. A., Mahabal, A., Masci, F. J., et al. 2019a, Monthly Notices of the
of Fourier space. This compactness allows for efficient data com- Royal Astronomical Society, 489, 3582
pression, significantly reducing the amount of non-informative Duev, D. A., Mahabal, A., Ye, Q., et al. 2019b, Monthly Notices of the Royal
data. Astronomical Society, 486, 4158
Förster, F., Cabrera-Vives, G., Castillo-Navarrete, E., et al. 2021, The Astronom-
Our experiment compared the performance of ML models ical Journal, 161, 242
using FoVs of different sizes for the three ZTF alert image Graham, M. J., Kulkarni, S. R., Bellm, E. C., et al. 2019, Publications of the
stamps (science, reference, difference) and assessed the impact Astronomical Society of the Pacific, 131, 078001
Itagaki, K. 2023, Transient Name Server Discovery Report, 2023-1158, 1
of adding a central cutout of the FFT of the difference image Ivezić, Z., Kahn, S. M., Tyson, J. A., et al. 2019, The Astrophysical Journal, 873,
as an input. As discussed in Sect. 4.1, incorporating the FFT 111
block significantly improved the classification metrics of model Karpov, S. & Peloton, J. 2022, arXiv e-prints, arXiv:2202.05719
C (which adopted a smaller 16′′ FoV) to the levels achievable Karpov, S. & Peloton, J. 2023, Contributions of the Astronomical Observatory
Skalnate Pleso, 53, 69
with the full 63′′ FoV. In model F, using 16′′ FoV stamps that in- Laher, R. R., Surace, J., Grillmair, C. J., et al. 2014, Publications of the Astro-
cluded an FFT image, the overall F1 score increased to the level nomical Society of the Pacific, 126, 674
of the best performing models, with the most notable improve- Liaw, R., Liang, E., Nishihara, R., et al. 2018, arXiv e-prints, arXiv:1807.05118
LSST Science Collaboration, Abell, P. A., Allison, J., et al. 2009, arXiv e-prints,
ment on the classification of satellites. The confusion between arXiv:0912.0201
SNe and asteroids, which can be the product of the small FoV, Masci, F. J., Laher, R. R., Rusholme, B., et al. 2019, PASP, 131, 018003
was not alleviated by the inclusion of the FFT. In this case, we Matheson, T., Stubens, C., Wolf, N., et al. 2021, The Astronomical Journal, 161,
found that the additional contextual information of larger FoVs 107
Moreland, K. & Angel, E. 2003, in Graphics Hardware, ed. M. Doggett, W. Hei-
was important, and the multiscale model introduced in RJ23 drich, W. Mark, & A. Schilling (The Eurographics Association)
(model B here) did a better job while keeping the size of the Murphy, K. P. 2012, Machine Learning: A Probabilistic Perspective, Adaptive
input low. Computation and Machine Learning Series (Cambridge, Massachusetts, Lon-
don, England: MIT Press)
Beyond the immediate scope of this work, the FFT’s poten- Möller, A., Peloton, J., Ishida, E. E. O., et al. 2020, Monthly Notices of the Royal
tial applications in internal data processing pipelines are vast. Astronomical Society, 501, 3272
As discussed in Sect. 4.2, integrating FFTs into existing models Nordin, J., Brinnel, V., van Santen, J., et al. 2019, A&A, 631, A147
can enable the analysis of larger image cutouts without a sub- NVIDIA Corporation. 2023, NVIDIA cuFFT Library, version 12.6
Patterson, M. T., Bellm, E. C., Rusholme, B., et al. 2018, Publications of the
stantial increase in computational cost, facilitating more accurate Astronomical Society of the Pacific, 131, 018001
and scalable contaminant detection. This approach could be par- Perley, D. A., Fremling, C., Sollerman, J., et al. 2020, The Astrophysical Journal,
ticularly impactful for LSST, where efficient data handling and 904, 35
Reyes-Jainaga, I., Förster, F., Arancibia, A. M. M., et al. 2023, The Astrophysical
precise contamination filtering will be paramount. Journal Letters, 952, L43
Moving forward, this method opens several avenues for fu- Sánchez-Sáez, P., Arredondo, J., Bayo, A., et al. 2023, A&A, 675, A195
ture optimization. For instance, since the signature of satellites Sánchez-Sáez, P., Lira, H., Martí, L., et al. 2021, AJ, 162, 206
Smith, K. W., Williams, R. D., Young, D. R., et al. 2019, Research Notes of the
in Fourier space is fringe-like, we found success in experiments AAS, 3, 26
to reduce the images into 1D arrays (“spectra”). The possibility Tyson, J. A., Snyder, A., Polin, D., Rawls, M. L., & Ivezić, Ž. 2024, ApJ, 966,
of deriving features related to this 1D frequency space spectra L38
or even coefficients that could capture the 2D nature may be of Vereš, P., Jedicke, R., Denneau, L., et al. 2012, PASP, 124, 1197
Voelz, D. G. 2011, Computational Fourier Optics, Tutorial Texts (Bellingham,
relevance as information to include in alert packets in these data WA: SPIE Press)
streams. Additionally, integrating this approach with real-time
Article number, page 8 of 8