TIRVolcH: Thermal Infrared Recognition of Volcanic Hotspots. A single band TIR-based algorithm to detect low-to-high thermal anomalies in volcanic regions

Abstract

Detecting early signs of impending eruptions and monitoring the evolution of volcanic phenomena are fundamental objectives of applied volcanology, both essential for timely assessment of associated hazards. Thermal remote sensing proves to be a cost-effective, yet reliable, information source for these purposes, especially for the hundreds of volcanoes still lacking conventional ground-based monitoring networks. In this work, we present an innovative and effective single band TIR-based (11.45 μm) algorithm (TIRVolcH), capable of detecting thermal anomalies in a broad range of volcanic settings, from low-temperature hydrothermal systems to high-temperature effusive events. Based on the processing of Visible Infrared Imaging Radiometer Suite (VIIRS) scenes, the algorithm offers an unprecedented trade-off between spatial (375 m) and temporal resolution (multiple acquisitions per day), having the potential to detect thermal anomalies for pixel-integrated temperatures as low as 0.5 K above the background, while maintaining a false positive rate of ∼1.8 %. The analysis of decadal time series of VIIRS data (2012−2023), acquired at three different volcanoes, reveals how the algorithm can: (i) detect hydrothermal crises at fumarolic fields (Vulcano, Italy), (ii) unveil thermal unrest preceding dome extrusions and explosive eruptions (Agung, Indonesia), and (iii) spatially trace lava flows extent and quantify their advancement rate, as well as track their long-term cooling behaviour (La Palma, Spain).
We envisage that the algorithm will prove instrumental for detecting early signs of volcanic activity and following the evolution of eruptive phenomena, providing a useful tool for hazard management and risk reduction applications. Furthermore, the compilation of statistically robust multidecadal thermal datasets will provide novel insights and new perspectives into volcano monitoring, laying the ground for forthcoming higher-resolution TIR missions.

Full text

TIRVolcH: Thermal Infrared Recognition of Volcanic Hotspots. A single
band TIR-based algorithm to detect low-to-high thermal anomalies in
volcanic regions.
S. Aveni
a,b,*
, M. Laiolo
b,c
, A. Campus
b
, F. Massimetti
b
, D. Coppola
b,c
a
Department of Civil, Constructional and Environmental Engineering (DICEA), Sapienza University of Rome, Via Eudossiana 18, 00184 Rome, Italy
b
Department of Earth Sciences, University of Turin, Via Valperga Caluso 35, 10125 Turin, Italy
c
NATRISK: Centro Interdipartimentale sui Rischi Naturali in Ambiente Montano e Collinare, Universit`
a di Torino, Largo Paolo Braccini, 2, 10095 Grugliasco, Italy
ARTICLE INFO
Edited by: Jing M. Chen
Keywords:
Thermal InfraRed (TIR)
Volcano monitoring
Hotspot detection algorithm
Precursory thermal activity
Lava ows
VIIRS
ABSTRACT
Detecting early signs of impending eruptions and monitoring the evolution of volcanic phenomena are funda-
mental objectives of applied volcanology, both essential for timely assessment of associated hazards. Thermal
remote sensing proves to be a cost-effective, yet reliable, information source for these purposes, especially for the
hundreds of volcanoes still lacking conventional ground-based monitoring networks. In this work, we present an
innovative and effective single band TIR-based (11.45
μ
m) algorithm (TIRVolcH), capable of detecting thermal
anomalies in a broad range of volcanic settings, from low-temperature hydrothermal systems to high-
temperature effusive events. Based on the processing of Visible Infrared Imaging Radiometer Suite (VIIRS)
scenes, the algorithm offers an unprecedented trade-off between spatial (375 m) and temporal resolution
(multiple acquisitions per day), having the potential to detect thermal anomalies for pixel-integrated tempera-
tures as low as 0.5 K above the background, while maintaining a false positive rate of ~1.8 %. The analysis of
decadal time series of VIIRS data (2012−2023), acquired at three different volcanoes, reveals how the algorithm
can: (i) detect hydrothermal crises at fumarolic elds (Vulcano, Italy), (ii) unveil thermal unrest preceding dome
extrusions and explosive eruptions (Agung, Indonesia), and (iii) spatially trace lava ows extent and quantify
their advancement rate, as well as track their long-term cooling behaviour (La Palma, Spain).
We envisage that the algorithm will prove instrumental for detecting early signs of volcanic activity and
following the evolution of eruptive phenomena, providing a useful tool for hazard management and risk
reduction applications. Furthermore, the compilation of statistically robust multidecadal thermal datasets will
provide novel insights and new perspectives into volcano monitoring, laying the ground for forthcoming higher-
resolution TIR missions.
1. Introduction
Volcanic eruptions and associated or cascading hazards pose a threat
to at least 800 million people living within a 100 km radius of one of the
~1400, potentially active, subaerial Holocene volcanoes on Earth
(Small and Naumann, 2001,Brown et al., 2015a,Lara et al., 2021).
Between 1600 and 2010, more than 270,000 fatalities were associated
with volcanic activity (Auker et al., 2013). Furthermore, these events
have the potential to cause widespread damage to infrastructures,
disrupt global aviation, lead to socio-economic losses, and cause adverse
effects on human health (Hansell et al., 2006,Horwell and Baxter, 2006,
Gudmundsson, 2011,Brown et al., 2015b,Brown et al. 2015c,Loughlin
et al., 2015). Given the transient nature of volcanic phenomena, the
capability of detecting, with early notice, variations in the equilibrium
of volcanic systems often draws the line between survival and fatality
rate (Garcia and Fearnley, 2012,Auker et al., 2013,Poland and
Anderson, 2020,Lowenstern and Ramsey, 2017).Even after an eruption
begins, the ability to monitor its progression (i.e., identifying active
vents, establishing directions and velocities of advancing lava ows,
etc.), remains essential for stakeholders and competent bodies to timely
review expected scenarios, update hazard maps, and issue exclusion
and/or evacuation orders (Ganci et al., 2012a;Harris et al., 2016;Harris
et al., 2017;Harris et al., 2019;Coppola et al., 2016a;Coppola et al.,
2020;Ganci et al., 2020).
* Corresponding author.
E-mail address: simonesalvatore.aveni@uniroma1.it (S. Aveni).
Contents lists available at ScienceDirect
Remote Sensing of Environment
journal homepage: www.elsevier.com/locate/rse
https://doi.org/10.1016/j.rse.2024.114388
Received 4 May 2024; Received in revised form 2 August 2024; Accepted 24 August 2024
Remote Sensing of Environment 315 (2024) 114388
Available online 3 October 2024
0034-4257/© 2024 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license (
http://creativecommons.org/licenses/by/4.0/ ).

Forefront instruments for detecting early signs of volcanic unrest, as
well as advancements in tracking eruptive episodes, play a major role in
reducing the risk associated with volcanic events, lowering the overall
vulnerability of those exposed to volcanic hazards (Sparks et al., 2012;
Donovan et al., 2012). Studies suggested that advancements in volca-
nology likely saved as many as 50,000 lives in the twentieth century
alone (Auker et al., 2013). On the other hand, recent works also revealed
that ~50 % of the ~1400 potentially active volcanoes still lack con-
ventional ground-based instruments capable of detecting impending
volcanic resurgence (Brown et al., 2015a,Pritchard et al., 2018,Delgado
et al., 2019).
In this context, satellite data proves to be a cost-effective, yet reli-
able, information source for detecting early signs of volcanic activity and
monitoring the evolution of eruptive events at remote, poorly moni-
tored, volcanoes. Yet, even at well-monitored volcanoes, remotely
sensed data provides a spatiotemporal perspective of eruptive phe-
nomena, mitigating for the gaps in information left by ground-based
monitoring networks (Ebmeier et al., 2018;Reath et al., 2019b;Cop-
pola et al., 2020).
Spaceborne data have been employed with a range of techniques to
monitor various parameters, encompassing volcanic gas/ash emissions
(e.g., Carn et al., 2017;Prata, 2009), deformation (i.e., Pritchard and
Simons, 2004a, 2004b;Biggs et al., 2014;Pritchard et al., 2018), and
thermal output (i.e., Wright et al., 2015;Wright, 2016;Coppola et al.,
2023). Amongst them, the investigation of volcanic phenomena through
thermal infrared observations has been a central focus in remote sensing
studies for decades, leading to the establishment of a dedicated disci-
pline since the 1960s (Ramsey and Harris, 2013,Harris, 2013,Blackett,
2017).
The progress made in recent decades has allowed the development of
the rst automatic volcano hotspot-detection systems such as MOD-
VOLC (Wright et al., 2002;Wright et al., 2004) and MIROVA (Coppola
et al., 2016b; see Steffke and Harris (2011) for a comprehensive review).
These systems, based on the joint availability of Thermal InfraRed (TIR)
and Mid-InfraRed (MIR) data acquired by the MODerate resolution
Imaging Spectroradiometer (MODIS) mounted aboard Terra (since
1999) and Aqua (since 2002) satellites are still used to detect eruptive
activity all over the world, exploiting the high sensitivity of MIR chan-
nels to high (i.e., magmatic) temperatures. With a resolution of 1 km and
a revisit time of approximately 12 h, MODIS sensors constitute an
important volcanic monitoring tool to support volcanological observa-
tories (Coppola et al., 2020).
However, due to the moderate spatial resolution and given that they
exploit the different spectral response of the MIR and TIR bands to hot
surfaces ⪆600 K (i.e., Normalised Thermal Index (NTI); see Wright
et al., 2002 for details), these algorithms are not designed to detect low-
temperature (⪅600 K) volcanic phenomena (Zhukov and Oertel, 2001,
Briess et al., 2003). Hydrothermal systems, crater lakes, and fumarolic
elds, for instance, as well as cooling lava bodies and cooler dome
carapaces, are often characterised by temperatures well below the MIR-
method-operational threshold (i.e., ⪆600 K) and, as such, the thermal
energy sourced by these volcanic features remains undetected and/or
unquantied. To track and quantify thermal radiations sourced from
these low-temperature volcanic features, scholars employed TIR chan-
nels, these better suited to track subtle thermal variations (Reath et al.,
2016;Ramsey et al., 2023). Yet, even in a high-temperature domain, TIR
observations typically prove more effective in accurately determining
the geometrical characteristics of emplaced and advancing lava ows. In
fact, the spectral response of MIR channels to hot surfaces is far greater
than that of TIR bands, meaning that, even a metrical subpixel hot
component lets the overall pixel-integrated temperature rise exponen-
tially (Steffke and Harris, 2011), following a power-law relationship
(Wooster et al., 2003). The substantial energy radiated from a relatively
small subpixel component causes the thermal signal to spread across
several adjacent pixels, following the convolution scheme dictated by
the sensor’s Point Spread Function (PSF) (Markham, 1985;
Schowengerdt, 2007;Calle et al., 2009;Zakˇ
sek et al., 2015a;Zakˇ
sek
et al., 2015). Blurring artefacts in the hotspot-contaminated MIR scenes
imply that even a small lava ow, factually contained within a few
pixels, may spread over a considerably larger area, thus signicantly
affecting the estimation and accuracy of length, width, and shape of the
volcanic products (Harris, 1996;Harris, 2013;Harris et al., 2017;
Ramsey et al., 2019). On the other hand, TIR acquisitions do retain
sensitivity to hot surfaces (Aveni and Blackett, 2022;Verdurme et al.,
2022), reducing PSF-related distortions and, in turn, allowing a more
detailed interpretation of the eruptive scenario.
In this regard, together with MODIS, the Advanced Spaceborne
Thermal Emission and Reection Radiometer (ASTER), aboard the Terra
satellite marked a transformative phase in the thermal remote sensing of
volcanic activity. With ve TIR (8–12
μ
m) bands, and a spatial resolu-
tion of 90 m ASTER has been instrumental in detecting early (or pre-
cursory) signs of volcanic activity (Pieri and Abrams, 2005;Reath et al.,
2016), track the evolution of volcanic unrest (Corradino et al., 2023;
Pailot-Bonn´
etat et al., 2023), assess the progression of effusive episodes
(Harris et al., 2019;Genzano et al., 2021;Ramsey and Flynn, 2020;
Ramsey et al., 2023), quantify the thermal energy sourced by volcanic
and hydrothermal targets (Mia et al., 2018;Mannini et al., 2019;
Ramsey et al., 2023), locate thermal anomalies in volcanic regions,
fumarolic elds, geothermal areas, and hydrothermal systems (Genzano
et al., 2021,Taryn et al., 2018,Uchˆ
oa et al., 2023,Hellman and Ramsey,
2004,Vaughan et al., 2012a,Vaughan et al., 2020,Chalik et al., 2019,
Braddock et al., 2017,Silvestri et al., 2019,Hilman et al., 2020,Caputo
et al., 2019, and references therein), and creating comprehensive mul-
tidecadal database of volcanic thermal behaviour (Reath et al., 2019a,
Urai and Pieri, 2011a, 2011b [https://gbank.gsj.jp/vsidb/image/Agun
g/volinfo.html]). Previous authors employed TIR bands to monitor
low-energy fumarolic elds (Vaughan et al., 2012b;Braddock et al.,
2017;Caputo et al., 2019;Reath et al., 2019a;Silvestri et al., 2019;
Ramsey and Flynn, 2020;Way et al., 2022), whilst others were suc-
cessful in quantifying the heat ux sourced by hydrothermal systems
(Harris and Stevenson, 1997a, 1997b;Mannini et al., 2019). Further-
more, TIR-based retrospective studies conducted on selected targets
revealed early evidence of variations in the thermal activity associated
with impending eruptions (e.g., Dehn et al., 2002;Pieri and Abrams,
2005;Reath et al., 2016). These studies revealed how TIR radiation
analyses are effective in measuring the heat sourced from volcanic tar-
gets. However, despite ASTER’s high spatial resolution, its low temporal
resolution (1 image every 16 days, at nadir over the equator) poses a
major limitation for volcano monitoring. This is further exacerbated by
the irregular acquisitions of ASTER scenes over several volcanic targets
(Reath et al., 2019b;Ramsey and Flynn, 2020) which makes it quasi-
impossible to timely detect potential manifestation of subtle thermal
anomalies. Even during eruptive crises, despite the off-nadir pointing
capabilities of the instrument and the Urgent Request Protocol (URP)
program (see Ramsey, 2016 for details) can be invoked to increase the
acquisition frequency up to 1 image every 4 days (depending on target
latitude), the revisit time of ASTER impedes a timely assessment of the
progression and variations in the eruptive dynamics, especially at vol-
canoes exhibiting short-living episodes (i.e., Waythomas et al., 2017;
Letourneur, 2008;Coppola et al., 2005;Coppola et al., 2021;Bonaccorso
and Aloisi, 2021;Marquez et al., 2022;Proietti et al., 2023;Calvari and
Nunnari, 2022;Guerrieri et al., 2023;Ganci et al., 2023;Bignami et al.,
2020;Werner et al., 2017) or those characterised by persistently
unfavourable meteorological conditions where the likelihood of
acquiring a cloud-free scene is lowered by more than 65 % (i.e., Mannini
et al., 2019,Coppola et al., 2022,Blackett and Wooster, 2011,Reath
et al., 2019a, 2019b,Henney, 2012,Carter and Ramsey, 2010,Gray
et al., 2019).
The limitations described above have meant that, at present, there is
no automatic satellite-based system for detecting low-to-high tempera-
ture anomalies using a single approach. To address the numerical
challenges related to the automatic or supervised recognition and
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
2

detection of anomalous pixels using a single TIR band, previous studies
applied in-scene contextual and/or statistical thresholding (Murphy
et al., 2011;Blackett, 2014;Rabuf et al., 2022;Pailot-Bonn´
etat et al.,
2023), long-term change-point detection (Tramutoli, 1998;Tramutoli,
2005;Genzano et al., 2021), combination of image processing and sta-
tistical techniques (Ramsey et al., 2023), Machine Learning (ML) ap-
proaches (Corradino et al., 2023), and supervised inspection and
selection routines (Reath et al., 2019a).
Yet, these studies and/or algorithms were exclusively based on the
processing and elaboration of ASTER-like scenes, thus constrained by
the temporal availability of these acquisitions. Furthermore, the desig-
nated lifetime of TERRA (6 years) has long passed, and its forthcoming
disposal must be taken into account (Wright et al., 2015). In this regard
and supporting the continuity of satellite-based volcano monitoring,
Corradino et al. (2019) and Campus et al. (2022), revealed how the
Visible Infrared Imaging Radiometer Suite (VIIRS) sensor aboard Suomi-
NPP and NOAA-20 platforms is an excellent candidate to mitigate for the
decommissioning of TERRA (and AQUA) platform.
Amongst the comprehensive spectral range embraced by VIIRS sen-
sors, the I-5 band, occupying the TIR portion of the spectrum with the
central peak placed at 11.45
μ
m, is of most interest (Table 1). The
intriguing compromise between the spatial (375 m) and temporal res-
olution (up to 4 acquisitions of the same target per day (in constellation;
at the equator)) of this sensor might provide innovative, yet crucial
advancements for the systematic monitoring of low-temperature volca-
nic settings and, in turn, might be key for detecting subtle thermal
anomalies associated with early evidence of volcanic unrest.
In this work, we present an innovative and effective single band TIR-
based algorithm, devised to elaborate VIIRS scenes to detect volcano-
genic thermal anomalies at specic volcanic targets. The algorithm, at
its current stage, is conceived to work solely on nighttime acquisitions,
to prevent contamination from solar irradiance and reectance which
may affect both the number of alerts and the quality of the retrieved
measures (Pieri and Abrams, 2004;Reath et al., 2019a;Pailot-Bonn´
etat
et al., 2023). The algorithm can detect thermally anomalous pixel(s) as
low as 0.5 K above the surrounding background pixels, located as far as
25 km from the volcano’s summit.
We present the results obtained from the analysis of more than 10
years of data acquired at three different volcanoes (Fig. 1a) that have
experienced various types of volcanic activity: (i) Vulcano Island (Italy)
which underwent a hydrothermal crisis in 2021–2022 (Fig. 1b), (ii)
Mount Agung (Indonesia) that erupted, explosively, on November 21st,
2017 preceded by mid-term thermal precursors (Fig. 1c), and (iii), La
Palma (Spain) which produced a large lava ow during the 85-day long
effusive eruption begun on September 19th 2021 (Fig. 1d). To determine
the reliability of our results, we rst conducted a visual selection of the
hotspot-contaminated scenes during the whole 10-year period, and then
compared the algorithm detections against the supervised outputs. To
further validate the results, we performed a cross-correlation against the
same parameters obtained from the higher-resolution ASTER scenes and
with independent measures collected via ground-truth instruments.
Finally, we demonstrate the benets of operating a single band TIR
based hotspot detection system for volcano monitoring.
2. Case studies
2.1. Vulcano
Vulcano Island, Italy (38.39◦N, 14.97◦E), is the southernmost
emerged volcano of the Aeolian Archipelago (Fig. 1b). The island was
formed in the geodynamical context of the Aeolian Arc, originating from
the subduction of the African plate underneath the European Plate
(Keller, 1980;Ellam et al., 1989). In this subduction regime, the
volcanism of Vulcano began ~130 ka (Keller, 1980,De Astis et al.,
2013). The last magmatic eruption occurred from 1888 to 1890 (Selva
et al., 2020). Following this event, a fumarolic eld reaching tempera-
tures up to 700 ◦C was established within the Gran Cratere area (Fig. 1b;
Diliberto, 2017 and Diliberto et al., 2021,Barberi et al., 1991,Capasso
et al., 1994,Chiodini et al., 1995). In 1987, after almost a century of low
and relatively stable activity, a 6-year-long period of unrest led to a
signicant increase in the fumarolic activity (Barberi et al., 1991,
Chiodini et al., 1996,Montalto, 1996), later followed by minor unrests
in 2004–2005, 2009, and 2017 (Granieri et al., 2006,Paonita et al.,
2013,Ricci et al., 2015,Selva et al., 2020).
In September 2021, variations in micro-seismicity associated with
hydrothermal uid circulations, ground deformation, increased fuma-
roles temperatures, and alteration in geochemical composition of
ground-exhaling gasses at La Fossa cone (Fig. 1b; Federico et al., 2023),
prompted the Italian Department of Civil Protection (DPC) to raise the
alert level from green to yellow, effectively announcing the beginning of
a new period of unrest (DPC, 2021).
2.2. Mount Agung
Mount Agung, located on the island of Bali, Indonesia (8.34◦S,
115.51◦E) (Fig. 1c), is considered one of the highest-risk volcanoes in the
country (Ardianto et al., 2021). The stratovolcano, located within the
Sunda arc, is the supercial manifestation of the geodynamic processes
characterising the subduction zone between the Indo-Australian plate
and the Sunda block (Syatri et al., 2022). Extending for 3142 m above
sea level, Mt. Agung is renowned for its explosive activity and, given its
proximity to populated areas, for the human, social, and economic
impact its eruptions have had on the inhabitants of rural villages located
along its slopes (Gunawan et al., 2020). According to the Center for
Volcanology and Disaster Hazard Mitigation of Indonesia (CVGHM),
Mount Agung has erupted four times in the last two centuries: in 1808,
1821, 1843, and 1963 (Gunawan et al., 2020). The 1963 –VEI 5 –
eruption claimed 1148 lives and injured 296 people (Zen and Hadiku-
sumo, 1964). The elevated death toll was mainly related to far-reaching
(10 to 14 km) pyroclastic ows, ejection of large ballistics up to ~6.5 km
from the summit, and ensuing lahars (Kusumadinata, 1964;Surjo, 1965;
Self and Rampino, 2012).
After a 53-year interval of dormancy, the volcano underwent an
eruptive phase between November 21, 2017, and June 13, 2019, rein-
vigorating the scientic focus on the Indonesian volcano (Andaru et al.,
2021). The most recent eruptive phase could have been anticipated
based on signicant ground deformation, occurrence of seismic swarms,
and increasing intra-crater thermal activity since 2017 (Ardianto et al.,
2021,Gunawan et al., 2020,Syahbana et al., 2019,Bemelmans et al.,
2023).
2.3. La Palma
La Palma, Canary Islands, Spain (28.71◦N, 17.91◦W) (Fig. 1d), is one
of the most active volcanoes of the intraplate hot-spot archipelago
(Romero et al., 2022;Montesinos et al., 2013). Subsurface volcanism at
Table 1
Main characteristics of VIIRS sensors.
1
Cao et al. (2014),
2
Cao et al. (2013b),
3
Schroeder and Giglio (2017),
4
Oudrari et al. (2016).
VIIRS (S-NPP/N20) / VIIRS (S-NPP/N20)
Orbit altitude (km) 824
Swath (km) 3060
Equator crossing time 12:40 LT / 13:30 LT
Pixel resolution at nadir (km) 0.375
Pixel resolution at the edge (km) 0.75
ID TIR Band I-5
Spectral range (
μ
m) 10.560–12.428
(1)
Central Wavelength (
μ
m) 11.45
T Min (K) ~ 205
(2)
T Max (K) ~ 380
(2)
NEΔT (K) @ 210 ~ 0.40
(3)
NEΔT (K) @ 267 ~ 0.05
(4)
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
3

(caption on next page)
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
4

La Palma started ~4–3 Ma ago (Carracedo et al., 2001). In the last 125
ka activity was mainly located in the southern part of the emerged is-
land, known as the Cumbre Vieja volcanic complex (Cabrera-P´
erez et al.,
2023). In the last ~500 years, at least 6 eruptions took place within the
Cumbre Vieja ridge, in 1585, 1646, 1677–1678, 1712, 1949, and 1971
(Guillou et al., 1998;Carracedo et al., 1998;Klügel et al., 2000;Carra-
cedo et al., 2001;Casillas Ruiz et al., 2020).
On September 19th, 2021, after a 2-year long period of unrest
characterised by seismic swarms, geochemical anomalies, and ground
ination (e.g., Torres-Gonz´
alez et al., 2020,Padr´
on et al., 2021,Carra-
cedo et al., 2022,Civico et al., 2022,D’Auria et al., 2022,Pankhurst
et al., 2022), a new eruption began. The event was characterised by the
simultaneous emission of lava ows and tephra plumes from multiple
vents located along a NW-SE-orientated ssure (Bonadonna et al., 2022;
Birnbaum et al., 2023). After 85 days of uctuating activity, the eruption
was ofcially announced to have ended on December 13th, 2021 (Plank
et al., 2023). Subaerial lava ows covered an area of ~11.8 km
2
,
affected a total of 3126 buildings (of which 2800 were eventually
destroyed), and led to the evacuation of ~7500 inhabitants (JRC, 2021;
Amonte et al., 2022;Civico et al., 2022). The effused lava volume was
estimated to be 177.6 ±5.8 Mm
3
, with a Mean Output Rate (MOR) of
~24.1 m
3
/s, and a maximum and average lava ow thickness of 65 m
and 15.2m, respectively (Civico et al., 2022;Bonadonna et al., 2022;
Plank et al., 2023).
3. VIIRS sensor and input data
3.1. Visible Infrared Imaging Radiometer Suite (VIIRS) sensors
The Suomi National Polar-Orbiting Partnership (SNPP) and the Joint
Polar Satellite System’s (JPSS) JPSS-1 (NOAA-20 or N20) have been in
orbit since October 2011 and November 2017, respectively (Goldberg,
2018;Xiong et al., 2018). Both Suomi-NPP and JPSS-1 satellites are
placed in a polar orbit at a nominal altitude of 824 km (Cao et al., 2017).
Boasting a cross-track eld-of-view (FOV) of 112.56◦, and a swath width
of 3060 km, each VIIRS sensor provides full coverage of the globe daily
(Cao et al., 2013a). The sensors gather information across 22 spectral
bands, encompassing the electromagnetic spectrum from 0.412
μ
m to
12.01
μ
m. This includes 16 moderate-resolution bands (M-bands), a
panchromatic Day-Night Band (DNB) characterised by a spatial resolu-
tion of 750 m, and 5 imaging resolution bands (I-bands) with a spatial
resolution of 375 m. Amongst the comprehensive spectral range of VIIRS
instruments, the I-5 TIR band centred at 11.45
μ
m (Table 1), is the one
used in this work to detect thermal anomalies of volcanic origin.
3.2. Input data
The proposed algorithm is currently based on VIIRS I-5 nighttime
scenes elaborated by the MIROVA system (Campus et al., 2022). The
datasets are made of daily acquisitions from both Suomi-NPP and
NOAA-20 VIIRS Level 1B radiances (VNP02IMG and VJ102IMG 6-Min
L1B Swath 375 m, respectively; atmospherically uncorrected) and
associated geolocation data products (VNP03IMG and VJ103IMG Im-
agery Resolution Terrain-Corrected Geolocation 6-Min L1 Swath 375 m,
respectively), covering a period of more than 10 years (since January
2012). These are freely distributed by NASA’s Level-1 and Atmosphere
Archive &Distribution System–Distributed Active Archive Center
(LAADS-DAAC) in netCDF4/HDF5 format. The original granules, as per
MIROVA workow, are resampled to a regular 134 ×134 UTM grid
centred on the volcano summit according to the coordinates provided by
the Global Volcanism Program (2023). This step ensures consistency
across all the scenes that now cover an area of ~2500 km
2
. Once the
scenes are processed, these are stored in a local database, together with
information regarding date and acquisition time, zenith angle, geo-
location information, etc., thus ready for further processing (see Coppola
et al., 2016b for details).
4. TIRVolcH algorithm
4.1. Overview
When satellite scenes of the same region are stacked together, at-
pixel resolution long-term statistics can be obtained and, in turn,
anomalous variations from the long-term behaviour can be detected
both in time and space (Tramutoli, 1998). To embrace both conditions,
this algorithm combines spatial and temporal checks to determine the
occurrence of anomalous pixels within the scene.
In the following paragraphs, the different steps of the algorithm will
be detailed. These are (i) initialisation, where the ancillary data and the
pre-processing phase necessary for identifying hotspot-contaminated
pixels (i.e., Conrmed Alerts) are created, (ii) hotspot detection, detail-
ing the distinction between land- and water-dominated scenes, together
with the higher sensitivity statistics applied to high-interest volcano-
specic features (i.e., VSROI) and (iii), outputs, where the volcanologi-
cally relevant parameters are retrieved. The whole process is scripted in
a MATLAB environment, following the workow summarised in Fig. 2.
4.2. Step 1 –Initialization
4.2.1. Ancillary data
To identify hotspot candidate pixels (i.e., Candidate Alerts) and
minimise the number of false alerts the hotspot detection phase requires
Fig. 1. a) Orthographic World projection (M_Map package; Pawlowicz, 2020) showing the location of the three studied volcanoes. b’)Digital Surface Model (DSM)
of Vulcano Island at 1 m spatial resolution, adopted from Ministero dell’Ambiente e della Tutela del Territorio e del Mare (MATTM under Creative Commons License
CC-BY-SA 3.0 IT). The red shape depicts the Main Fumarolic Zone (MFZ) as described by Mannini et al. (2019). The blue line approximates the perimeter of La Fossa
cone. Bold and regular black lines represent equidistant contour intervals at 250 m and 50 m, respectively. b”)Location map of Vulcano island and the Aeolian Arc.
b”’)VIIRS I5 band at 375 m resolution acquired on May 8th, 2014, at 00:48 (UTC), superimposed the DSM and centred over Gran Cratere area. c’)Digital Elevation
Model (DEM; Demnas) of Bali island at ~8.25 m spatial resolution, adopted from the Geospatial Information Agency (Badan Informasi Geospasial—BIG). Available
from: https://tanahair.indonesia.go.id/demnas/#/ (Accessed July 4th, 2022). The red shape depicts the fumarolic and solfataric zone as inferred from (Syahbana
et al., 2019;Andaru and Rau, 2019;Bemelmans et al., 2023). The blue line approximates the perimeter of the main volcanic edice. Bold and regular black lines
represent equidistant contour intervals at 1000 m and 250 m, respectively. c”)Location map of Bali Island and the central portion of the Indonesian archipelago. c”’)
VIIRS I5 band at 375 m resolution acquired on May 19th, 2016, at 17:36 (UTC) superimposed the DEM, centred over Mount Agung. d’)Digital Terrain Model (DTM)
of La Palma island at 2 m spatial resolution, adopted from the Autonomous body National Center for Geographic Information (CNIG) under Creative Commons
License CC-BY 4.0 (Accessed January 31st, 2023). Available from: https://centrodedescargas.cnig.es/CentroDescargas/busquedaSerie.do?codSerie=MDT02. Bold
and regular black lines represent equidistant contour intervals at 1000 m and 250 m, respectively. The yellow star depicts the location of the Tajogaite cone. The red
shape demarks the extent of the lava ow (September 19th-December 13th, 2021) as provided by Copernicus Emergency Management Service (2024). Available
from: https://emergency.copernicus.eu/mapping/ems-product-component/EMSR546_AOI01_GRA_MONIT63_r1_VECTORS/1 (Accessed January 31st, 2023). d”)
Location map of La Palma Island and the Canarian archipelago. d”’)VIIRS I5 band at 375 m resolution acquired on October 17th, 2021, at 02:48 (UTC) superimposed
the DTM and centred over the Cumbre Vieja ridge. (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of
this article.)
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
5

some ancillary parameters that must be provided before running the
whole workow. These parameters, stored in a structure named Vol-
canoes Info, are:
•Coordinates of Interest: the revised location of the volcano’s summit
(or active vent) to compensate for any offset in the GVP list (Latitude
and Longitude coordinates, in degrees; GVP, 2023). These correc-
tions, where needed, ensure that coordinates are representative of
the active vents and/or thermal target locations, where higher
sensitivity statistics are applied (see Section 4.3).
•Land/Water Mask (LW Mask): Land/Water mask of the region (binary
matrix 134 ×134 pixels) as provided with L1B VIIRS products.
•Monthly Reference Scenes (REF): mostly cloud-free scenes of the
investigated region (one per each month of the year; 12 REF per
volcano), representing the average temperature conditions of the
area. These scenes are compared against the observed ones (i.e.,
satellite acquisitions) to detect variations from the normal behaviour
(see Section 4.2.2 and 4.3).
•Regions of Interest (ROIs): Four regions are dened as ROI
1
, extending
for ~1 km from the volcano’s summit, ROI
2
, from ~1 to ~5 km,
ROI
3
, from ~5 to ~12.5 km, and ROI
4
, beyond 12.5 km.
•Volcano-Specic Region of Interest (VSROI): an ad hoc mask within
which higher sensitivity statistics are applied for volcanoes
Fig. 2. Workow of TIRVolcH. The top panel shows input data, followed by (Step 1) Initialisation, (Step 2) Hotspot Detection, and (Step 3) Outputs (i.e., parameters
retrieval). Note, the dashed and grey shaded box in Step 2 indicates iteration of the loop (dashed arrow) until no (further) Candidate Alerts are detected (see text for
details). (For interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
6

exhibiting persistent activity such as crater lakes, fumarolic elds,
etc. (x,y coordinates; see Section 4.3.4).
•Volcano-specic exclusion Region of Interest (VSExcROI ): to reduce the
number of false alerts (binary matrix 134 ×134 pixels); This, simi-
larly to the data-removal mask previously applied by Tramutoli
(1998), is generated for each volcano based on the type of volcanism
and on the distance from the summit reached by lava ows and/or
magmatic products within the last 2 decades. This is established by
evaluating the volcano’s long-term behaviour observed by the
MIROVA system together with ground-based volcanological obser-
vations outlined in bulletins issued by volcanic observatories and
available literature. With that in mind, VSExcROI are generated
following the conditions outlined below:
(i) For targets featuring quasi-exclusively conned activity (i.e.,
crater lakes, lava lakes, fumarolic elds, etc.), anomalous pixels
are searched within a 2 km radius from the summit area;
(ii) For targets featuring effusive and/or explosive episodes conned
within the summit area, anomalous pixels are searched within a
5 km radius from the volcano’s summit;
(iii) For targets featuring far-reaching lava ows (>5 km) or for those
where no thermal activity was detected in the past two decades,
no spatial restrictions apply.
It must be noted that a VSExcROI , where needed, can be promptly
resized to accommodate the ongoing eruptive dynamics.
4.2.2. Pre-processing and generation of monthly reference scenes (REF)
The pre-processing phase enables retrieval of monthly reference
matrices (REF), these representing the average, monthly-based, Bright-
ness Temperature (BT) for each investigated region. The rst time a new
target is processed, the initial step consists of extracting and stacking
together all VIIRS nighttime TIR acquisitions for the designated volcano
from the MIROVA database. The resampled I-5 radiance scenes (134 ×
134 pixels) are converted into BT and then stored in a single cubic
matrix (Fig. 3a), accompanied by temporal and spatial information. A
rst reference image is then generated by averaging all the stacked
scenes, to obtain a time-averaged BT matrix of the investigated region
(Fig. 3b). Notably this reference image, although obtained by averaging
several images contaminated by clouds, has the property of maintaining
the texture and pattern of the cloud-free scene and can be used as a rst,
temporary reference matrix. Hence, each scene is compared against this
reference image and a coefcient of determination (R
2
) is computed.
This coefcient is empirically assumed to be a proxy of the cloud frac-
tion, since the mostly cloud-free images show high R
2
values, while poor
correlation is found for cloudy scenes (Fig. 3c). Images with an R
2
co-
efcient <0.5 are discarded and a new reference scene, containing
mostly cloud-free acquisitions, is generated following the stacking
approach discussed above. The process is repeated a second time to
further reduce the number of unsuitable scenes in the stacked compi-
lation. The remaining acquisitions are divided into months (Fig. 3d) and
the BT time series of each pixel is retrieved. Datapoints (pixels)
exceeding three scaled Median Absolute Deviation (MAD) from the
monthly median BT (i.e., outliers; Leys et al., 2013) are removed from
the matrix and replaced by interpolating the remaining values along the
third dimension (Fig. 3e). This step ensures that most of the remaining, if
any, sparse clouds, wildres, or anomalous pixels are removed. The nal
reference package includes a cubic matrix of 12 scenes, one per every
month of the year (REF; Fig. 3f), and it is stored locally, ready to be used
when recalled within the hotspot detection step described below.
For volcanoes characterised by persistent activity, such as those
related to fumarolic activity (i.e., Mt. Agung or Vulcano) or crater lakes,
the construction of monthly reference matrices requires a further step.
For these targets, the hotspot-contaminated pixels are visually identied
and removed from the matrix; the gaps are lled by assigning the BT
value of the surrounding, non-thermally anomalous pixels, using a bi-
cubic interpolation (Fig. 4).
4.3. Step 2 –Hotspot detection
Within this step, each observed image (OBS, in K) is compared
against its associated reference scene (REF, in K), to detect the presence
of hotspots-contaminated pixels (Conrmed Alerts). This is made through
a series of tests to identify potentially hotspots-contaminated pixels
(Candidate Alerts). The same approach is employed to detect cloud-
contaminated pixels (Cloud Pixels). After the initial set of tests (Initial
thresholding setting; see Section 4.3.1) the algorithm splits the processing
into two separate workows, one for Land-dominated scenes and one for
Water-dominated scenes. The algorithm checks the percentage of land in
the scene using the information contained in the LW Mask le (see
Section 4.2.1). Scenes containing more than 20 % of land pixels (~500
km
2
) are agged as ‘Land-dominated’, whilst those showing less than 20
% are agged as ‘Water-dominated. This main distinction is necessary
because of the very different temperature distribution in the two set-
tings. In particular, while in Land-dominated contexts the algorithm ex-
ploits clustering analysis of the heterogeneous temperature distribution
within the investigated region, this is not possible in Water-dominated
scenes, as water surfaces are mainly characterised by a homogenous
temperature distribution. Furthermore, at volcanoes exhibiting persis-
tent, yet subtle, thermal activity conned within a well-dened area
such as fumarolic elds or crater lakes, the algorithm employs a volcano-
specic ROI (VSROI). These features are characterised by considerably
lower temperatures as compared to those of magmatic bodies, so that
clustering or thresholding approaches usually fail to detect these subtle
volcanic features.
4.3.1. Initial thresholds setting
In this initial step, OBS is analysed and pixels exceeding a xed
threshold (ABSBT) are considered Candidate Alerts:
OBS ≥ABSBT [test 1]
where ABS
BT
is equal to 313.15 K, this being consistent with the
maximum nighttime temperature recorded on Earth, not contaminated
by a hotspot (NOAA, 2024).
Then, a scene of residues (RES, in K; Fig. 5c,f,i) is computed by
subtracting the monthly REF scene (Fig. 5a,d,g) from the BT of the
observed image (Fig. 5b,e,h), so that:
RES =OBS −REF (1)
The 99.5 percentile of RES (pRES99.5) is used to label Candidate Alerts
and Cloud Pixels according to test 2 and test 3, respectively:
RES >ABSDT [test 2]
RES <ABSCL [test 3]
where ABSDT and ABSCL result from the following conditions:
if pRES99.5>10 then 


ABSDT =20
ABSCL =0
,else
if 5<pRES99.5<10 then 


ABSDT =15
ABSCL = − 5
,else
if pRES99.5<5then 


ABSDT =10
ABSCL = − 10
Once tests 1, 2, and 3 are executed, the Cloud Pixels are removed from
RES and, a Z-Score matrix (Z−RES) is computed as:
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
7

Fig. 3. Workow for generating REF matrices (Mount Agung example). a) All scenes stacking b) Temporary REF scene c) Scatter plot showing the typical rela-
tionship between cloud-free and cloud-contaminated scenes. Note that scenes with R
2
<0.5 are discarded. d) Monthly division (only scenes with R
2
>0.5). e) Pixel-
by-pixel outlier removal in monthly REF matrices. f) REF output. See text for details. (For interpretation of the references to colour in this gure legend, the reader is
referred to the web version of this article.)
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
8

Z−RES =RES −RES
σ
RES
(2)
where RES and
σ
RES are the mean and the standard deviation of RES,
respectively. Accordingly, pixels not detected by previous test but
satisfying test 4 are added to the Candidate Alerts:
Z−RES >7 [test 4]
4.3.2. Hotspot detection for Land-dominated scenes
Within the hotspot detection for Land-dominated scenes, two distinct
blocks are executed. The rst operates on a cartesian domain by inves-
tigating the temperature distribution in a scatter plot to identify
anomalous (i.e., outliers) pixels (i.e., datapoints). The second block
works on the images anew, performing contextual statistics on the
scenes.
First Block: On the assumption that an approximately linear rela-
tionship exists between clear sky and hotspot-free OBS and REF, if we
plot these two variables in a scatter plot, most of the data should lie
around the 1:1 ratio. In contrast, scattered and low-density data points
(i.e., outliers) located above the main cluster should represent hot
anomalies (i.e., volcanic surfaces). Datapoints representing clearly
contaminated pixels are usually comprised between the ~0.005 % and
the ~1 % of the whole scene (i.e., 1 to 200 pixels or ~ 0.1 to ~30 km
2
)
and signicantly deviate from the main cluster of uncontaminated data.
This implies that their density distribution is much lower than pixels
representing average conditions.
At rst, REF and OBS matrices are reshaped into vectors and are
plotted on a scatterplot in the x and y axis, respectively (Fig. 6). Any
pixel(s) previously agged as Candidate Alerts or Cloud Pixels are
removed. Then, for each pair of data (x,y) in the scatterplot, the
Euclidean Distance (EDxy) with respect to the eight smallest distances (i.
e., the eight nearest datapoints in the plot with respect to the examined
x,ypair) is calculated as:
EDxy =
(OBSx−REFx)2+OBSy−REFy2
(3)
The choice of selecting eight neighbouring datapoints was made to
resemble the in-scene conditions, namely the number of pixels sur-
rounding any given pixel. Datapoints with EDxy <1 are considered part
of the Main Cluster (MC; non-anomalous data), and a polygon, hereafter
named PolyED, is automatically drawn around them (Fig. 6). At this
point, four concentric envelopes (PolyROIn) are drawn around PolyED by
adding a ROI-dependent buffer (bΔTROIn)so that:
PolyROIn =PolyED +bΔTROIn (4)
with bΔTROIn equal to 0.5, 1, 2 and 4 K for ROI
1
, ROI
2
, ROI
3
, and ROI
4
,
respectively (Fig. 6).
Any datapoint in the scatterplot is also associated with its specic
ROInin the spatial domain (matrix; see Fig. 7a,b,c). Hence, any data-
point (x,y)ROIn in the scatterplot (or its corresponding pixel in the ma-
trix), is considered a Candidate Alert if the following condition is
satised:
(x,y)ROIn ∕∈ PolyROIn [test 5]
where (x,y)ROIn is a datapoint of a specic ROI and PolyROIn is the
associated envelope.Graphically this test is used to identify the data
points located outside the polygon dened for each ROI (PolyROIn ).
Second Block:This block operates on the spatial domain (matrix)
anew, to detect remaining, or missed anomalous pixels within ROI
1
. This
test increases the capability to detect subtle thermal anomalies within
ROI
1
, the region where there is the highest likelihood of encountering
volcanogenic anomalies. To do this, after temporarily removing any
detected Candidate Alert resulting from all previous tests (Test 1 to 5),
the mean and standard deviation of the non-alerted pixels within ROI
1
are computed both for the OBS (OBSROI1and
σ
OBSROI1, respectively) and
for the RES images (RESROI1and
σ
RESROI1, respectively). Pixels of
OBSROI1satisfying the following test are agged as Candidate Alerts:
OBSROI1≥OBSROI1+3
σ
OBSROI1&OBSROI1≥RESROI1+3
σ
RESROI1
[test 6]
The Candidate Alerts are therefore those that have passed at least one
of the tests 1 to 6.
4.3.3. Hotspot detection for Water-dominated scenes
Over Water-dominated scenes the normal temperature distribution
implies that clustering approach described above (Section 4.3.2) fails
because of the homogeneous temperature distribution over sea and/or
water bodies. Hence, we only apply a contextual approach based on the
statistical analysis of the investigated region. The average temperature
(OBS) of the observed scene and its standard deviation (
σ
OBS) are
computed and the following test is performed to detect new Candidate
Alerts:
OBS ≥OBS +10
σ
OBS or OBS ≥pOBS99.5[test 7]
where pOBS99.5is the 99.95th percentile of the whole scene.
Candidate Alerts are removed from the scene, and a new ROI
(ROIIsland) of ~2 km centred on the volcano’s summit is dened. Hence,
the mean (OBSROIIsland ) and standard deviation (
σ
OBSROIIsland ) of ROIIsland
are computed to perform the last test:
Fig. 4. Mount Agung. a) REF scene as resulting from Fig. 3f. Note how a thermal anomaly persists within Agung’s crater (a’). b) REF scene with the persistent
thermal anomaly visually identied and removed (b’). (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of
this article.)
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
9

OBSROI1≥OBSROI1+3
σ
OBSROI1[test 8]
The Candidate Alerts are therefore those that have passed at least one
test amongst tests 1 to 4 and/or 7 to 8.
4.3.4. Hotspot detection within Volcano-Specic ROIs (VSROI )
If a persistent thermal anomaly can be distinguished, a VSROI –
usually extending for ~1 ×1 or ~ 2 ×2 km depending on the size of the
thermal feature(s) –is centred on the hottest pixel (Fig. 8). To detect
Candidate Alerts within the VSROI region(s), a similar approach to that
presented in tests 4 and 6 is applied, yet thresholds are lowered to in-
crease the detection sensitivity exclusively within the VSROI:
Z−RESVSROI ≥5 [test 9]
The already detected Candidate Alerts (tests 1 to 9), if any, are tem-
porary removed from the scene, and the mean and standard deviation of
both observed (OBSVSROI and
σ
OBSVSROI , respectively) and RES (RESVSROI
and
σ
RESVSROI , respectively) scenes are computed. Hence:
Fig. 5. Reference (REF), Observed (OBS), and Residual (RES) scenes for Vulcano (a,b,c), Agung (d,e,f), and La Palma (g,h,i), respectively. Blue patches represent
cloudy pixels resulting from Test 3. Red and orange shapes in f(and f’) and i(and i’) depict the anomalous pixels detected after Test 1 and 2, and after Test 4,
respectively. Note how, in gure c(and c’), tests based on xed thresholds (i.e., tests 1, 2, and 4), failed to detect the thermal anomaly, although a thermally
anomalous pixel can be visually identied. b) acquired on August 17th, 2019, at 01:00 (UTC). e) acquired on June 15th, 2018, at 17:42 (UTC). h) acquired on
October 21st, 2018, at 03:12 (UTC). (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
10

OBSVSROI ≥OBSVSROI +2
σ
OBSVSROI or OBSVSROI ≥RESVSROI +2
σ
RESVSROI
[test 10]
Where pixels satisfying test 10 are added to the Candidate Alerts.
It is worth it to mention that in a single scene multiple VSROI can be
placed over different location, enabling a greater coverage of sparse
thermal features (i.e., fumarolic elds) over the entire volcanic edice.
Detailed evidence on the role of a VSROI over low-temperature targets is
provided as Supplementary Material (S1).
4.3.5. Removal of residual false alerts, background BT computation and
iterations
In this nal step, all Candidate Alerts are subjected to a series of tests
to remove potential false alerts and dene Conrmed Alerts (hotspot-
contaminated pixels).
The algorithm initially veries that the BT of Candidate Alerts
(BTcandidate)is at least 0.5 K above the background temperature (BTbg ).
This ensures that pixels approaching the sensor’s noise limits (see
Table 1) are not identied as hotspot-contaminated. To further minimise
the number of distal false alerts, we test whether their BT is higher than
the theoretical BTbg plus a ROI-dependent threshold (ΔTbg ):
BTcandidate >BTbg +ΔTbg [test 11]
where ΔTbg takes one of the following values: 0.5 K (for VSROI , ROI
1
,
ROI
2
), 0.75 K (for ROI
3
) and 1 K (for ROI
4
).
The BTbg is computed iteratively by removing the Candidate Alerts
from the scene and interpolating the gaps using a bi-cubic interpolation
of the surrounding non-contaminated pixels (Fig. 9). Hence all pixels
failing test 11 are unagged and return to the original matrix. The
procedure is iterated until all Candidate Alerts satisfy test 11.
Finally, a set of spatial conditions is applied to minimise the number
of false alerts by taking into consideration the typical spatial pattern
characterising volcanic phenomena. These are:
If ROI
1
and ROI
2
contain no Candidate Alerts &there are less than 10
Candidate Alerts within ROI
3
and/or ROI
4
&these all fail Test 1 and
2, then these are assumed to be false alerts and are ‘unagged’
[test 12]
If ROI
1
, ROI
3
, and ROI
4
contain no Candidate Alerts &ROI
2
, contains
a single Candidate Alert,then this is assumed to be a false alert and
is ‘unagged’
[test 13]
If Candidate Alerts are located on water bodies (as per LW Mask)&are
not connected to inland Candidate Alerts,then these are assumed to
be false alerts and are ‘unagged’
[test 14]
If Candidate Alerts are located outside the VSExcROI ,then these are
assumed to be false alerts and are ‘unagged’[test 15]
If Candidate Alerts are located more than ~1 km from the shoreline
(as per LW Mask), then these are assumed to be false alerts and are
‘unagged’
[test 16]
Candidate Alerts, namely pixels satisfying at least one Test between 1
and 11 and failing Test 12 to 16 are nally labelled Conrmed Alerts, thus
pixels most likely contaminated by hotspots of volcanic source.
If Conrmed Alerts are detected and, after temporarily removing
these from the scene, the whole process is repeated until no further
Conrmed Alerts are detected between two consecutive runs. At this
stage, following the approach described above, Conrmed Alerts are
removed from the scene and the nal BTbg is calculated to allow esti-
mation of the volcanologically relevant parameters detailed in Section
4.4 and Supplementary Material S2.
4.4. Step 3 –Radiative power and parameters retrieval
With the Conrmed Alerts retrieved, a nal step is dedicated to esti-
mating and storing the volcanological relevant parameters (see Sup-
plementary Material S2). Amongst them, the quantication of thermal
energy sourced by volcanic targets is a key metric to assess the status of
the volcanic system (Wang and Pang, 2023). Previous studies attempted
to estimate the intensity of the volcanic events by mean of statistical
and/or normalised indices (i.e., Tramutoli, 1998;Tramutoli, 2005;
Rabuf et al., 2022), others adopted the maximum temperature above
the background of the alerted pixel(s) as a proxy of the energy involved
(i.e., Ball and Pinkerton, 2006;Calvari et al., 2020;Ramsey and Dehn,
2004;Reath et al. 2019a and 2019b). Nonetheless, as advised by Ramsey
et al. (2023), a maximum pixel temperature may not be representative of
the energy associated with thermal anomalies spreading over an
extensive area. As such, following the approach previously employed by
others (i.e., Ramsey et al., 2023;Corradino et al., 2023;Blackett, 2014;
Mia et al., 2018;Thompson et al., 2022) we estimate the radiative power
(ΦRad; in watt) of each alerted scene as:
ΦRad =Npix
i=1
σ
•
ε
•BT4
alert,i−BT4
bg,i•A(5)
where Npix is the number of Conrmed Alerts,
σ
is the Stefan-Boltzmann
constant (5.67 ×10
−8
W m
−2
K
−4
),
ε
is the surface spectral emissivity,
here assumed to be unity for sake of simplicity (i.e., Pieri and Abrams,
2005;Kervyn et al., 2008;Coppola et al., 2016b), BTalert and BTbg are the
temperature of the alerted pixel and its corresponding background
value, respectively, and Ais the pixel surface area, namely 140,625 m
2
for VIIRS I5 pixels.
Fig. 6. Scatter plot showing the relationship between the vectorised REF (x-
axis) and OBS (y-axis) matrices presented in Fig. 5d and 5e, respectively. The
grey dashed line shows the 1:1 ratio. Note how most of the datapoints (i.e.,
Main Cluster - MC) lie around this line. Orange stars depict the pixels already
agged as candidate alerts during tests 1–4 (see Fig. 5f’; Note, these pixels are
factually excluded from the ED clustering approach (the same applies to cloudy
pixels (blue dots); see text) but have been included in the gure to show their
typical distribution). The black line (Poly
ED) shows the result of the ED-
clustering approach, delimiting the MC of non-anomalous data. Red, yellow,
green, and cyan envelopes depict the buffer (PolyROIn) added to PolyED for ROI
1
,
ROI
2
, ROI
3
, and ROI
4
, respectively. Note, the same plot after running Test 5 and
6 is shown in Fig. 7 (note how anomalous data points have been detected, thus
added amongst the Candidate Alerts). (For interpretation of the references to
colour in this gure legend, the reader is referred to the web version of
this article.)
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
11

5. Algorithm performance
The efciency of the algorithm was assessed on three case studies
(Vulcano, Agung, La Palma), to evaluate its performance both on land-
and water-dominated scenarios, as well as on low- (i.e., fumarolic) and
high- (i.e., effusive) thermal regimes. Validation of the algorithm’s
outputs was conducted by visually inspecting the ~23,000 scenes ac-
quired within the investigated period, labelling the outputs according to
the criteria presented by Massimetti et al. (2020), namely:
(i) True Volcanic Alert (TVA): an anomaly detected by the algorithm
explicitly related to volcanic activity (hot degassing, lava body
exposed, hot eruptive materials exposed and possibly conrmed
by literature or consistent with the background knowledge of
volcano activity), showing a distinguishable thermal inconsis-
tency with the surrounding environment;
(ii) Fires or Anthropogenic alert (FoA): an anomaly detected by the
algorithm expressly and visually related to wildre occurrence
and/or located near human-settled areas;
(iii) False Positive Alert (FPA): an anomaly detected by the algorithm,
visually related to cloud coverage and/or cloud edges, secondary
cloud fringes, geometrical artefacts, or clearly not related to
known or ongoing volcanic processes;
(iv) False Negative Alert (FNA): a visually distinguishable thermal
anomaly undetected by the algorithm, clearly related to volcanic
processes, having a ΔT, with respect to their surrounding pixels,
higher than the thresholds outlined in Section 4.3.5.
Following Genzano et al. (2020), we dene the False Positive Rate
(FPR) as the ratio between the number of false detections (NFPA ), the
latter resulting from the visual inspection, and the total number of
scenes (NScenes) available for the two investigated regions:
FPR =NFPA
NScenes
(6)
From now on, we also name True Volcanic Alerts (TVA) the alerts
conrmed to be of volcanogenic source via visual inspection, thus those
not agged as FPA as:
TVA =NAlerts −NFPA (7)
Finally, we provide the False Negative Rate (FNR) as the number of
missed alerts (NFNA) divided by the number of visually selected scenes
displaying evidence of volcanogenic thermal anomalies (NManual) as:
FNR =NFNA
NManual
(8)
As evinced from Table 2, the algorithm well performs both on land
and islands, with an intrinsic FPR of 1.77 %, 1.90 % and 1.79 % for
Vulcano, Agung, and La Palma, respectively. A detailed investigation
revealed that most of these false alerts (~85 %) are associated with
cloud-edge effects, with the remaining due to geometric distortions
(high satellite zenith resulting in a mismatch between REF and OBS) (see
Supplementary Material S3). Nonetheless, as summarised in Table 2, the
already-minimal occurrence of FPAs can be simply, yet drastically
reduced to 0.47 % by applying distance (from the summit of the volcano,
Fig. 7. a,d) Residual (RES) scenes of Mount Agung and La Palma, respectively, as seen in Fig. 5f,I with ROIs overimposed. Blue patches represent cloudy pixels (Test
3). b,e) ED-clustering. Orange stars represent Candidate Alerts identied with Test 1–4, red stars show Candidate Alerts detected with Test 5–6. Blue datapoints show
the location of cloud-contaminated pixels. The horizontal red dotted line in (e) shows the I-5 channel saturation threshold (see Table 1). c,f) Zoom up over the
eruptive scene of Agung and La Palma, respectively (for acquisition dates see Fig. 5). Filled orange patches show the pixels detected with tests 1–4. Red outlines show
the Candidate Alerts detected after ED clustering and contextual analyses (Tests 5–6). Note, ROIs distortions (attening) in (d) is a graphical artefact due to WGS84
projection. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)
S. Aveni et al. Remote Sensing of Environment 315 (2024) 114388
12

or from the thermally anomalous zone) and geometrical ltering, while
avoiding removal of any relevant TVA (see Supplementary Material S4).
These results are in excellent agreement with those reported for other
hotspot detection systems, featuring FPR ranging from ⪅1 % to >20 % (i.
e., Kervyn et al., 2008,Steffke and Harris (2011),Massimetti et al.,
2020,Coppola et al., 2014,Genzano et al., 2023,Marchese and Gen-
zano, 2022,Lacava et al., 2018,Ramsey et al., 2023,Chu et al., 2020).
6. Thermal trends
6.1.1. Vulcano
Since the late 19th century, following the latest magmatic eruption
(1888–1890; Selva et al. 2020), a fumarolic eld established within
Gran Cratere (see Fig. 1b; Diliberto, 2017 and Diliberto et al., 2021,
Barberi et al., 1991,Capasso et al., 1994,Chiodini et al., 1995). Several
studies provided estimates of the thermal outputs associated with the
fumarolic activity, both collected via ground instruments and space-
borne platforms. Amongst them, combining ASTER and ground truth
data, Mannini et al. (2019) presented the results of a detailed investi-
gation conducted between 2000 and 2019. These authors suggested that
the average diffuse radiative power, at the time of sampling (see Man-
nini et al., 2019 for details), was 1.22 ±0.39 MW (1
σ
). These values are
largely in agreement (R
2
=0.86; Fig. 10c) with the VIIRS-derived
average radiative power obtained proximally to Mannini et al. (2019)
eld surveys (Fig. 10a), namely 1.07 ±0.37 MW (1
σ
). Extending the
period of analysis, Pailot-Bonn´
etat et al. (2023) presented the
ASTER-derived thermal outputs for 2021, thus including the beginning
of the unrest