Remote Sensing of Environment 82 (2002) 135 – 155

[Link]/locate/rse

Automated volcanic eruption detection using MODIS

Robert Wright *, Luke Flynn, Harold Garbeil, Andrew Harris, Eric Pilger
Hawaii Institute of Geophysics and Planetology, University of Hawaii, Honolulu, HI, USA
Received 12 October 2001; received in revised form 22 February 2002; accepted 4 March 2002

Abstract

The moderate resolution imaging spectroradiometer (MODIS) flown on-board NASA’s first earth observing system (EOS) platform,
Terra, offers complete global data coverage every 1 – 2 days at spatial resolutions of 250, 500, and 1000 m. Its ability to detect emitted
radiation in the short (4 Am)- and long (12 Am)-wave infrared regions of the electromagnetic spectrum, combined with the excellent
geolocation of the image pixels ( f 200 m), makes it an ideal source of data for automatically detecting and monitoring high-temperature
volcanic thermal anomalies. This paper describes the underlying principles of, and results obtained from, just such a system. Our algorithm
interrogates the MODIS Level 1B data stream for evidence of high-temperature volcanic features. Once a hotspot has been identified, its
details (location, emitted spectral radiance, satellite observational parameters) are written to an ASCII text file and transferred via file transfer
protocol (FTP) to the Hawaii Institute of Geophysics and Planetology (HIGP), where the results are posted on the Internet (http://
[Link]). The global distribution of volcanic hotspots can be examined visually at a variety of scales using this website, which
also allows easy access to the quantitative data contained in the ASCII files themselves. We outline how the algorithm has proven robust as a
hotspot detection tool for a wide range of eruptive styles at both permanently and sporadically active volcanoes including Soufriere Hills
(Montserrat), Popocatépetl (Mexico), Bezymianny (Russia), and Merapi (Java), amongst others. We also present case studies of how the
system has allowed the onset, development, and cessation of discrete eruptive events to be monitored at Nyamuragira (Congo), Piton de la
Fournaise (Réunion Island), and Shiveluch (Russia). D 2002 Published by Elsevier Science Inc.

1. Introduction Mouginis-Mark, & Horton, 1994; Wright, Flynn, & Harris,

2001), lava domes (e.g. Oppenheimer, Francis, Rothery,
Volcanism is synonymous with heat, and over the last Carlton, & Glaze, 1993; Wooster, Kaneko, Nakada, &
two decades, much research has been dedicated to studying Shimizu, 2000), lava lakes (e.g. Harris, Flynn, Rothery,
how thermal infrared satellite data can be used to monitor Oppenheimer, & Sherman, 1999), and fumarole fields (e.g.
the eruptive state of active volcanoes. This research can be Harris & Stevenson, 1997).
conveniently divided into two groups (Flynn, Harris, & Although the high-spatial-resolution (30 – 120 m per
Wright, 2001). Initial research was concerned with the use pixel) data provided by TM, ETM+, and the recently
of high-spatial-, low-temporal-resolution data from the launched advanced spaceborne thermal emission and reflec-
Landsat thematic mapper (TM), with Francis and McAllister tion radiometer (ASTER) are useful for spatially detailed
(1986) the first to use TM data to detect short-wave infrared thermal analysis, their relatively poor temporal resolution
(SWIR) radiance emitted from a growing lava dome at the (16-day revisit period) means that they do not lend them-
summit of Lascar Volcano in northern Chile. Subsequently, selves to routine thermal monitoring. For example, although
several authors have used TM and its successor, the Landsat the 1991– 1993 eruption of Mount Etna lasted for 473 days,
7 enhanced thematic mapper plus (ETM+), to analyse the only five cloud-free TM acquisitions were possible (Wright,
detailed thermal properties of active lava flows (e.g. Flynn Rothery, Blake, & Pieri, 2000). As a result, volcanologists
have looked to low-spatial- but high-temporal-resolution
data from sensors such as the advanced very high resolution
* radiometer (AVHRR), geostationary operational environ-
Corresponding author. Present address: Department of Geography
and Earth Sciences, Brunel University, Uxbridge UB8 3PH, UK. Fax: +44-
mental satellite series (GOES), and the along-track scanning
1895-203217. radiometer (ATSR) to satisfy the need for an operational
E-mail address: [Link]@[Link] (R. Wright). volcano monitoring capability.

0034-4257/02/$ - see front matter D 2002 Published by Elsevier Science Inc.

PII: S 0 0 3 4 - 4 2 5 7 ( 0 2 ) 0 0 0 3 0 - 5
136 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

AVHRR, GOES, and ATSR have proven to be effective fully global system. AVHRR is only used routinely in the
eruption monitoring tools, in spite of the fact that the North Pacific, while the use of GOES is limited to the
volcanic features of interest are often much smaller than Americas and the Pacific. As a result, volcanically active
the pixel size of the instruments (AVHRR f 1 km; ATSR areas of the globe such as Iceland, the Mediterranean, and the
f 1 km; GOES f 4 km). This success stems from their African Rift remain unmonitored by existing systems. The
relatively high sampling frequency and their ability to moderate resolution imaging spectroradiometer (MODIS),
measure emitted radiance in the SWIR region of the electro- launched on-board NASA’s first earth observing system
magnetic spectrum, a region in which volcanic hotspots emit (EOS) platform ‘‘Terra’’ in December 1999, is the latest
prodigious amounts of energy. By utilising nighttime and space-based sensor to provide the kind of low-spatial-,
daytime passes, AVHRR data can be acquired four times per high-temporal-resolution data previously described. The pur-
day for any given ground target, although this frequency pose of this paper is to show how MODIS has been used as
increases if more than one NOAA satellite is orbiting at any the basis for a global volcanic eruption detection system.
one time and at high latitudes where adjacent swaths overlap
significantly. ATSR data are acquired twice in every 24-h
period, while GOES provides a fan image of the Earth, on 2. Principles of low-spatial-resolution volcano
average, every 15 min. This high sampling frequency monitoring
reduces the chance that a ground target will be obscured
by clouds and is ideal for monitoring dynamic volcanic Before describing the MODIS detection algorithm in
phenomena. AVHRR has been used to provide detailed detail, we will first review the physical principles upon which
chronologies of effusive eruptions at Mount Etna, Sicily low spatial thermal monitoring of hotspots is based. AVHRR
(Harris, Blake, Rothery, & Stevens, 1997; Harris & Neri, in and GOES form the basis of two existing volcano monitoring
press), and is used by the Alaskan Volcano Observatory systems (e.g. Dehn et al., 2000; Harris, Flynn, et al., 2000)
(AVO) to routinely monitor the many active volcanoes that primarily because they detect radiance in spectral wavebands
exist in the North Pacific region (Dehn, Dean, & Engle, centered at 4 and 11 –12 Am. These wavebands coincide with
2000; Schneider, Dean, Dehn, Miller, & Kirianov, 2000). the wavelengths of peak emission for high-temperature
Since 1998, GOES data have been used as the basis for volcanic heat sources and ambient Earth surface temper-
an automated volcano monitoring system based at the HIGP atures, respectively. Research into hotspot detection using
(Harris, Flynn, et al., 2000; Harris et al., 2001). Using low-spatial-resolution satellite data has exploited the fact that
direct-broadcast GOES data provided by the Naval Research the presence of a high-temperature, subpixel-sized hotspot,
Laboratory (Monterrey, CA), the system searches for vol- within an otherwise thermally homogenous pixel radiating at
canic hotspots within several 500  500 pixel image ‘‘tiles’’ ambient temperature, causes the amount of radiance emitted
covering areas including the Lesser Antilles, Hawaii, South in the SWIR to increase dramatically when compared to that
and Central America, New Zealand, and the continental in the long-wave infrared (Dozier, 1981). For example, a
USA. The results are posted on the Internet ([Link] thermally homogenous 1-km pixel containing ground at 300
[Link]) usually within 15 min of initial image acqui- K will emit 0.4 and 9.5 W m 2 sr 1 Am 1 at 4 and 11 Am,
sition. The GOES-based system has provided valuable respectively. However, if an active lava body radiating at 850
insights into the development of lava flow-fields and erup- K occupies 0.05% of this pixel (500 m2) the integrated
tion timings at Kilauea Volcano in Hawaii (Harris, Keszthe- radiance (i.e. from the lava feature and the adjacent back-
lyi, et al., 1997; Harris & Thornber, 1999) and Fernandina ground) emitted at 4 and 11 Am increases to 1.3 and 9.6 W
Volcano in the Galapagos Islands (Mouginis-Mark, Snell, & m 2 sr 1 Am 1, respectively. Thus, although the presence
Ellisor, 2000). It has also been used to elucidate processes of a subpixel-sized hotspot causes emission at 11 Am to
governing the behaviour of potentially explosive silicic lava increase by only 1%, the presence of the same feature causes
domes at Popocatépetl, Mexico (Wright, de la Cruz-Reyna, emission at 4 Am to increase by more than 200%.
Flynn, Harris, & Gomez-Palacios, in press) and Lascar Remotely sensed infrared radiance data therefore provide
(Harris, Pilger, & Flynn, 2002). a tool that is well suited to identifying volcanic hotspots at
Although not used in an operational mode at this time, the pixel or subpixel scale. Section 3 reviews the properties
several studies have illustrated the value of ATSR for mon- of MODIS that make it suitable for hotspot monitoring and,
itoring volcanic activity at Lascar, Chile (Wooster, 2001), in certain aspects, preferential to GOES and AVHRR for this
Unzen, Japan (Wooster & Kaneko, 1998), and Mount Etna purpose.
(Rothery, Coltelli, Pirie, Wooster, & Wright, 2001). ATSR
acquires data in four infrared wavebands at a spatial resolution
of f 1 km at nadir. One nighttime and one daytime image are 3. The Moderate Resolution Imaging Spectroradiometer
acquired for each point on the Earth’s surface, on average, (MODIS)
once every 3 days.
Although AVHRR and GOES are commonly used to MODIS was launched into a sun-synchronous polar orbit
monitor active volcanoes, neither has been integrated into a on-board NASA’s first EOS platform, Terra, in December
 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 137

1999. Terra (previously called EOS AM-1) has a 10:30 a.m.

descending-node equatorial crossing time, providing one
daytime and one nighttime image in a 24-h period. Although
adjacent imaging swaths overlap significantly at high lat-
itudes, gaps between adjacent swaths occur close to the
equator, and, as a result, complete global coverage is
achieved every 2 days. A second MODIS sensor is due to
be launched in April 2002 on the next EOS platform
‘‘Aqua’’ (previously called EOS PM-1). As its former name
suggests, Aqua will have an afternoon (2:30 p.m.) descend- Fig. 1. MODIS across-track response and the definition of the MODIS
pixel. A hotspot at a given ground location (asterisk) can be observed by
ing-node equatorial crossing time.
two adjacent pixels (arrow). Adapted from Fig. 4 in Kaufman et al. (1998).
In terms of hotspot detection, several features are of
particular relevance. MODIS acquires data in the important
4- and 11-Am regions of the electromagnetic spectrum. (Fig. 1). This does not, however, lead to ‘‘double counting’’
However, whereas AVHRR and GOES provide only three of radiance, as although the radiance from a hotspot may be
suitable spectral channels, MODIS provides up to 10. Of shared between two pixels, the sampling period (dwell time)
these, three are used by the MODVOLC algorithm that we is always the same ( f 333 As).
describe in this paper; Bands 21, 22, and 32. Bands 21 and
22 both detect radiance in the same spectral interval (3.929 –
3.989 Am), with a radiometric accuracy (expressed as noise 4. The MODIS volcano detection algorithm:
equivalent temperature variation; NEDT) of 2.0 and 0.07 K, MODVOLC
respectively. Significantly, although Band 22 has a satura-
tion temperature of f 330 K, Band 21, also known as the The algorithm currently uses nighttime MODIS data
‘‘fire channel,’’ was designed to have a saturation temper- only, and it is the development of the nighttime MODVOLC
ature of f 500 K (Barnes, Pagano, & Salomonson, 1998). algorithm that we describe here. However, in order to fully
Unfortunately, Band 21 was only calibrated to a temperature understand the rationale upon which the algorithm was
of f 340 K (3.5% of its full dynamic range) prior to designed, it is first necessary to outline the operational
launch, and, as a result, Band 21 data acquired during the constraints that guided its development. MODVOLC had
first 9 months of MODIS operation are useless for precise to be designed to run at an early stage in the MODIS data
quantitative purposes. However, thanks to a vicarious lunar chain with no other data dependencies (Kaufman et al.,
calibration performed in November 2000, Band 21 is now 1998) and had to be simple enough so as not to impose a
calibrated to a temperature of f 400 K (the brightness drain on the Goddard Space Flight Center (GSFC) Distrib-
temperature equivalent to Lunar reflectance at 4 Am). uted Active Archive Center’s (DAAC) computer resources.
Despite these initial problems, Band 21 is a welcome The MODVOLC algorithm had to be a point operation (i.e.
addition to the MODIS sensor, given the ease with which pixel by pixel), using no more than eight mathematical
the AVHRR and GOES 4-Am channels saturate over high- operations and five bands of MODIS data. The resultant
temperature sources. Band 32 (which images between data files had to small enough to not take up significant
11.770 and 12.270 Am; NEDT = 0.05 K) also has a higher amounts of space at the DAAC and could not be archived on
saturation brightness temperature (420 K) than its long- site (Flynn, Wright, Garbeil, Harris, & Pilger, 2002).
wave infrared counterparts on AVHRR and GOES.
Although the specified geometric accuracy of the MODIS 4.1. A review of previous automatic hotspot detection
pixels was 300 m, postlaunch corrections for bias in the roll, methods
pitch, and yaw of the Terra platform mean that an accuracy
of 200 m can be expected. The geolocation of MODIS Although manual identification of hotspots is relatively
pixels (which is 1 km for Bands 21, 22, and 32) is therefore simple using 4-Am data, automatic (i.e. noninteractive)
well suited to the task of mapping the spatial distribution of detection is more complex because an appropriate threshold
volcanic hotspots. radiance value must be chosen to distinguish those pixels
Each MODIS image covers an area of approximately that contain hotspots from those that do not. This difficulty
1354  2030 km. Scanning is achieved by 10 simultaneous stems from the complex nature of the 4-Am radiance signal,
1-km strips (i.e. one strip per detector giving a 10-km because although 4-Am data are effective in highlighting the
footprint) or 20 and 40 strips for the 500- and 250-m hotspots, they also record variations in the background
resolution channels, respectively (Kaufman et al., 1998). temperature of the surfaces surrounding the hotspot. This
Although the nominal MODIS pixel size at 4 and 11– 12 Am makes it difficult to identify automatically hotspots at the
is 1 km, a point heat source is likely to be represented by global scale using a single 4-Am radiance threshold, as any
one or two across-track pixels. This is because the response such value will fail to account for seasonal and geographical
function of MODIS is triangular in the across-track direction variations in the ambient background temperature. The
138 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

situation is complicated further when using daytime data algorithm to threshold automatically volcanic hotspots in
due to the contribution that reflected sunlight makes to the MODIS data within the imposed operational constraints.
4-Am signal, which can result in anomalously reflective
surfaces appearing ‘‘hot’’ in daytime 4-Am imagery. 4.2. MODVOLC: theoretical background and algorithm
These problems provided the impetus for more robust development
techniques for automatically detecting hotspots in satellite
data to be developed. Such algorithms generally combine The nature of the relationship between temperature and
two distinct approaches: the spectral comparison method radiance at 3.959 Am (MODIS Band 21/22) and 12.02 Am
and the spatial comparison method. The spectral comparison (Band 32) offers the possibility for distinguishing thermally
method is based on the fact that 4- and 11– 12-Am bright- homogenous surfaces of differing temperature based on
ness temperatures should be very similar for surfaces at differences in the slope of the Planck curve between these
ambient temperature, varying by only a few degrees due to points (Fig. 2). Over the range 80 to + 100 jC (sufficient
differences in atmospheric transmissivity and target emis- to cover the full range of ambient Earth surface temper-
sivity (although wide departures can occur over clouds). As atures), the difference between the radiance detected at
such, it requires that the satellite radiance data are converted 3.959 and 12.02 Am increases as the temperature of the
to brightness temperatures, as measured radiance at 4 and surface increases.
11– 12 Am is markedly different, even for thermally homo- Fig. 3 shows contrast-enhanced Band 22 data from a
genous surfaces. The presence of a high-temperature radi- MODIS Level 1B nighttime image of the Big Island of
ator causes the 4-Am brightness temperature to increase Hawaii, acquired on 2 February 2001 (MODIS granule ID:
dramatically when compared to the 11-Am brightness tem- [Link]).
perature recorded for the same pixel. The magnitude of this For several cover types identified in the image, a sample of
temperature difference largely depends on the hotspot size 30 pixels was chosen at random for each group, and the
and its temperature. The spectral comparison method is a difference between the radiance measured in Bands 22 and
much more robust way to detect hotspots than simply 32 (i.e. 22 32) was recorded. An active lava flow at
thresholding raw 4-Am radiance data, because it automati- Kilauea Volcano was also present on this date (Fig. 3,
cally accommodates changes in the ambient background inset), and Band 22 32 values were calculated for pixels
temperature and solar heating effects. The method has been that appeared thermally anomalous. We use Band 22 data
used successfully as a basis for discriminating wildfires because of its higher radiometric resolution when compared
(Prins & Menzel, 1992) and active lava flows (Harris, to Band 21, which improves its ability to discriminate
Swabey, & Higgins, 1995). between surfaces emitting very similar radiances. However,
The spatial comparison method assumes that the 4-Am when Band 22 was saturated, as it was for several lava flow
brightness temperature of a hotspot pixel will be signifi- pixels in this image, we used the unsaturated Band 21 values
cantly different to that of neighbouring background pixels (i.e. 21 32). Fig. 4 shows the results. Clearly, simple
(Flasse & Ceccato, 1996). By assuming that each pixel in an subtraction of the radiance measured in Band 32 radiance
image is a ‘‘potential’’ hotspot and statistically comparing from that measured in Band 22 or 21 does not distinguish
its 4-Am brightness temperature to that of its neighbours, the hot lava flows from the other cover types, with the lava
positive discrimination of hotspots can be achieved when pixels being particularly confused with the ‘‘cold’’ cloud
the temperature difference exceeds a certain threshold. pixel group. This is because although thermally homoge-
Sometimes called the ‘‘contextual approach,’’ it accommo- nous blackbody surfaces can be discriminated using a
dates the influence of seasonal and geographical effects on simple band subtraction operation (Fig. 2), the presence of
the 4-Am signal by tailoring the threshold for positive subpixel-sized hotspots within the lava pixels causes this
detection to the prevailing local background. simple relationship to break down.
Hybrids of the spectral and contextual approaches form Fig. 5 shows a series of Planck curves for blackbodies
the basis of several existing hotspot detection algorithms, radiating at a range of temperatures from 50 to + 13 jC,
such as the GOES automated biomass burning algorithm encompassing the range of nonvolcanic temperatures found
(ABBA; Prins & Menzel, 1994), the MODIS fire algorithm in the nighttime Big Island MODIS data set. Low-temper-
(Kaufman et al., 1998), and the HIGP GOES hotspot ature surfaces (such as clouds and mountain tops) exhibit
algorithm (Harris et al., 2001). However, the constraints relatively small 22[21] 32 values (i.e. a shallow sloping
placed on the development of MODVOLC meant that we Planck curve between 3.959 and 12.02 Am), while warmer
could not utilise either approach. Restrictions on the number surfaces (such as land at sea level) exhibit relatively larger
of mathematical operations available prevented conversion 22[21] 32 values (a steeper sloping curve). Land at a
of the MODIS Level 1B radiance products to brightness homogenous temperature of 13 jC (such as that immedi-
temperatures (thus excluding the spectral discrimination ately adjacent to the active lava flow in this data set)
method), while the necessity for the algorithm to run as a therefore exhibits a greater MODIS Band 22[21] 32
point operation meant that contextual analysis could not be radiance value than do cold clouds. However, the introduc-
performed. As a result, we have developed an adapted tion of a subpixel-sized, high-temperature lava flow to such
 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 139

Fig. 2. Planck curves calculated for thermally homogenous surfaces at 3.959 and 12.02 Am. The 3.959 12.02 Am difference is plotted on the right-hand
ordinate.

Fig. 3. Contrast-enhanced MODIS Band 22 image of the Big Island of Hawaii, acquired at night on 2 February 2001. The image has been rotated 180j. Main
image measures f 2354  1354 km. Active lava flows associated with the ongoing eruption at Kilauea Volcano are obvious as a group of bright pixels in the
second inset image.
140 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

Band 22[21] 32 values for same pixels portrayed in Fig. 4,

normalised by the sum of the radiance’s (i.e. MODIS band
[22 32/22 + 32], or [21 32/21 + 32], where the Band 22
detectors were saturated). This approach, which, for con-
venience, we term a normalised thermal index (NTI), allows
much better discrimination between the lava flows and the
other cover types within the image, offering the possibility
that active hotspot pixels can be automatically discriminated
on the basis of an NTI threshold.
Fig. 7 shows the histogram of NTI values calculated for
all 2 748 620 pixels in the 2 February 2001 Big Island data
set. The distribution is slightly bimodal with lower values
corresponding to relatively cold clouds and the bulk of the
Fig. 4. MODIS band 22[21] 32 values calculated for several different higher NTI values corresponding to the ocean pixels that
pixel cover types within the 2 February 2001 Big Island data set.
dominate the scene. The inset shows the extreme right-hand
tail of the distribution. Manual inspection of the data
a pixel causes the amount of radiance emitted at 3.959 Am to revealed that each of the 13 pixels with an NTI value of
increase at a much greater rate than at 12.02 Am, as > 0.80 contained active lava associated with the ongoing
discussed in Section 2. This has the effect of reducing the eruption at Kilauea Volcano. In this example, hotspot pixels
slope of the Planck curve between 3.959 and 12.02 Am as emitting more radiance at 3.959 Am are characterised by
the size of the hotspot increases. As a result, very cold higher NTI values. This makes them easier to distinguish
surfaces and pixels containing subpixel-sized lava flows than the less intense lava pixels. These pixels, which occupy
cannot be discriminated on the basis of the MODIS Band positions close to the edge of the flow-field and almost
22[21] 32 radiance subtraction. certainly contain larger proportions of ground at ambient
By normalising by the sum of the radiances, however, an temperatures, have NTI values closer to those that character-
index value can be obtained that is weighted to those ise the ocean, land, and cloud pixels that comprise the bulk
surfaces that emit substantial amounts of radiance at 3.959 of the histogram.
Am (i.e. the lava flows) rather than those that emit lower This initial analysis indicated that the NTI, which sat-
amounts of radiance at this wavelength. Fig. 6 shows the isfied the imposed operational constraints, could be an

Fig. 5. Planck curves for a range of blackbody surfaces. Planck curves for thermally homogenous surfaces exhibit an increasing slope between 4 and 12 Am
over the range of temperatures presented. However, this relationship breaks down for pixels containing subpixel-sized radiators, as the amount of radiance
emitted at 4 Am increases dramatically when compared to the amount emitted at 12 Am. As a result, hotspot pixels can exhibit similar slopes to cold, thermally
homogenous pixels.
 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 141

NTI threshold that would be suitable for detecting volcanic

hotspots at the global scale.
Some of the results are shown in Fig. 8. Targets were
chosen to exhibit as much variation in background temper-
atures as possible, with Mount Erebus (Antarctica) and Erta
Ale volcano (Ethiopia) end-member examples. In each
instance, the extreme right-hand tail of the NTI distribution
was composed solely of hotspot pixels (Fig. 8, insets). This
confirmed that the NTI approach could be used to automati-
cally identify volcanic hotspots and allowed us to empiri-
cally establish a threshold that would identify as many
volcanic hotspots as possible without resulting in false alerts.
At Mount Erebus (Fig. 8a), two thermally anomalous
Fig. 6. NTI values calculated for the same pixels as depicted in Fig. 4. The pixels associated with the active summit lava lake were
NTI allows active lava pixels to be distinguished from nonvolcanic pixels apparent through visual inspection of the Band 22 data. A
much more easily than a simple MODIS Band 22[21] 32 difference. normalised index threshold of 0.80 was sufficient to flag
one of these. In order to identify both, a lower threshold of
appropriate means for detecting hotspots with MODIS. As 0.84 would need to be implemented. However, such a
the NTI is based on absolute radiance values, variations in value would be unstable as a means for detecting hotspots at
geography and season will influence its value, as MODIS the global scale. The fact that land and ocean temperatures
Bands 21, 22, and 32 are all sensitive to variations in the in Ethiopia are much higher than those in Antarctica causes
ambient background temperature. However, operational the histogram of NTI values for Erta Ale to shift to more
constraints dictated that we establish a single threshold positive values when compared to that of Erebus. This
value appropriate for the entire globe. For this reason, we positive shift in the NTI histogram means that a threshold
analysed MODIS nighttime images for a range of active of 0.84, although appropriate for Erebus (and by exten-
volcanoes around the world to empirically establish a single sion the many active volcanoes that reside in Kamchatka

Fig. 7. NTI histogram of all 2 748 620 pixels in the 2 February 2001 Big Island MODIS data set. The inset shows the same data at a different scale, to highlight
the fact that pixels containing active lava form a distinct tail to the NTI distribution.
142 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

Fig. 8. NTI histograms and Band 22 images derived from nighttime MODIS granules of (a) Mount Erebus (acquired on 1 May 2001), (b) Erta Ale (acquired on
13 November 2000), (c) Etna (acquired on 8 November 2000), and (d) Soufriere Hills (acquired on 23 February 2001). In each case, the right-hand tail of the
distribution is composed solely of hotspot pixels.
R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 143

Fig. 8 (continued).
144 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

and Alaska), is inappropriate for volcanoes near the equator, respectively. These dimensions and temperatures are more
such as Erta Ale, where such a threshold would result in than sufficient to cover the range of lava lakes typically
>100,000 alerts. In the Big Island data set, where back- observed on Earth (Flynn, Mouginis-Mark, Gradie, &
ground temperatures (and therefore radiances) are somewhat Lucey, 1993; Harris, Flynn, et al., 1999).
lower than those characteristic of Ethiopia, a threshold of The curves presented in Fig. 9 illustrate how complex the
0.84 would result in only 44 alerts. However, visual relationship between the value of the NTI and the nature of
analysis of the image data confirmed that 23 of these would the radiating surface can be. When the lava lake area is less
be false. Fig. 8c and d reveals that an NTI threshold of than 10 5 of the MODIS 1-km pixel, it is essentially
0.80 would also be appropriate for detecting the hotspots invisible to the sensor, and the NTI approaches an asymp-
evident at Soufriere Hills, Montserrat (probably hot pyro- tote that corresponds to that of a thermally homogenous
clastic flows emplaced on and around the 23 March 2001; surface radiating at 25 jC (NTI f 0.86) and 35 jC
GVN, 2001a), and the summit crater hotspots at Mount (NTI f 0.97) for cases A and B, respectively. As the size
Etna, without encroaching on the main body of the NTI of the lava lake increases, the amount of radiance emitted
histogram and resulting in false alerts. increases at a faster rate at 3.959 Am than at 12.02 Am
Based on the analysis of these and many other MODIS (Section 2), MODIS begins to ‘‘see’’ the lake, and the NTI
images, we determined an empirical NTI threshold of increases as a result. As the 3.959-Am bandwidth is respon-
0.80 to be appropriate for nighttime global volcanic sive to the subpixel-sized hotspots and the ground that
hotspot detection. Nevertheless, using a single threshold surrounds them, this occurs at a smaller lake size in Case
value has implications for the sensitivity of our algorithm B than in Case A, where the lake must become larger before
to the presence of hotspots in different geographical it makes itself distinct from the relatively ‘‘warm’’ back-
locations. Fig. 9 provides a simple illustration of this. ground. Potentially, therefore, the Case B lava lake could be
Here, we have modelled the NTI response of MODIS to detected at a smaller size than the Case A lake. However, as
a subpixel-sized lava lake, of the kind present at Erta Ale discussed in Section 4.1, to do so would require setting an
(Case A) and Mount Erebus (Case B). In each case, the NTI threshold that would be so low as to result in many
curves show how the NTI varies as a function of back- thousands of false alerts in Case A. Interestingly, Fig. 9
ground temperature, lava lake size, and lava temperature, indicates that similar lava lakes in equatorial and polar
assuming that the lava lake is composed of a chilled lava regions should result in almost the same NTI. However,
crust (radiating at 400 jC) within which cracks expose the results presented in Fig. 9 are difficult to generalise upon
molten lava from the lake interior (radiating at 1000 jC). as they are entirely model dependent.
The size of the lake varies between 10 and 10,0000 m2, Fig. 9 makes an important point: While the NTI index is
while the area of high-temperature cracks on the lake useful to detect volcanic hotspots, it is not in itself a basis
surface was allowed to vary between 0.01% and 100% for quantifying the intensity of the hotspot, or as a basis for
of the lake surface itself. Background temperatures were comparative analysis of geographically disparate volcanoes.
assumed to be 25 and 35 jC for Cases A and B, Interpreting the absolute NTI value is fraught because of the

Fig. 9. Modelled variation in the NTI calculated for a pixel containing a subpixel-sized active lava lake. Case A refers to the NTI of a lava lake of the kind that
might exist at warm equatorial latitudes (e.g. Erta Ale), while Case B models the NTI response of an identical lava lake at cold Polar latitudes (e.g. Mount
Erebus). The horizontal dashed line denotes the empirically determined NTI threshold value of 0.80, used by the MODVOLC algorithm.
 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 145

effect of variations in background temperature on the 3.959- The MODVOLC algorithm checks whether a pixel is a
and 12.02-Am signals, as well as variations in the detailed daytime or a nighttime pixel by reference to its solar zenith
temperature characteristics of the hotspots themselves. For angle (solar zenith, satellite zenith, and satellite azimuth
example, the variations in background temperature that we
model in this instance can result in a difference in NTI of
50% for a given lava lake. Moreover, variations in the
composition of the hotspot will also result in large variations
in NTI. For example, a 1000-m2 Case B lava lake, which
was vigorously overturning at the time of MODIS overpass
(i.e. the chilled lake surface crust foundering en masse,
exposing large areas of molten lava), would be indistin-
guishable, on the basis of the NTI value, from a 5000-m2
Case B lava lake in a state of relative quiescence (i.e. stable
crust, relatively few cracks). It is therefore important to bear
in mind that while the NTI is a reliable way to automatically
detect volcanic hotspots, it should not be used as the basis
for quantitative analysis. As we will go on to describe in
Section 4.3, there is in fact no need for reliance on the NTI
value as an indicator of hotspot intensity, because absolute
radiance in Bands 21, 22, 28, 31, and 32 is recorded and
reported for each pixel designated as a hotspot by the
MODVOLC algorithm.

4.3. Implementation of the algorithm

The MODVOLC algorithm has been incorporated into

the EOSDIS Core System (ECS), used by NASA’s Goddard
Earth Science (GES) DAAC to process the MODIS data
stream. Fig. 10 illustrates the processing steps diagrammati-
cally. The algorithm uses the MODIS Level 1B data stream.
Although MODIS is a 12-bit instrument (i.e. 4096 discrete
measurement levels), the Level 1B data are converted from
this 12-bit representation to a 16-bit unsigned scaled integer
(SI) format (i.e. 0 – 65 535). These unsigned integers are
subsequently converted to radiance (in W m 2 sr 1
Am 1) using calibration parameters stored within each
MODIS hierarchical data format (hdf) Level 1B granule.
Although the SI representation allows 65 535 measure-
ment levels, only the lower half of the scale (i.e. 0 –32 767)
is used to store science data. SIs in excess of 32 767 are
designated ‘‘reserve values,’’ and are used to indicate
various types of ‘‘bad’’ data. For example, a SI of 65 531
indicates a dead detector, whereas a value of 65 534 indi-
cates that the original Level 1A data were missing from the
raw MODIS granule. The MODVOLC algorithm identifies
bad pixels on the basis of these reserved SI values and
excludes them from further processing. Although Bands 22
and 21 were afflicted by dead detectors at launch, they do
not occupy the same relative position on the MODIS
imaging plane. As a result, Band 22 or 21 data are always
available for each of the ten 1-km strips that comprise the
MODIS footprint, and no hotspots are missed as a result.
Fig. 10. Flow diagram illustrating the implementation of the MODVOLC
The problems caused by dead detectors, electronic crosstalk,
algorithm. Hotspot details written to the MODVOLC file include time
and other instrument noises have subsequently been alle- detected (year/month/day/hour/minute) location (latitude/longitude), spec-
viated by a switch from the initial ‘‘Side A’’ to the redundant tral radiance in Bands 21, 22, 28, 31, and 32 (in W m 2 sr 1 Am 1),
‘‘Side B’’ MODIS electronics in November 2000. satellite zenith, satellite azimuth, and solar zenith (all in degrees).
146 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

are recorded for each pixel in the MODIS granule and m 2 sr 1 Am 1), satellite zenith, satellite azimuth, and
stored in the .hdf file). This is done on a pixel-by-pixel solar zenith (all in degrees).
basis to account for those situations in which the termi- The MODVOLC files are transferred via file transfer
nator falls within the 1354-km MODIS swath. This sorting protocol (FTP) to HIGP, University of Hawaii, where the
procedure is included in preparation for a daytime version results are displayed on the Internet ([Link]
of the MODVOLC algorithm, which is currently under [Link]). The front page of the website provides a global
development, and will be implemented as soon as com- view of all hotspots detected during the previous 24-h
puter processing resources at the GSFC DAAC permit. If period and can be interrogated at larger scales for the
the pixel is a nighttime pixel, the algorithm establishes incidence of hotspots at individual volcanoes. The appear-
whether it is saturated in Band 22. This is done, again, on ance and functionality of the site are constantly being
the basis of a SI value reserved for detector saturation developed and improved. However, all revisions allow the
(SI = 65 533). If Band 22 is unsaturated, the NTI is data contained in the original MODVOLC files to be easily
calculated using Bands 22 and 32. If Band 22 is saturated, retrieved, and allow quick access to the MODVOLC hotspot
the NTI is calculated using Bands 21 and 32. The value is data archive.
then compared against the empirically determined NTI
threshold value of 0.80. Pixels with NTI > 0.80 are
considered as hotspots, and the details are written to an 5. Results
ASCII text file given the GSFC Earth Science Data Type
name ‘‘MODVOLC.’’ For each hotspot pixel, the MOD- 5.1. Volcanic hotspots
VOLC file records the time when it was detected (year/
month/day/hour/minute), its location (latitude/longitude), MODVOLC has detected hotspots (Table 1) associated
spectral radiance in Bands 21, 22, 28, 31, and 32 (in W with a wide range of eruptive styles at 33 different volca-

Table 1
Volcanoes that have been identified as ‘‘hotspots’’ by MODVOLC and the number of times they have been detected for the periods 1 October 2000 to 30
November 2000 and 1 February 2001 to 31 May 2001
Volcano name (location) Latitude/Longitude MODVOLC alerts Nature of activity
Ambrym (Vanuata) 16.25jS, 168.12jE 8 Lava lake
Arenal (Costa Rica) 10.46jN, 84.70jW 1 Vent activity
Bagana (Papua New Guinea) 6.14jS, 155.19jE 10 Lava dome
Bezymianny (Russia) 55.98jN, 160.59jE 10 Lava dome/pyroclastic flow
Big Ben (Heard Island) 53.10jS, 73.51jE 5 Vent activity/lava flows
Cleveland (Aleutian Islands) 52.82jN, 169.95jW 3 Vent activity
Erebus (Antarctica) 77.53jS, 167.17jE 58 Lava lake
Erta Ale (Ethiopia) 13.60jN, 40.67jE 49 Lava lake
Etna (Sicily) 37.73jN, 15.00jE; 63 Vent activity, lava flows
Ibu (Halmahera) 1.48jN, 127.63jE 2 Vent activity
Karangetang (Siau Islands) 2.47jN, 125.29jE 10 Lava dome
Kilauea (Hawaii) 19.42jN, 155.29jW 87 Lava flows
Láscar (Chile) 23.37jS, 67.73jW; 4 Lava dome
Lopevi (Vanuata) 16.50jS, 168.34jE 1 Lava flows/vent activity
Mayon (Philippines) 13.25jN, 123.68jE 2 Lava dome
Merapi (Java) 07.54jS, 110.44jE 20 Lava dome/pyroclastic flow
Michael (South Sandwich Islands) 57.78 S, 26.45 W 2 Unknown
Miyake-jima (Japan) 34.08jN, 139.53jE 1 Vent activity
Nyamuragira (Zaire) 1.40jS, 29.20jE 18 Lava flows
Pacaya (Guatemala) 14.38jN, 90.60jW 1 Vent activity
Piton de la Fournaise (Réunion Island) 21.23jS, 55.71jE 21 Lava flows
Popocatépetl (Mexico) 19.02jN, 98.62jW 38 Lava dome
Rabaul (Papua New Guinea) 4.27jS, 152.20jE 3 Vent activity
Santa Maria (Guatemala) 14.75jN, 91.55jW 39 Lava dome, block lava flow
Semeru (Java) 8.11jN, 112.92jE 14 Vent activity
Shiveluch (Russia) 56.65jN, 161.36jE 13 Lava dome/pyroclastic flow
Soufriere Hills (Montserrat) 16.72jN, 62.18jW 28 Lava dome/pryoclastic flow
Stromboli (Italy) 38.79jN, 15.21jE 1 Vent activity
Tinakula (Santa Cruz Islands) 10.38jS, 165.80jE 3 Vent activity/lava flows
Tofua (Tonga) 19.75jS, 175.07jW 1 Caldera
Ulawun (Papua New Guinea) 5.04jS, 151.34jE 3 Lava flows
Villarrica (Chile) 39.42jS, 71.93jW 1 Lava lake
Yasur (Vanuatu) 19.52jS, 169.43jE 2 Vent activity
 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 147

noes worldwide since October 2000 (Flynn et al., 2002 MODVOLC algorithm for monitoring global volcanism. As
provide a summary of some of these results). In this section, noted in a previous section, gaps between adjacent MODIS
we review some of these results to illustrate the utility of the swaths occur close to the equator, reducing the frequency of

Fig. 11. Image maps taken from the MODVOLC website ([Link] showing hotspots (green dots) detected by the MODVOLC algorithm
at (left to right, top to bottom); Erebus, Kilauea, Erta Ale, Popocatépetl, Soufriere Hills, Mount Etna, Merapi, and Bezymianny. The location of each hotspot as
reported by MODVOLC is shown alongside the published location (Simkin & Seibert, 1994) for comparison. Dashed boxes are 5j in latitude and longitude.
The MODVOLC hotspots have a colour gradation, with green – yellow – red indicating increasing intensity of the 3.959-Am emitted radiance.
148 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

Table 2 Hawaii, a fact reflected in the large number of alerts

Location of the hotspots detected at Erta Ale by the MODVOLC algorithm
reported by MODVOLC for this volcano (87 between 1
during February 2001. Simkin and Siebert (1994) report a location of
40.67jE, 13.60jN for Erta Ale’s summit caldera, within which these October 2000 to 30 November 2000, and 1 February 2001
hotspots lie to 31 May 2001). The hotspots observed to date have all
Date MODVOLC hotspot location been sited close to the coast, several kilometers southeast of
(latitude/longitude) the current active vent (at Pu’u ’O’o). This is entirely
2 February 2001 40.653jE, 13.605jN consistent with the nature of the activity at the time of
4 February 2001 40.659jE, 13.604jN image acquisition, whereby surface pahoehoe lava flows fed
8 February 2001 40.662jE, 13.601jN by tubes extending from the vicinity of Pu’u ’O’o are often
13 February 2001 40.661jE, 13.603jN most abundant on the coastal plain.
15 February 2001 40.660jE, 13.601jN
Mount Etna is the largest and most active volcano in
17 February 2001 40.660jE, 13.600jN
19 February 2001 40.663jE, 13.602jN Europe with a documented history of eruptions stretching
22 February 2001 40.659jE, 13.600jN back almost 2000 years (Chester, Duncan, Guest, & Kil-
24 February 2001 40.660jE, 13.602jN burn, 1985). Given this level of activity, it is not surprising
that MODVOLC regularly detects hotspots at its summit.
data collection for volcanoes at these latitudes. Cloud cover Historically, eruptions at Etna can be divided into two main
also prevents detection of volcanic thermal anomalies. groups: those that initiate within its summit crater region,
However, in spite of these sampling limitations MODVOLC and those that begin at flank vents on the lower slopes of the
has reliably detected hotspots at both permanently and edifice (Chester et al., 1985). The ability of MODVOLC to
sporadically active volcanoes. As noted in Section 2, the locate hotspot pixels to an accuracy of f 200 m will
MODVOLC algorithm detects hotspots, but it does not therefore prove valuable in determining whether detected
discriminate the nature of the hotspot. As a result, their hotspots are associated with the former or the latter. In the
interpretation can be fraught with difficulty, as the spectral example presented in Fig. 11, the hotspot is clearly confined
signature of biomass burning could be mistaken for eruptive to the summit region when, consistent with field reports
activity. Accurate location of where the hotspots occur describing the predominance of high-temperature Strombo-
therefore becomes critical for determining the cause of the lian explosions within the Bocca Nuova, one of four active
anomaly. The accuracy of the MODIS Level 1B geolocation craters at Etna’s summit at this time (GVN, 2001b).
( f 200 m) provides a valuable means for assessing the Since eruptive activity began in 1995, Soufriere Hills
veracity and reliability of the MODVOLC algorithm, as we Volcano has devastated the southern sector of the tiny
can confidently compare MODVOLC detected hotspot Caribbean Island of Montserrat. The volcanic threat is posed
locations to the positions of known volcanic activity. Fig. by the intermittent collapse or explosive disruption of a lava
11 presents maps compiled from the MODVOLC website dome growing at the volcano’s summit, which results in
showing detected hotspots for a range of volcanoes. For pyroclastic flows sweeping down valleys on the flanks of
comparison, the latitude and longitude of the hotspot as the volcano (Young et al., 1998). On 23 February 2001,
recorded by MODVOLC is shown alongside the coordinates MODVOLC detected six hotspot pixels at Soufriere Hills
of the summit of the volcano as published by Simkin and Volcano (Figs. 8d and 11). Although the growing lava dome
Seibert (1994). is much smaller than a single 1-km MODIS pixel, this
Persistently active lava lakes are known to have existed relatively large number of hotspots is easily reconciled with
at Erta Ale (Ethiopia) and Mount Erebus (Antarctica) for at reference to field reports that describe the emplacement of
least 200 years, although ground observations are rare extensive pyroclastic flows on and around the 23 February
because of their geographical isolation (Harris, Wright, & eruption (GVN, 2001a).
Flynn, 1999; Tazieff, 1994). MODVOLC regularly detects The previous discussion has illustrated the value of
the hotspots associated with these lava lakes (Fig. 11; Table MODVOLC as a hotspot detection tool. We now recount
1). The lakes in question are significantly smaller than the 1- three case studies that show how MODVOLC can be used
km MODIS pixel size and, as a result, only one or two to monitor the onset, development and cessation of discrete
hotspot pixels are detected by MODIS at these volcanoes. eruptive events. We do this using case studies taken from
Table 2 shows the locations reported for hotspots observed eruptions at Nyamuragira, Shivulech, and Piton de la Four-
at Erta Ale during February 2001. The accuracy is clearly naise.
good enough to be confident that the hotspots are volcanic
in origin. 5.1.1. Nyamuragira
Spatially extensive lava flows are an almost permanent Nyamuragira is situated on the border between Rwanda
characteristic of the ongoing eruption at Kilauea Volcano in and the Democratic Republic of Congo. On 6 February

Fig. 12. Top: The total amount of spectral radiance detected by MODVOLC (i.e. summed for all pixels identified as hotspots by the algorithm) during the
February to March 2001 eruption of Nyamurgira, Democratic Republic of Congo. Bottom: Image maps taken from the MODVOLC website showing the
distribution of the detected hotspots, with a field map of the location of the lava flows responsible for comparison (map adapted from GVN, 2001c).
R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 149
150 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

2001, a large eruption began, which lasted until mid-March VOLC. This illustrates a significant limitation of our algo-
2001, during which time lava flows were erupted extending rithm. Without the original image data or, as in this instance,
10 –20 km along the south and west flanks of the volcano reliable field observations, it is impossible to know how the
(GVN, 2001c). Direct field observations are difficult presence of meteorological or volcanic ash clouds hinders
because the area is held by the rebel group ‘‘Congo Rally the detection process. This poses problems when interpret-
for Democracy.’’ As a result, MODVOLC provided an ing MODVOLC radiance time series as a short-term
important chronology of the spatial and temporal develop- decrease in detected hotspot radiance may not necessarily
ment of the eruption (GVN, 2001d). Fig. 12 shows how the be due to changes in the intensity of the volcanic activity.
total amount of radiance emitted at 3.959 Am, as reported by
the MODVOLC algorithm, varied over the duration of the 5.1.2. Shiveluch
eruption. Field reports indicate that the eruption began on 6 Shivulech is one of Kamchatka’s largest and most active
February 2001 with the opening of two fissures on the north volcanoes, with a long history of dome growth events and
and south flanks of the volcano. The northern vent erupted explosive eruptions. From relatively low levels in January
flows that extended 15– 20 km due north of the volcano, and February 2001, activity increased during April 2001,
while the southern vent erupted flows that extended a and a sharp increase in seismicity prompted the Kamchatka
similar distance to the southwest. The amount of radiance Volcanic Eruptions Response Team (KVERT) to increase
detected by MODVOLC rises quickly to a peak within the hazard status of the volcano on 24 April 2001 (GVN,
several days of the beginning of the eruption and then falls 2001e). MODVOLC reported hotspot activity on 30 April
by almost an order of magnitude. This is consistent with the (Fig. 13, confirmed by manual analysis of AVHRR data by
common tendency of large effusive eruptions to exhibit AVO scientists; GVN, 2001e). MODVOLC reports a strong
rapidly increasing levels of lava production during the early increase in the intensity of hotspot activity on 1 May. This
stages of the eruption (a short ‘‘waxing’’ phase; Wadge, thermal anomaly preceded intense seismic tremor on 7 May,
1981) before reverting to a much longer phase of exponen- believed to have been caused by the growth of a new lava
tially decreasing lava production (a long ‘‘waning’’ phase; dome (GVN, 2001e). On 19 May 2001, at 04:46 GMT,
Harris, Murray, et al., 2000). However, the waning phase is Shiveluch erupted explosively for 40 min, producing a 10-
interrupted by a second sharp rise in emitted radiance km-high ash cloud and emplacing pyroclastic flows and hot
detected by MODVOLC on 2 March. Field reports indicate avalanches in the immediate vicinity of the active dome.
that a new eruption that began on Nyamuragira’s southern The radiance peak detected by MODVOLC some 7 h later
flank on the 2 March 2001 (GVN, 2001c), thus explaining (point iv on Fig. 13) almost certainly relates to the emplace-
this perhaps unexpected radiance increase. Although we ment of these hot pyroclastic flows.
have no ground confirmation regarding when the eruption
ended, MODVOLC reports a decreasing radiance trend over 5.1.3. Piton de la Fournaise
the first half of March, with the last hotspot recorded on 13 Piton de la Fournaise is a basaltic shield volcano that
March 2001. This is consistent with fields reports that forms the southeastern half of Réunion Island, 700 km east
describe how the intensity of the lava flows decreased in of Madagascar. It is one of the most active oceanic volca-
mid-March (GVN, 2001c). noes in the world, having erupted more than 100 times in
Fig. 12 also shows field maps of the location of the lava the last 300 years (GVN, 2001f). Increasing volcanic tremor
flows produced during this eruption (GVN, 2001c) along- culminated in a lava eruption on 27 March 2001 from a
side two hotspot maps taken from the MODVOLC website. fissure on the southeastern flank of the volcano (GVN,
Although the MODVOLC algorithm reliably documents the 2001f). MODVOLC immediately detected the hotspots
relative position and dimensions of the two major flows to associated with the onset of the eruption (Fig. 13). The
the north and east of the volcano, a third flow to the hotspot alignment MODVOLC reports (inset, Fig. 13) is
southwest goes unrecorded. Field reports (GVN, 2001c) clearly consistent with the reported direction of the flows,
show that the majority of the ash erupted during the episode which extended along the southeastern flank of the volcano
was deposited to the southwest of the volcano, exactly (GVN, 2001f). The eruption ended on 4 April 2001 (GVN,
coincident with the location of this undetected flow. It is 2001f). MODVOLC records a declining trend in emitted
likely that the cloud that transported this ash obscured the radiance between 30 March and 2 April. No hotspots were
flow from MODIS, thus preventing its detection by MOD- detected by MODVOLC after this time.

Fig. 13. Top: The total amount of spectral radiance detected by MODVOLC (i.e. summed for all pixels identified as hotspots by the algorithm) during the
April – May 2001 eruption of Shiveluch, Russia. Image maps taken from the MODVOLC website showing the spatial location of the detected hotspots are also
included. Images i, ii, iii, and iv were acquired on dates corresponding to points i, ii, iii, and iv on the adjacent graph. Bottom: The total amount of spectral
radiance detected by MODVOLC (i.e. summed for all pixels identified as hotspots by the algorithm) during the March 2001 eruption of Piton de la Fournaise,
Réunion Island. The inset shows an image map taken from the MODVOLC website on 28 March 2001. The orientation of the hotspot is clearly consistent with
reported east – south – east direction of the lava flows.
R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 151
152 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

5.2. Further applications: wildfires and industrial hotspots section, although they are clearly consistent with distribu-
tions determined independently (Fig. 14b,c). A notable
The MODIS Fire Team have developed a complex exception is Rothery, Thorne, and Flynn (in press), which
algorithm to detect, monitor, and characterise wildfires in provides detailed ground validation (Table 1, therein) of
order to quantify pyrogenic trace gas and aerosol emissions MODVOLC hotspots observed in Britain over a 6-month
during biomass burning events (Kaufman et al., 1998). period at the beginning of 2001, attributable to sources as
Although the MODVOLC algorithm was primarily intended diverse as petrochemical factories, steel works, warehouse
as a means for rapid detection of volcanic eruptions, it can fires, and ‘‘foot-and-mouth’’ pyres. Our confidence in the
also be used to detect wildfires and establish how their performance and stability of the MODVOLC algorithm is
distribution changes through time. Fig. 14a shows all further supported by several sources of evidence. Firstly,
hotspots detected by MODVOLC during October 2001. hotspots detected for volcanoes such as Erta Ale, Erebus,
Although vegetation fires occur all over the planet, there and Popocatépetl (where the source of radiance is known to
are certain areas where they are particularly frequent and be contained within a relatively small and spatially invariant
intense, with 80% of forest fires occurring within the summit crater) consistently fall within a few seconds of arc
tropics. The distribution of hotspots detected by MOD- of the accepted location of the summits. Secondly, spatially
VOLC reflects this pattern, with high concentrations occur- extensive contiguous hotspots, such as those detected at
ring in the savanna regions of central Africa, Central Nyamuragira and Piton de la Fournaise, have dimensions
America, the Orinoco and Amazonia Basins of South and orientations wholly consistent with field-based obser-
America, coastal regions of Australia and Southeast Asia. vations. Thirdly, although every granule for the entire globe
These are all fire-prone regions traditionally associated with is subjected to the MODVOLC algorithm, we only get
high incidences of satellite-detected hotspots. Fig. 14b and c hotspots in areas where we expect to see them. That there
shows hotspots detected by MODVOLC over South Amer- are few, if any, alerts in regions that we do not expect to
ica on 23 March and 7 June 2001. Wildfires detected by the contain hotspots (i.e. oceans, uninhabited polar regions)
Brazilian National Institute for Space Research (INPE) while consistently detecting large numbers of hotspots in
using AVHRR data on the same day are shown for compar- areas where we do expect to find them (e.g. South America,
ison. Although there are differences, the fire distribution central Africa, Australia) indicates that the algorithm is both
detected by MODVOLC bears close resemblance to that efficient and stable.
reported by INPE, in particular, the shift in the main burning Although the MODVOLC algorithm appears to provide a
zone from the Orinoco basin (Venezuela and Colombia) in robust tool for detecting hotspots at the global scale, there
March 2001 to the Amazon basin (Brazil) in June 2001. are several limitations that should be addressed. The first is
Although we do not attempt a thorough crossvalidation of that the algorithm does not detect all thermally anomalous
the relative accuracy of the two products, it appears that pixels associated with a particular anomaly. For example,
MODVOLC can be used to monitor the spatial and temporal manual inspection of the Big Island MODIS data set
distribution of wildfires. depicted in Fig. 3 showed that although 21 pixels were
MODVOLC has also proven itself capable of identifying thermally anomalous with respect to the surrounding back-
a range of persistent industrial hotspots, especially gas and ground, MODVOLC only classified 13 as hotspots. This
oil flares. Fig. 14d – g provides a range of examples. problem is common to all automatic hotspot monitoring
Hotspots are routinely detected by MODVOLC in the methodologies and stems from a desire to eliminate false
Campos Basin off the coast of Brazil (United States Geo- alerts by setting relatively high threshold values. However,
logical Survey Geological Province Code No. 6035; Klett, the effect is compounded by the sampling regime of the
Ahlbrandt, Schmoker, & Dolton, 1997), the Villahermosa MODIS sensor. It is likely that the radiance from any
Uplift (Gulf of Mexico; No. 5305), the Rub Al Khali basin hotspot will be shared between two adjacent pixels (Fig.
(Persian Gulf; No. 2019), and the Niger Delta (No. 7192), 1), and this dilution can result in only one of the two pixels
amongst others. exceeding the NTI threshold or, in certain cases, the hotspot
not being detected at all. We must stress, however, that the
5.3. Validation and limitations of the algorithm MODVOLC algorithm serves merely to facilitate rapid
detection of hotspots, and for this purpose, it is not neces-
It is difficult to validate the accuracy, for example, of sary to detect all thermally anomalous pixels associated with
each of the wildfire hotspots presented in the previous a given thermal anomaly. Although the algorithm returns

Fig. 14. (a) Composite image showing all hotspots detected by MODVOLC during October 2001. (b and c) Wildfires (green dots) detected by MODVOLC in
South America on (b) 27 March and (c) 7 June 2001. Wildfires (red dots) detected by INPE using AVHRR data on the same day are shown for comparison
(Source, INPE-DSA). (d – g) Hotspots associated with gas and oil flares detected by MODVOLC during February 2001 at: Campos Basin off the coast of
Brazil, the Villahermosa Uplift in the Gulf of Mexico, the Rub Al Khali basin in the Persian Gulf, and the Niger Delta off the coast of West Africa, respectively.
Each dashed box measures 5j in both latitude and longitude.
R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 153
154 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155

basic quantitative information for each hotspot detected, it is operational, we anticipate that MODVOLC and its associ-
not intended to be an all-encompassing volcanic thermal ated website will provide an integrated resource for detect-
analysis tool, and the original image data should be ing, archiving, and disseminating information regarding the
requested from the GSFC DAAC if the user intends to global occurrence of volcanic eruptions.
perform detailed thermal analysis. That the algorithm does
not return image data as an output is itself a source of
confusion when interpreting hotspot occurrence. Clearly, Acknowledgements
clouds prevent detection of hotspots, but without actual
image data, it is impossible to deduce their impact. The authors were supported by NASA Pathfinder grant
Implementation of the MODVOLC algorithm was ini- NAG5-9413. MODIS data were made available via the
tially hampered by data processing bottlenecks at the GSFC GSFC DAAC. We would like to thank Rich Hucek (NASA,
DAAC (Flynn et al., in press), and as a result, several data GSFC) for, thus far, 7 years of valuable assistance without
gaps exist in our hotspot archive. However, this situation which MODVOLC would not have been implemented. We
has improved markedly and at the time of writing (February would also like to thank members of the MODLAND Fire
2002), the time lag between acquisition of the raw MODIS Team (Chris Justice, Yoram Kaufman, Elaine Prins, Louis
data and posting of the hotspot results on the MODVOLC Giglio, Jackie Kendall) for insightful discussions regarding
website has fallen to between 24 and 48 h. Thus, although global hotspot detection. The paper benefited from informal
we are not yet detecting volcanic eruptions in real time, we discussions with D. Rothery (Open University) and from
are certainly approaching a situation of near-real-time oper- formal reviews provided by M. Wooster (King’s College,
ation for the MODVOLC algorithm. London) and an anonymous referee. This is HIGP
contribution number 1216 and SOEST contribution number
5976.
6. Conclusions

We have outlined how MODIS data can be used as the References

basis for an automated system to detect and monitor
volcanic eruptions for the entire globe. Case studies from Barnes, W. L., Pagano, T. S., & Salomonson, V. V. (1998). Prelaunch
several volcanoes have shown the algorithm to be robust characteristics of the Moderate Resolution Imaging Spectroradiometer
(MODIS) on EOS-AM1. IEEE Transactions in Geoscience and Remote
enough to detect a range of contrasting eruption styles and Sensing, 36, 1088 – 1100.
intensities, while yielding very few, if any, false alerts. By Chester, D. K., Duncan, A. M., Guest, J. E., & Kilburn, C. R. J. (1985).
designing the algorithm to eliminate erroneous detection Mount Etna: the anatomy of a volcano (p. 404). Stanford: Stanford
events, we realise that we forgo the detection of low- University Press.
intensity hotspots, but we find this preferential to the Dehn, J., Dean, K., & Engle, K. (2000). Thermal monitoring of North
Pacific volcanoes from space. Geology, 28, 755 – 758.
situation in which the results of the algorithm cannot be Dozier, J. (1981). A method for satellite identification of surface temper-
trusted. The algorithm simultaneously provides a tool to ature fields of subpixel resolution. Remote Sensing of Environment, 11,
monitor the distribution of wildfires and industrial hot- 221 – 229.
spots. Flasse, S. P., & Ceccato, P. S. (1996). A contextual algorithm for AVHRR
Although this paper presents results pertaining to the fire detection. International Journal of Remote Sensing, 17, 419 – 424.
Flynn, L. P., Harris, A. J. L., & Wright, R. (2001). Improved identification
analysis of nighttime MODIS data, we are currently testing of volcanic features using Landsat 7 ETM + . Remote Sensing of Envi-
a version of the MODVOLC algorithm that will run on the ronment, 79, 1 – 14.
daytime MODIS data stream. When operational, this will Flynn, L. P., Mouginis-Mark, P. J., Gradie, J. C., & Lucey, P. G. (1993).
allow two hotspot observations per day for most areas of the Radiative temperature measurements at Kupaianaha lava lake, Kilauea
Volcano, Hawai’i. Journal of Geophysical Reasearch, 98, 6461 – 6476.
Earth. The launch of the second MODIS sensor on-board
Flynn, L. P., Mouginis-Mark, P. J., & Horton, K. A. (1994). Distribution of
Aqua will further enhance the utility of the MODVOLC thermal areas on an active lava flow field: landsat observations of
system, by allowing up to four hotspot observations to be Kilauea, Hawai’i, July 1991. Bulletin of Volcanology, 56, 284 – 296.
made (morning, afternoon, evening, and night) in a 24-h Flynn, L. P., Wright, R., Garbeil, H., Harris, A. J. L., & Pilger, E. (2002). A
period. This is important in terms of fire detection, as the global thermal alert using MODIS: initial results from 2000 – 2001.
morning and late-night overpass of Terra MODIS will miss Advances in Environmental Monitoring and Modeling, 1, 5 – 36.
Francis, P. W., & McAllister, R. (1986). Volcanology from space; using
the short-duration afternoon fires common in South America Landsat thematic mapper data in the Central Andes. Eos, Transactions
and Africa (Kaufman et al., 1998). It also has important of the American Geophysical Union, 67, 170 – 171.
implications for monitoring dynamic volcanic phenomenon, GVN (2001a). Soufriere Hills. Bulletin of the Global Volcanism Network,
such as advancing lava flows, by providing a more dense 26(2), 7 – 9.
temporal sample. GVN (2001b). Etna. Bulletin of the Global Volcanism Network, 26(5), 6 – 7.
GVN (2001c). Nyamuragira. Bulletin of the Global Volcanism Network,
The examples described in this paper show that MODIS 26(3), 7 – 8.
can be used to automatically detect and monitor spatially GVN (2001d). Nyamuragira. Bulletin of the Global Volcanism Network
and temporally dynamic thermal phenomenon. When fully 26(1), 7.
 R. Wright et al. / Remote Sensing of Environment 82 (2002) 135–155 155

GVN (2001e). Shiveluch. Bulletin of the Global Volcanism Network, 26(4), Mouginis-Mark, P. J., Snell, H., & Ellisor, R. (2000). GOES satellite and
2 – 4. field observations of the 1998 eruption of Volcan Cerro Azul, Galápa-
GVN (2001f). Piton de la Fournaise. Bulletin of the Global Volcanism gos Islands. Bulletin of Volcanology, 62, 188 – 198.
Network, 26(5), 4 – 6. Oppenheimer, C., Francis, P. W., Rothery, D. A., Carlton, R. W. T., &
Harris, A. J. L., Blake, S., Rothery, D. A., & Stevens, N. F. (1997). A Glaze, L. S. (1993). Infrared image analysis of volcanic thermal fea-
chronology of the 1991 to 1993 Etna eruption using advanced very high tures: Lascar Volcano, Chile, 1984 – 1992. Journal of Geophysical Re-
resolution radiometer data: implications for real-time thermal volcano search, 98, 4269 – 4286.
monitoring. Journal of Geophysical Research, 102, 7985 – 8003. Prins, E. M., & Menzel, W. P. (1992). Geostationary satellite detection of
Harris, A. J. L., Flynn, L. P., Dean, K., Pilger, E., Wooster, M., Okubo, C., biomass burning in South America. International Journal of Remote
Mouginis-Mark, P., Garbeil, H., Thornber, C., De La Cruz-Reyna, S., Sensing, 13, 2783 – 2799.
Rothery, D., & Wright, R. (2000). Real-time monitoring of volcanic hot Prins, E. M., & Menzel, W. P. (1994). Trends in South American biomass
spots with satellites. In P. J. Mouginis-Mark, J. A. Crisp, & J. H. Fink burning detected with the GOES visible infrared spin scan radiometer
(Eds.), Remote sensing of active volcanism. American Geophysical atmospheric sounder from 1983 to 1991. Journal of Geophysical Re-
Union Monograph: Vol. 116. (pp. 139 – 159). (Washington, DC). search, 99, 16719 – 16735.
Harris, A. J. L., Flynn, L. P., Rothery, D. A., Oppenheimer, C., & Sherman, Rothery, D. A., Coltelli, M., Pirie, D., Wooster, M. J., & Wright, R. (2001).
S. B. (1999). Mass flux measurements at active lava lakes: implications Documenting surface magmatic activity at Mount Etna using ATSR
for magma recycling. Journal of Geophysical Research, 104, 7117 – remote sensing. Bulletin of Volcanology, 63, 387 – 397.
7136. Rothery, D. A., Thorne, M. T., & Flynn, L. P. (in press). MODIS hot spot
Harris, A. J. L., Keszthelyi, L., Flynn, L. P., Mouginis-Mark, P. J., Thornb- alerts in Britain and the North Sea during the first half of 2001. Interna-
er, C., Kauahikaua, J., Sherrod, D., Trusdell, F., Sawyer, M. W., & tional Journal of Remote Sensing.
Flament, P. (1997). Chronology of the episode 54 eruption at Kilauea Schneider, D. J., Dean, K. G., Dehn, J., Miller, T. P., & Kirianov, V. U.
Volcano, Hawai’i, from GOES-9 satellite data. Geophysical Research (2000). Monitoring and analyses of volcanic activity using remote sens-
Letters, 24, 3281 – 3284. ing data at the Alaska Volcano Observatory: case study for Kamchatka,
Harris, A. J. L., Murray, J. B., Aries, S. E., Davies, M. A., Flynn, L. P., Russia, December 1997. In P. J. Mouginis-Mark, J. A. Crisp, & J. H.
Wooster, M. J., Wright, R., & Rothery, D. A. (2000). Effusion rate Fink (Eds.), Remote sensing of active volcanism. American Geophysical
trends at Etna and Krafla and their implication for eruptive mechanisms. Union Monograph: Vol. 116. (pp. 65 – 85). (Washington, DC).
Journal of Volcanology and Geothermal Research, 102, 237 – 270 Simkin, T., & Siebert, L. (1994). Volcanoes of the world (2nd ed.) (p. 368).
(Washington, DC) . Tucson, AZ: Geoscience Press.
Harris, A. J. L., & Neri, M. (in press). Volumetric observations of the 1999 Tazieff, H. (1994). Permanent lava lakes: observed facts and induced mech-
Bocca Nuova eruption: pressurized drainage of a shallow chamber or anisms. Journal of Volcanology and Geothermal Research, 63, 3 – 11.
pulsed supply? Journal of Volcanology and Geothermal Research. Wadge, G. (1981). The variation of magma discharge during basaltic erup-
Harris, A. J. L., Pilger, E., & Flynn, L. P. (2002). Web-based hot spot tions. Journal of Volcanology and Geothermal Research, 11, 139 – 168.
monitoring using GOES: what it is and how it works. Advances in Wooster, M. J. (2001). Long-term infrared surveillance of Lascar Volcano:
Environmental Monitoring and Modeling, 1, 37 – 69. contrasting activity cycles and cooling pyroclastics. Geophysical Re-
Harris, A. J. L., Pilger, E., Flynn, L. P., Garbeil, H., Mouginis-Mark, P. J., search Letters, 28, 847 – 850.
Kauahikaua, J., & Thornber, C. (2001). Automated, high temporal res- Wooster, M. J., & Kaneko, T. (1998). Satellite thermal analyses of lava
olution, thermal analysis of Kilauea volcano, Hawai’i, using GOES dome effusion rates at Unzen Volcano, Japan. Journal of Geophysical
satellite data. International Journal of Remote Sensing, 22, 945 – 967. Research, 103, 20935 – 20947.
Harris, A. J. L., & Stevenson, D. S. (1997). Thermal observations of open Wooster, M. J., Kaneko, T., Nakada, S., & Shimizu, H. (2000). Discrim-
degassing conduits and fumaroles at Stromboli and Vulcano using re- ination of lava dome activity styles using satellite-derived thermal struc-
motely sensed data. Journal of Volcanology and Geothermal Research, tures. Journal of Volcanology and Geothermal Research, 102, 97 – 118.
76, 175 – 198. Wright, R., de la Cruz-Reyna, S., Flynn, L. P., Harris, A. J. L., & Gomez-
Harris, A. J. L., Swabey, S. E. J., & Higgins, J. (1995). Automated thresh- Palacios, J. J. (in press). Infrared satellite monitoring at Popocatepétl:
olding of active lavas using AVHRR data. International Journal of explosions, exhalations, and cycles of dome growth. Journal of Geo-
Remote Sensing, 16, 3681 – 3686. physical Research.
Harris, A. J. L., & Thornber, C. R. (1999). Complex effusive events at Wright, R., Flynn, L. P., & Harris, A. J. L. (2001). Evolution of lava flow-
Kilauea as documented by the GOES satellite and remote video cam- fields at Mount Etna, 27 – 28 October 1999, observed by Landsat 7
eras. Bulletin of Volcanology, 61, 382 – 395. ETM + . Bulletin of Volcanology, 63, 1 – 7.
Harris, A. J. L., Wright, R., & Flynn, L. P. (1999). Remote monitoring of Wright, R., Rothery, D. A., Blake, S., & Pieri, D. C. (2000). Improved
Mount Erebus Volcano, Antarctica, using Polar orbiters: progress and remote sensing estimates of lava flow cooling: a case study of the
prospects. International Journal of Remote Sensing, 20, 3051 – 3071. 1991 – 1993 Mount Etna eruption. Journal of Geophysical Research,
Kaufman, Y. J., Justice, C. O., Flynn, L. P., Kendall, J. D., Prins, E. M., 105, 23681 – 23694.
Giglio, L., Ward, D. E., Menzel, W. P., & Setzer, A. W. (1998). Potential Young, S. R., Sparks, R. S. J., Aspinall, W. P., Lloyd, L. L., Miller, A. D.,
global fire monitoring from EOS-MODIS. Journal of Geophysical Re- Robertson, R. E. A., & Shepherd, J. B. (1998). Overview of the erup-
search, 103, 32215 – 32238. tion of Soufriere Hills Volcano, Montserrat, 18 July 1995 to December
Klett, T. R., Ahlbrandt, T. S., Schmoker, J. W., & Dolton, G. L. (1997). 1997. Geophysical Research Letters, 25, 3389 – 3392.
Ranking of the world’s oil and gas provinces by known petroleum
volumes. USGS Open-File Report 97-463.