RESEARCH ARTICLE

Time series analysis of infrared satellite data for detectingthermal anomalies: a hybrid approach

W. C. Koeppen & E. Pilger & R. Wright

Received: 3 March 2010 /Accepted: 5 November 2010 /Published online: 4 December 2010# Springer-Verlag 2010

Abstract We developed and tested an automated algorithmthat analyzes thermal infrared satellite time series data todetect and quantify the excess energy radiated from thermalanomalies such as active volcanoes. Our algorithm enhan-ces the previously developed MODVOLC approach, asimple point operation, by adding a more complex timeseries component based on the methods of the RobustSatellite Techniques (RST) algorithm. Using test sites atAnatahan and Kīlauea volcanoes, the hybrid time seriesapproach detected ~15% more thermal anomalies thanMODVOLC with very few, if any, known false detections.We also tested gas flares in the Cantarell oil field in theGulf of Mexico as an end-member scenario representingvery persistent thermal anomalies. At Cantarell, the hybridalgorithm showed only a slight improvement, but it dididentify flares that were undetected by MODVOLC. Weestimate that at least 80 MODIS images for each calendarmonth are required to create good reference imagesnecessary for the time series analysis of the hybridalgorithm. The improved performance of the new algorithmover MODVOLC will result in the detection of lowtemperature thermal anomalies that will be useful in

improving our ability to document Earth’s volcanic erup-tions, as well as detecting low temperature thermalprecursors to larger eruptions.

Keywords MODIS . Time series analysis .MODVOLC .

GOES . Kilauea volcano . Anatahan volcano . Cantarell oilfield

Introduction

Satellite-based thermal infrared (TIR) instruments havebeen a boon for monitoring the thermal behavior of theEarth’s surface. Multiple TIR instruments are currently inorbit offering near-global coverage of the Earth at afrequency of at least once per day. These instrumentsprovide data over potentially dangerous, high-temperaturephenomena, such as volcanic eruptions and forest fires,with relatively little cost and no risk to the end user.However, the data volume produced by spaceborne TIRinstruments is too large to be manually processed andanalyzed on a daily basis and global scale. Therefore, anautomated or semi-automated way to accurately detectthese thermal anomalies in TIR satellite data is an essentialcomponent to any global thermal monitoring system.

Automated algorithms have successfully detected ther-mal anomalies in TIR data from the Advanced Very HighResolution Radiometer (AVHRR) (e.g., Harris et al. 1995;Dehn et al. 2000; Webley et al. 2008), the Along-TrackScanning Radiometer (ATSR) (e.g., Wooster et al. 1997),Geostationary Operational Environmental Satellites(GOES) (e.g., Prins et al. 1998; Harris et al. 2001; Wrightet al. 2002a,b; Xu et al. 2010), the Moderate ResolutionSpectroradiometer (MODIS) (Wright et al. 2002a,b; Giglioet al. 2003), the Advanced Spaceborne Thermal Emission

Editorial responsibility: A. Harris

W. C. Koeppen (*)Earth and Ecological Science Institute,1 University Place,Orono, ME 04473, USAe-mail: william.koeppen@earthesi.org

E. Pilger :R. Wright (*)Hawaii Institute of Geophysics and Planetology,University of Hawaii,1680 East West Road, POST 602, Honolulu, HI 96822, USAe-mail: wright@higp.hawaii.edu

Bull Volcanol (2011) 73:577–593DOI 10.1007/s00445-010-0427-y

and Reflection Radiometer (ASTER)(e.g., Pieri andAbrams 2004; Trunk and Bernard 2008),and the SpinningEnhanced Visible and Infrared Imager (SEVIRI) (Robertsand Wooster 2008).Unlike manual approaches, the algo-rithm approach is quantitative from start to finish, and it canbe refined to reprocess the historical record of TIR datarelatively easily. Additionally, the same automated algo-rithm can be applied to multiple sensors in order tocompare results from instruments with differing spatialand temporal resolutions. Finally, TIR data that are used formonitoring hazards should be processed and sent on toappropriate agencies as soon as images become available.On a global scale, this type of turn around can only beachieved using an automated approach.

One example of an automated volcano monitoringsystem is the MODVOLC algorithm (Wright et al. 2002a,b). MODVOLC uses a combination of MODIS bands at 4and 12 μm in order to calculate a normalized thermal index(NTI) (Wright et al. 2002a,b, 2004). Wright et al. (2002a,b)found that NTI values could distinguish active lava fromland and atmospheric clouds more effectively than simplyusing the difference between the 4- and 12-μm bands, andthey empirically determined that NTI values greater than athreshold of −0.8 consistently identified active lava andminimized false alarms. The system began operation inFebruary 2000, and it has made quantitative informationregarding the level of thermal unrest at all of Earth’s activevolcanoes available, via the internet, typically within12 hours of each MODIS overpass.

Although the fixed threshold technique of MODVOLCtakes minimal computing power, is easy to understand, hasaphysical basis, and can be applied globally, it also limitsthe amount of information that can be gleaned from theMODIS dataset. When MODVOLC was developed, it waslimited to single pixel operations, eight mathematicalcalculations, and five MODIS bands—restrictions imposedby the computer resources then available at the GoddardSpace Flight Center processor where MODVOLC wasbeing run (Wright et al. 2002a,b). The single threshold thatMODVOLC applies to NTI values avoids false detectionsfrom the warmest and most variable surfaces on Earth (e.g.,where active lava flows enter the ocean in Hawai‘i).However, the same threshold may be unnecessarily conser-vative in geographic areas that have well defined seasonaltemperature variations, especially in regions that arerelatively cold and atmospherically clear (e.g., the MaunaLoa summit caldera).

Tramutoli (1998) proposed using long-term satelliteobservations of volcanoes in order to statistically charac-terize their thermal behavior and identify data that falloutside the normal distribution of values. This approachwas used by a number of authors to study AVHRR data(Tramutoli 1998; Cuomo et al. 2001; Pergola et al. 2001;

Tramutoli et al. 2001; Di Bello et al. 2004; Pergola et al.2004) and was christened Robust AVHRR Techniques(RAT), a name later changed to Robust Satellite Techniques(RST) to reflect its general applicability to other sensors(Pergola et al. 2008; Pergola et al. 2009). Pergola et al.(2004) proposed that RST could be applied globally inorder to detect volcanic thermal anomalies in TIR data.They presented a demonstrative analysis of eruptive eventsat Mt. Etna from July to August 2001 and January toFebruary 1999 as well as an analysis of both Mt. Etna andStromboli volcanoes from October to November 2002.

Because RST uses a long time series of TIR data, it hasthe potential to identify relatively low temperature thermalanomalies (Pergola et al. 2004) compared to the MOD-VOLC algorithm, which uses a conservative fixed thresh-old. This makes a time series approach particularlyappealing to the field of volcano monitoring becausesatellite thermal infrared data have recorded low tempera-ture (5–20 K) increases in various volcanic environmentssuch as on the flanks of volcanoes (Bonneville and Gouze1992; Di Bello et al. 2004; Pieri and Abrams 2005),fumarolic fields (Harris and Stevenson 1997; Lagios et al.2007), warm-mud volcanoes (Patrick et al. 2004), summitcraters(Dehn et al. 2002; Pieri and Abrams 2005), andsummit crater lakes (Trunk and Bernard 2008), some ofwhich have preceded larger eruptions. Non-supervisedidentification and tracking of these small and/or low-temperature features in satellite data require an algorithmthat is sensitive to low-temperature or spatially smallthermal anomalies.

Unfortunately, using only a time series approach isinadequate for identifying all types of thermal anomaliespresent in volcanic environments.The RSTalgorithm relies onthe detection of statistically anomalous behavior compared toa surface’s past thermal history (i.e., change detection), and itwas not developed to provide a continuous record ofpersistent, stationary thermal anomalies. For example, thePu‘u ‘Ō‘ō-Küpaianaha eruption has effused high-temperaturelavas onto Kīlauea’s southeast flank almost continuously since1983 (Heliker et al. 2003). In TIR satellite images over thisarea, it is “normal” to expect high thermal radiance valuesbecause many pixels have historically contained active lava,some over long durations. Satellite identification of allpresently occurring thermal anomalies, regardless of theirpast thermal history, is not only useful to monitoringagencies, it is essential when using flagged pixels to estimatethe excess radiant power (a proxy for flux rate) frompersistent thermal anomalies such as Kīlauea.

In this study we developed a hybrid time seriesalgorithm that uses the MODVOLC approach enhancedby a time series component. Our aim was to produce amethodology that could exploit the decadal MODIS dataarchive while requiring a bare minimum, if any, user

578 Bull Volcanol (2011) 73:577–593

interaction in the processing chain. In our implementation,data are first processed by MODVOLC and thermalanomalies are flagged. The data are then passed into a timeseries analysis, which identifies pixels that did not exceedthe NTI threshold but which are statistically anomalouscompared to the historical data set.

We tested the hybrid approach over three differentlocales: Anatahan volcano, Kīlauea volcano, and theCantarell oil field in the Gulf of Mexico. Thermalanomalies at these locations have markedly differentspatio-temporal characteristics, and they provided anobjective test of the method’s performance over a widerange of conditions. Thermal anomalies at Anatahanvolcano are spatially limited (one to two MODIS pixels),temporally sporadic, and of low spectral radiance. Those atKīlauea volcano are spatially extensive (several to tens of1-km MODIS pixels), temporally persistent, and exhibithigh levels of spectral radiance. The Cantarell site wasincluded as it represents the most extreme case with respectto temporal persistence. The gas flares are very small at theMODIS pixel scale, but their flaring is nearly continuouslyso that elevated radiance values over the flares are notstrictly anomalous, they are the norm.We determined thatthe hybrid time series algorithm is capable of monitoring alltypes of thermal anomalies, from sporadic to very persis-tent, and that it is more sensitive to low-temperature eventsthan MODVOLC, the currently applied global volcanomonitoring system.We investigated the hybrid algorithmparameter-space as well as an additional semi-automatedcorrection that is necessary for the algorithm to produce itsbest results over the most temporally persistent thermalanomalies. Finally, we compared the results obtained fromthe hybrid time series approach using MODIS data to thoseobtained from GOES data and show that it is portable tomultiple sensors with different spatial resolutions.

Methods

MODIS data

The first MODIS instrument was launched on board Terra,NASA’s first Earth Observing System (EOS) satellite inDecember 1999. The Terra satellite is in a polar, sun-synchronous orbit, and MODIS typically achieves nearlyglobal daytime and nighttime coverage of the Earth everyday and complete global coverage of the Earth every 2 days(Barnes et al. 1998). A second MODIS instrument was putinto orbit on board NASA’s Aqua platform in May 2002,and the Aqua data can be used in conjunction with the Terradata to achieve twice the temporal resolution of Terra alone(Xiong et al. 2008). For this study we only used data fromthe Terra MODIS instrument because development of the

algorithm did not require the full temporal resolution ofboth MODIS instruments. Additionally, we focused onnighttime MODIS data, which avoid complications due toreflected sunlight. For this work, this amounts to incorpo-rating one Terra MODIS image each night for ~ 95% of thedays between February 2000 and December 2008.

MODIS collects data in 36 wavebands of which 10 aresuitable for monitoring thermal behavior on the surface ofthe Earth. The hybrid algorithm uses MODIS bands 21, 22,and 32, all of which have spatial resolutions of ~1 km perpixel. MODIS bands 21 and 22 cover the same wavelengthregion, 3.930–3.989 μm; however, band 21 has a highersaturation temperature than band 22 (~500 and ~330 K,respectively) (Kaufman et al. 1998) at the expense of ahigher noise equivalent temperature (2.0 and 0.07 K,respectively) (Barnes et al. 1998). MODIS band 32 coverslonger wavelengths, 11.770–12.270 μm, and has a noiseequivalent temperature of 0.05 K and saturation tempera-ture of ~420 K (Barnes et al. 1998).

We retrieved all nighttime Terra MODIS radiance(MOD021KM) and geographic information (MOD03) filesfor Anatahan volcano, the Big Island of Hawai‘i, and theCantarell oil field in the Gulf of Mexico acquired betweenFebruary 2000 and December 2008. We then used bands 21,22, and 31 of each MODIS radiance image along with itscompanionMOD03 file to generate over-sampled (0.5 km perpixel), time-ordered, georeferenced data cubes that werecropped to our areas of interest. We primarily used band 22radiance values as the 4-μm inputs and band 32 radiancevalues as the 12-μm inputs. However, radiance values in band22 can saturate in pixels over very high temperature thermalanomalies (e.g., active volcanic eruptions and fires) andtruncate the measured radiance. Therefore, in saturated pixelswe used band 21 radiance values in order to better quantify thetrue value of the 4-μm radiance.

GOES data

GOES data are collected by a series of geostationarysatellites in five wavelength bands that include a mid-infrared band (band 2, 3.78–4.03 μm) and two far-infraredbands (band 4, 10.2–11.2 μm, and band 5, 11.5–12.5 μm).Although the spatial resolution of GOES data are relativelylow compared to MODIS (~4 km per pixel), images aretypically acquired every 15 min. GOES Imager data arereceived at the Naval Research Laboratory in Monterey,California where they are calibrated and converted tobrightness temperature. These data are then forwarded tothe Hawai‘i Institute of Geophysics and Planetology(HIGP) for processing, and they are converted back intoradiance space for comparison to the MODIS data.Thermalinfrared data from GOES are available over Kīlaueavolcano and the Cantarell oil field in Mexico; however,

Bull Volcanol (2011) 73:577–593 579

Anatahan volcano is outside the field of view of any GOESsatellite, and there are no corresponding data for thatlocation.

MODVOLC algorithm

MODVOLC was developed as a non-interactive algorithmto be applied globally to MODIS data in order to flagthermal anomalies, in particular volcanic eruptions (Wrightet al. 2002a,b, 2004). MODVOLC is described in detail inWright et al. (2002a,b), and we provide only a briefdescription here. MODVOLC uses a combination ofspectral wavelengths centered at 4 and 11–12 μm, whichcorrespond to the peak radiance emitted from high-temperature volcanic sources and ambient Earth surfaces,respectively. In a typical volcanic eruption, the hightemperatures associated with the event rarely fill the entirepixel. The radiance measured from the pixel is a combina-tion of the blackbody curves for the high- and ambient-temperature surfaces weighted in proportion to their surfacearea in the pixel. In pixels that contain a thermal anomaly,the radiance measured in the mid-infrared (~4 μm) isdramatically higher than in nearby pixels without a thermalanomaly, but the radiance measured in the far-infrared (11–12 μm) shows very little difference from nearby pixels(Wright et al. 2004). Many efforts to identify pixelscontaining active lava using remote sensing have exploitedthis fact, and they often create temperature “difference”images between the 4 and 12-μm bands to search forthermal anomalies (e.g., Harris et al. 1995; Higgins andHarris 1997; Wright et al. 2002a,b). However, this simpledifferencing technique breaks down when using just tworadiance bands to analyze images that contain a largedistribution of temperatures. For example, if only thedifference between the 4 and 12-μm measured radiancesis considered, a surface (or cloud) at −10°C is indistin-guishable from a surface at 13°C containing a 1000°Cvolcanic thermal anomaly that covers 0.03% of the surfacearea of the pixel (Wright et al. 2002a,b).

The MODVOLC algorithm takes one of many possibleapproaches to improve on the ambiguity of the simpledifferencing approach by normalizing the difference imagesby the sum of the 4- and 12-μm radiances to generate theNormalized Thermal Index (NTI) (Wright et al. 2002a,b).

NTI ¼ L4mm � L12mmL4mm þ L12mm

ð1Þ

The NTI effectively takes into account the absolute valueof the surface radiances, and MODVOLC is then able todistinguish between the surface at −10°C and the surface at13°C containing the 1000°C thermal anomaly (Wright et al.2002a,b). In most situations NTI is calculated using

MODIS band 22 as the 4-μm radiance value; however, ifband 22 is saturated, then band 21 is used. Because the 4-μm radiances are always lower than 12-μm radiances, NTIis always negative, and, when a high-temperature thermalanomaly is present in the pixel, NTI increases andapproaches zero. Wright et al. (2002a,b) empiricallyidentified a global NTI threshold of −0.80 as the minimumcutoff that suggests the presence of a high temperaturethermal event. MODVOLC calculates NTI, flags potentialthermal anomalies, and reports their positions on HIGP’sMODVOLC thermal alert website (http://modis.higp.hawaii.edu).

Times series analysis

The time series component of the hybrid algorithm issimilar to the RST algorithm in the way that it aims tostatistically characterize the thermal history of image dataand identify pixels that do not match the typical behavior.The time series analysis has five main steps: 1. Co-locatethe georeferenced satellite observations, 2. Collect theobservations into groups of data from the same time ofyear and similar viewing conditions, 3. Create referenceand variability images for each data group that encompassthe “normal” thermal behavior of each pixel in the region ofinterest, 4. Calculate the deviation of each original imagefrom the reference and variability images, and 5. Flagpixels that fall outside of the envelope of statisticallynormal thermal behavior (Tramutoli 1998; Pergola et al.2004). This analysis is run on a single wavelength bandthrough time, and, like NTI, this has the advantage ofavoiding the inter-band ambiguity between cold clouds andsurface pixels with heterogeneous temperatures found insimple differencing techniques (Pergola et al. 2004).

We used the time-ordered, georeferenced MODIS cubesto compute reference and variability images for the 4-μmMODIS data over Anatahan, the Big Island of Hawai‘i, andthe Cantarell oil field. MODIS images over our study siteswere acquired at roughly the same time of night (within a2 hour window) throughout the month. Data were groupedby calendar month to calculate one set of reference (mean)and variability (standard deviation) images for each monthusing all of the available nighttime MODIS images fromnearly 8 years of nightly observations (February 2000 toDecember 2008). For example, we calculated one mean andstandard deviation image for January using of all nighttimeMODIS data from January 2001, January 2002, January2003, etc. Figure 1a–c shows an example of the raw dataand resulting reference and variability images calculated forHawai‘i during the month of January.

For the final steps of the time series component we usedthe notation of the RST algorithm and calculated theAbsolutely Llocal [sic] Index of Change of the Environ-

580 Bull Volcanol (2011) 73:577–593

ment (ALICE, or ⊗V), which represents the degree towhich each pixel in an individual observation V(x,y,t)deviates from its normal behavior (VREF), normalized by itsnatural variability (σV) (Pergola et al. 2004):

�n x; y; tð Þ ¼ V x; y; tð Þ � VREF x; yð Þsn x; yð Þ ð2Þ

Pergola et al. 2004 point out that in areas containing highnatural variability, the resulting ⊗V value will be smaller,which, in effect, reduces the detection threshold within thealgorithm for that area. This is in contrast to fixed-thresholdalgorithms, which have the problem of needing to use aconservative, globally-applicable threshold value in order tosuppress false detections in areas of high natural variability.We then set a limit for ⊗V, above which radiances aredetermined to be thermally anomalous compared to theirnormal background values. Choosing a value for the ⊗V

limit corresponds to the number of standard deviationsaway from the mean that values must be to be consideredanomalous. For this work, we compared results from thehybrid time series algorithm using ALICE thresholds of2.0, 2.6, and 3.0 for MODIS data, which correspond to

statistical confidences of 95.4, 99.0, and 99.7% (respec-tively) that pixels not flagged by MODVOLC were stilloutside the normal distribution of radiances for thatlocation.

Preprocessing the time series inputs

The original RAT (now RST) scheme described byTramutoli (1998) applies three separate data-removal masksbefore calculating the reference and variability images fromthe time series. They include the following: (1) a discardingmask, which removes data outside of the region of interest,(2) an events mask, which removes outliers associated withthermal anomalies, and (3) a cloud mask, which removes alldata thought to be influenced by atmospheric water and iceclouds.

The hybrid time series approach tested here makes nouse of the discarding mask, because it is intended to be aglobally applicable algorithm. However it is necessary toremove large outliers associated with thermal anomalies(the events mask) present in the time series, because theycan greatly influence the resulting mean and standarddeviation images (Fig. 1b and c). These outliers include,

b)a)

c)

d) e)

Fig. 1 a Eight MODIS images from the first twelve days of January,2008 over the Big island of Hawai‘i showing the variability due tocloud cover in individual MODIS images. b Reference and c

variability images for January data with no corrections applied. dReference and e variability images for January data with MODVOLCpreprocessing applied

Bull Volcanol (2011) 73:577–593 581

for the most part, fires or volcanic eruptions, and the hybridalgorithm uses the results from MODVOLC as an effectivemeans of identifying and removing high-radiance outliersfrom the image cubes before calculating the reference andvariability images in its time series component (e.g., Fig. 1dand e). MODVOLC is a proven algorithm that identifiesmajor thermal anomalies with very few false detections(Wright et al. 2002a,b, 2004), and it has already been run inthe first portion of the algorithm so no additionalcalculations are necessary.

We tested the viability of two separate cloud removalmechanisms for use with the hybrid algorithm, the “k-σclipping filter” described by Tramutoli (1998) and thederived MODIS cloud mask (MOD35_L2). The k-σclipping filter iteratively removes data points from the timeseries of each pixel that fall above and below two and threestandard deviations (for regions of land and water,respectively) from the time series of each pixel. However,the filter alone cannot be used to remove cloud radiancevalues over geographic areas and/or times that containabundant cloud cover, and in all three of our test areascloud cover was an issue. For example, in one MODISpixel over the eastern crater of Anatahan, cloud radiancevalues account for >40% of the time series between themonths of July and October every year from 2000 to 2008.Similarly, full or partial cloud-influenced radiance valuestypically account for >50% of the radiance values measuredfrom the eastern quarter of the Big Island of Hawai‘i andthe western flank of Mauna Loa between the months ofDecember and February.

The MODIS cloud mask also proved inappropriate forfiltering the 4-μm time series.It has been acknowledged thatthe MODIS cloud-masking algorithms have difficultydiscerning clouds over non-vegetated surfaces, in high-elevation regions, and in nighttime data over areas withstrong surface temperature inversions (Platnick et al. 2003).Unfortunately, these situations are common when imagingmany of the Earth’s active or potentially active volcanoes.Two simple tests of the MODIS cloud mask over the BigIsland of Hawai‘i confirmed that the MODIS cloud maskcould not be used. In the first test we calculated thepercentage of clear sky data (MODIS cloud mask results of“clear” and “probably clear”) over the region for eachmonth (e.g., Fig. 2). The results show a distinct boundarybetween cloudy (dark shades) and clear (light shades) thatcorresponds exactly with the 2000 ft elevation contour.Elevation is a component of the MODIS cloud maskderivation (Platnick et al. 2003); however, a manualexamination of cloud patterns in all of the raw radiancedata do not show the distinct boundary that is inferred inthe MODIS cloud mask. Additionally, the summits of bothMauna Kea and Mauna Loa are shown in the MODIS cloudmask to be verging on >70% cloudy, a result that is

unrealistic as these locations are two of the most cloud-freeplaces on Earth. As a second test, we manually selected 34of the clearest 4-μm radiance images and inspected theirMODIS cloud mask results (e.g., Fig. 3). All of the selectedimages were virtually cloud free over their land surfacesand individual lava flows were clearly visible on the flanksof the volcanoes. However, the MODIS cloud mask showslarge portions of the Big Island as having “uncertain/probably cloudy” or “confident cloudy” values, which areerroneous.

Given the issues with the MODIS cloud mask, we used asimple and transparent approach wherein all of the MODISscenes (both clear and cloudy) are processed and included.Therefore, in the hybrid time series approach, the calculatedreference and variability images do not represent surface-only values. Rather, they define a statistical envelope thatincludes all of the natural thermal variation that a sceneelement undergoes, from surface temperature changes tofull or partial cloud cover. This envelope, encompassing allthe pixel’s natural variation, is used by the algorithm toidentify additional thermal anomalies that went undetectedby MODVOLC. Because we analyze a nine-year archive,we reduce the impact of inter-year variations in surfacetemperature and cloud cover.In fact, no approach using 1-km scale data can be truly considered to isolate surface-leaving radiance, regardless of the cloud screening methodemployed; sub-pixel-sized clouds or volcanic plumes arevirtually impossible to identify in such data. Compromisesin algorithm design are almost always necessary if one is toconvert a methodology applied semi-manually to onevolcano, to a methodologythat operates autonomously, in

100%0% 50%

Fig. 2 Percentage of values labeled as “clear” and “probably clear”according to the MODIS cloud mask (MOD35_L2) over the BigIsland of Hawai‘i for June, 2000 to 2008

582 Bull Volcanol (2011) 73:577–593

near-real-time, at the global scale. Our algorithm isdesigned with such an operational environment in mind.

12-μm correction for very persistent thermal anomalies

Very persistent thermal anomalies (e.g., volcanoes such asKilauea that erupt nearly continuously) pose a uniqueproblem for any time series analysis. In the time series of apixel over a persistent thermal anomaly, it is common toobserve many very high radiance values interspersed withcold, cloud radiance values. Preprocessing the data with

MODVOLC to remove the thermal events discards nearlyall data, and the pixel is left with a time series of only thecold, cloud values where the thermal anomaly wasobscured. Running a time series analysis on this pareddown time series has the potential to underestimate both themean and standard deviation values for the area comparedto areas adjacent to the thermal anomaly. To make mattersworse, this could lead to false detections over the locationof a persistent thermal anomaly (i.e., in a location where itis least likely to be recognized as false).

The hybrid time series approach applies a correction tothese pixels based on the fact that although sub-pixelthermal anomalies tend to show high 4-μm radiances, their12-μm radiances are generally representative of thebackground temperatures within the pixel (Wright et al.2002a,b). However, for most individual multi-band thermalinfrared images, plotting the 4-μm radiances against the 12-μm radiances produces a series of overlapping linear andsub-linear trends (Fig. 4). Empirically, the shape, slope, andy-intercept of eachtrend depend on the type of cover withinthe scene (e.g., different types/elevations of land, ocean, orvarious cloud types), presumably because of emissivitydifferences.If thermal anomalies are present in the image,they extend off the linear trends in 4-μm radiance, but fallin the normal range of 12-μm radiances. This effect is alsoseen in the reference and variability images that the hybridalgorithm calculates from the MODIS and GOES nighttimedata cubes.

confidentclear

probablyclear

uncertain/probablycloudy

confidentcloudy

Nov-10-2000, 9:10 GMT

Dec-26-2001, 8:45 GMT

Dec-13-2002, 8:40 GMT

Fig. 3 Representative MODIS radiance images (left) from very cleardays and corresponding MODIS cloud mask values (right) over theBig Island of Hawai‘i reported in the MOD35_L2 files

1.0

0.8

0.6

0.4

0.2

5.5 6.0 6.5 7.0

12-µm Radiance (Band 32)

4-µm

Rad

ianc

e (B

and

22/2

1)

Fig. 4 MODIS 12-μm radiance vs. MODIS 4-μm radiance for theuncorrected January reference image over Hawaii showing multipleoverlapping trends. Persistent thermal anomalies flagged by the hybridalgorithm (open circles) deviate from the normal distribution of points(grey circles) and are corrected by using the linear trend defined bypoints within the training region (black circles)

Bull Volcanol (2011) 73:577–593 583

One potential method to correct reference images wouldbe to use the 12-μm radiance measurements to calculate the4-μm values directly, a method that could be unsupervised(e.g., Wooster 2001; Webley et al. 2008). Unfortunately, thevariety of 12-/4-μm trends exhibited in Fig. 4 suggests thatthis method may misrepresent the 4-μm values up to 0.05radiance units (~15%), and it may not be appropriate tocalculate the 4-μm variability images from mostly 12-μmdata with those potential errors.

Therefore, the hybrid algorithm uses an empiricalscheme. It selects the pixels that are affected by persistentthermal anomalies by recording the pixel locations that hadmore than 5% of their data flagged and removed byMODVOLC. The algorithm uses a ~20 pixel squaretraining region in the images, which has similar surfacecover to the pixels being corrected (e.g., nearby land orocean with a similar distribution of elevations) but contain-ing no thermal anomalies. Although the algorithm deter-mines the pixels for which correction is necessary and alertsthe user, it does not, at present, automatically select thetraining region. It may be possible to automate this process,but there are many difficulties to algorithmically choosingtraining pixels. For example, over a subaerial islandvolcano it is necessary to select nearby land withoutselecting ocean pixels, which fall along a different 12-/4-μm trend. It is also appropriate to select a region withpixels that have a similar distribution of elevations to thosein the pixels being corrected in order to reproduce the 12-/4-μm trend that is obscured by the thermal anomaly.

Points within the training region are used to calculate theslope and y-intercept of the 4- to 12-μm trend line, and thealgorithm calculates the corrected 4-μm radiance values of thepixels by using the equation of the trend line and their 12-μmradiance values. Figure 5 shows an example of the raw Junereference images at Cantarell, the reference image withMODVOLC preprocessing but no applied 12-μm correction(where cloud values weight the resulting mean values), andthe reference image with the 12-μm correction applied.

Comparison of GOES and MODIS results

One of our aims was to show the portability of the hybridtime series approach to multiple sensors by comparingresults obtained from GOES and MODIS data. This isimportant because data from spacecraft operating in bothgeosynchronous (GEO) and low Earth orbits (LEO) can beused for monitoring thermal anomalies such as activevolcanism, with the usual trade-offs between frequency ofobservation and spatial resolving power. The spatially andtemporally dynamic nature of volcanic eruptions alsomeans that a multi-sensor/resolution approach provides theoptimal amount of information regarding the timing ofvolcanic eruptions (or changes in intensity during erup-

tions) as well as the spatial distribution of the eruptedproducts. To demonstrate under what circumstances thetemporal information available from GEO can be combinedwith the spatial information available from LEO, we havecompared the radiance detected by MODIS and GOES forKilauea and Cantarell study sites. In some circumstancesthe thermal anomalies will be too small (and/or cool) to bereliably detected in coarse resolution 4-km data; in thesecases GEO and LEO do not serve to complement eachother. However, for targets that are sufficiently radiant to beobservable in GOES-class data, the high temporal resolu-tion record of emittance can be used synergistically with thehigh spatial resolution of MODIS-class data.

In order to directly compare results generated by thehybrid approach from GOES data to results obtained fromMODIS data, we applied the algorithm to MODIS andGOES data cubes for two of our three target areas. TheGOES image acquired closest in time to each nighttimeTerra MODIS acquisition was selected, and thus the imageswere coincident in time to within 30 min. This image stackwas then processed by the hybrid algorithm in exactly thesame way as the MODIS data were processed. That is, byrunning the MODVOLC algorithm on the data, makingmonthly reference and variability images using the datapreprocessed by MODVOLC, comparing the original

a)

b)

c)

Fig. 5 Reference images forJune data over the Cantarell oilfield showing a raw referenceimage (no correction applied), breference with MODVOLC pre-processing, c reference imagewith MODVOLC preprocessingand 12-μm correction

584 Bull Volcanol (2011) 73:577–593

image stack against the reference and variability images,and adding the thermally anomalous pixels detected by thetime series analysis to those identified by MODVOLC. Theonly variable in this analysis was spatial resolution, giventhat scene subject and spectral band were held constant.

We compared the results from MODIS to those ofGOES, which has a much larger pixel size, by calculatingthe total excess power emitted by the thermal anomalies,which is a sensor-independent measure of its radiant flux.We determined the total excess power by calculating the4-μm radiance for each flagged pixel, subtracting thereference value (i.e., the mean 4-μm radiance) for thatlocation and time, summing the pixels over the features ofinterest, and multiplying by the pixel size of theinstrument.

Results

We compared the results of the hybrid time series approachto the MODVOLC algorithm using nighttime MODIS dataover Anatahan volcano, Kīlauea volcano, and the Cantarelloil field in the Gulf of Mexico. Primarily, we used thefollowing parameters to analyze the timing and distributionof detected thermal anomalies: a) the total number of pixelsflagged in each image, which is a proxy for the total areaand duration of the thermal anomaly, and b) a two-dimensional histogram that shows the spatial distributionof pixels (summed by month)flagged by the differentalgorithms, which can be used to determine if the algorithmproduces false detections in areas known to have nothermal activity. We then applied the hybrid time seriesap-proach to GOES data over Kīlauea and Cantarell andcompared the results to those obtained from MODIS.

Anatahan volcano

Anatahan is a volcanic island in the Marianas Island Arc,located at 16.35N, 145.68E. Anatahan is a particularlyuseful test of thermal anomaly detection algorithms becausethe volcano was inactive until it erupted suddenly on May10, 2003, and since that time it has displayed relativelysporadic eruptions (Trusdell personal communication;Trusdell et al. 2005). This provides us with over 2 yearsof known volcanic quiescence for Anatahan (February 2000to April 2003) and a clear beginning to an extended thermalanomaly that can be used to identify any false detectionsmade by the algorithm. Anatahan also exhibits the leastpersistent eruption chronology that we studied, and there-fore it had the most potential for improvement byincorporating a time series analysis.

We calculated the total number of thermally anomalouspixels identified by the MODVOLC algorithm overAnatahan and summarized them by month (Fig. 6). Thepeaks in the number of thermal anomalies detected byMODVOLC correspond very well to periods of knownvolcanic activity. MODVOLC identified no thermal anoma-lies at Anatahan prior to May 2003, consistent with a 0%rate of false detections during the period of knownquiescence. The first major eruption at Anatahan occurredin May 2003, an event well documented by MODVOLC(Wright et al. 2005) and the hybrid time series approach.Thermal anomalies recorded from May to July 2004 andJanuary to July 2005 correspond to moderate to highStrombolian volcanic events that occurred at Anatahan(Trusdell personal communication).

In order to compare the results from the hybrid algorithmalong with the multiple ALICE thresholds we tested, weused the MODVOLC results over Anatahan as a baseline.

2000 2001 2002 2003 2004 2005 2006 2007 20080

100

50

150

Δ A

nom

alie

s (p

ixel

s)

0

200

100

300

# A

nom

alie

s (p

ixel

s)

Year

ALICE 2.0ALICE 2.6ALICE 3.0

Anatahan Volcanoa)

b)

Fig. 6 a Total number of ther-mal anomalies detected by theMODVOLC algorithm overAnatahan volcano, summarizedby month. b Increase in numberof thermal anomalies detected bythe hybrid algorithm over MOD-VOLC (note changed y-axis)using ALICE thresholds of 2.0,2.6, and 3.0. Lightly shaded areaindicates period of volcanic qui-escence at Anatahan, and blackarrows indicate false detectionsby the hybrid algorithm usingALICE threshold of 2.0

Bull Volcanol (2011) 73:577–593 585

We calculated the total number of thermal anomaliesdetected by the hybrid algorithm, and we subtracted thenumber of thermal anomalies detected by MODVOLC(Fig. 6). When using an ALICE threshold of 2.0, the hybridalgorithm flagged a small number (~5 total) of thermalanomalies at Anatahan in the period prior to May 2003. Wemanually inspected the individual images that containedthese pixels, but no anomalous pixels were found indicatingthat those identifications were false detections.The hybridalgorithm did not flag any pixels during this volcanicallyquiescent period when the ALICE threshold was set to 2.6or 3.0.

As expected at Anatahan, the hybrid algorithm detectedmany more thermal anomalies than MODVOLC. Thisincrease is most apparent in the total number of thermalanomalies detected in May 2003 and January to April 2005,periods with significant volcanic activity on the island(Trusdell et al. 2005). The hybrid algorithm also identifieda number of thermal anomalies correlated with Anatahan’scontinuing activity in 2006, 2007, and 2008 that were notdetected by MODVOLC (Fig. 6). These anomalies primar-ily correspond to ash and steam plumes coming fromAnatahan, which were also detected by other satellite andground observations during these time periods (Trusdellpersonal communication).

Two-dimensional histograms of the thermal anomaliesidentified by the hybrid algorithmshow that almost all ofthe flagged pixels occur over the island of Anatahan(Fig. 7). In particular, thermal anomalies occur primarilyover the eastern crater of Anatahan where the bulk of theknown volcanic activity has occurred (Trusdell et al. 2005).The spatial distribution of thermal anomalies in thehistogram is evidence that the algorithm is not identifyingfalse detections in the form of random noise throughout thescene. In addition, we confirmed the good agreementbetween thermal anomalies present in the MODIS radiance

data to the pixels flagged by the hybrid time seriesalgorithm through manual inspection of the time series.

Kīlauea volcano

Kīlauea, on the Big Island of Hawai‘i at 19.40N, −155.27E,erupts much more frequently than Anatahan. The effusivelavas that emanate from Pu‘u ‘Ō‘ō typically move down theslopes of the volcano to enter the ocean, and excessradiance is commonly observed by MODIS during allstages of the lava’s aboveground movement. As withAnatahan, we summarized the total number of thermalanomalies identified per month over Kīlauea by MOD-VOLC and the hybrid algorithm (Fig. 8). The scale onFig. 8 shows the increased number of anomalies identifiedat Kīlauea over Anatahan, and every month betweenFebruary 2000 and December 2008 displayed some pixelsflagged.

At Kīlauea, there are significant differences betweenresults obtained using the hybrid algorithm with differentALICE thresholds (Fig. 8b). Using an ALICE threshold of2.0, the algorithmflagged many more thermal anomaliesthan MODVOLC, but large spikes were seen in the resultsobtained from the summer months of every year of thestudy. In the two-dimensional histograms for results usingan ALICE threshold of 2.0 (not shown) it was clear that thetime series component of the hybrid algorithm flaggedpixels unrealistically across the entire Big Island. Evenusing an ALICE threshold of 2.6 caused the hybridalgorithm to produce systematic false detections on theflanks of Mauna Kea and Mauna Loa that were provenerroneous by manual inspection of the data. In order toachieve zero known false detections, we were required touse an ALICE threshold of 3.0 for the hybrid algorithmover the Big Island. Figure 9 shows that the hybridalgorithm detected more thermal anomalies on the BigIsland of Hawai‘i than those produced solely in the vicinityof Kīlauea. However, we manually inspected and verifiedthe presence of these thermal anomalies in individualMODIS images, and these occurrences likely representvegetation fires associated with agriculture.

Although it goes beyond the scope of this paper toanalyze correlations between the thermal radiance and theoccurrence of all of the individual volcanic eventsatKīlauea, fluctuations in the time series do reflect realchanges at the volcano based on geologic observationsfrom the USGS Hawai‘i Volcano Observatory (Orr andPatrick personal communication). For example, the highexcess radiances calculated for January to May 2005correspond to a significant increase in flow field activityand surface breakouts associated with the PKK flows atKīlauea. Similarly, high radiance values from August toNovember 2007 correspond to a fissure eruption that fed

1 1005025 75

Number of Anomalies

25 km

Fig. 7 Two-dimensional histogram of thermal anomalies identifiedover Anatahan volcano by the hybrid time series algorithm using anALICE threshold of 3.0

586 Bull Volcanol (2011) 73:577–593

multiple perched lava ponds east of Pu‘u ‘Ō‘ō at that time.Ongoing work is making use of the hybrid algorithm resultsand correlating them to specific volcanic episodes acrossmultiple time scales.

Cantarell oil field

Cantarell is the largest oil field in Mexico, and the burningof associated gas at Cantarell produces flares that are almostalways visible in cloud-free MODIS data (Wright et al.2002a,b). Although the primary use of the hybrid time

series algorithm would be for detecting volcanic activity,Cantarell represents an end-member scenario that tests thealgorithm over very persistent thermal anomalies. Individ-ual MODIS images show that at least 17 established flaregroups can be distinguished at Cantarell, and the number ofthermal anomalies visible to MODIS is indicative both ofthe volume of gas being flared and the number of wells thatare flaring at any given time. The number of thermalanomalies identified and summarized per month byMODVOLC over the entire field is consistently high(Fig. 10). Broad variations in MODVOLC results showthat flaring at Cantarell decreased in 2003 and stayedrelatively low until 2006 when flaring increased to itshighest levels in the 8 years of our study.

The hybrid time series approach detected slightly morethermal anomalies than MODVOLC over Cantarell, thoughthe results from the hybrid algorithm also show a clearseasonal dependency (Fig. 10). During our study period, thetime series portion of the hybrid algorithm identified morepixels than MODVOLC during the winter and spring(November to May). During these months, variabilityimages over the Gulf of Mexico show relatively lowvalues, a scenario that gives the time series component ofthe hybrid algorithm an enhanced ability to detect thermalanomalies. However, during the summer and fall (June toOctober) the time series portion of the hybrid algorithmprovided no benefit because the envelope of normal thermalbehavior is relatively large due to higher natural variability.

For the most part, two-dimensional histograms of thehybrid algorithm results (Fig. 11) look very similar to thosecreated for MODVOLC (not shown). However, some of thelow frequency, spatially dispersed thermal anomalies seenoutside of the V-shape formation of the main wells wereonly flagged by the hybrid algorithm. We manually

1 >25612864 192Number of Anomalies

50 km

Fig. 9 Two-dimensional histogram of thermal anomalies identifiedover the Big Island of Hawai‘i by the hybrid time series algorithmusing an ALICE threshold of 3.0. Arrows indicate thermal anomaliesdetected by the algorithm that were undetected by MODVOLC

Δ A

nom

alie

s (p

ixel

s)

2000 2001 2002 2003 2004 2005 2006 2007 2008

Year

3000

1000

0

2000

# A

nom

alie

s (p

ixel

s)

0

2000

1000

3000

ALICE 2.0ALICE 2.6ALICE 3.0

K lauea Volcanoa)

b)

Fig. 8 a Total number of ther-mal anomalies detected by theMODVOLC algorithm overKīlauea Volcano, summarizedby month. b Increase in numberof thermal anomalies detectedby the hybrid algorithm overMODVOLC using ALICEthresholds of 2.0, 2.6, and 3.0

Bull Volcanol (2011) 73:577–593 587

inspected the MODIS time series to confirm the presence ofthese high-radiance pixels in the data, and they mayrepresent small exploration wells in the area.

Comparison of MODIS and GOES results

As a test of the hybrid algorithm on data sets with differentspatial resolutions, we compared the excess energy calcu-lated using data from the MODIS and GOES instruments.Although Anatahan volcano is outside of any GOES

satellites’ fields of view, we were able to compare resultsover Kīlauea volcano and the Cantarell oil field.

At Kīlauea the mean excess power per month calculatedfrom GOES is relatively similar to that calculated fromMODIS (Fig. 12). There were gaps in the GOES dataset (e.g., April to September 2002, and February to April 2007)and some bad GOES data had to be removed (e.g.,September to December 2001), but most of the 2000–2008 observations by MODIS had nearly concurrentobservations by GOES. Relative month-to-month changesin excess radiance calculated for MODIS and GOES arealso usually in agreement. Plotted against each other, themean excess power per month derived from MODIS andGOES data at Kīlauea show a relatively simple trend(Fig. 13). This trend is below the one-to-one line, andappears to fall off at high MODIS powers indicating thathigh radiances measured by MODIS generally show adecreasing return in GOES data. This may be because theGOES instrument saturates at high radiance values,whereas the high saturation values achieved by combiningMODIS bands 21 and band 22 allow MODIS to record thepeak radiance even at very hot locations. Therefore, as theGOES pixels over an thermal anomaly saturate, the meanexcess power calculated for GOES asymptotes towards avalue dictated by the spatial extent of the anomaly ratherthan its actual excess thermal energy.

At Cantarell the mean excess power per month calculat-ed from GOES is clearly not a good match to the excesspower calculated from MODIS, and even gross trends inthe MODIS data are not visible in the GOES data (Fig. 14).GOES data were only available from October 2002onwards, and most of the available data show that GOESseverely underestimated the excess power calculated fromMODIS. However, there are also large spikes in the excess

0

1500

1000

2000

500

# A

nom

alie

s (p

ixel

s)

2000 2001 2002 2003 2004 2005 2006 2007 2008

Year

500

0

1000

Cantarell Oil Field

ALICE 2.0ALICE 2.6ALICE 3.0

Δ A

nom

alie

s (p

ixel

s)

a)

b)

Fig. 10 a Total number of ther-mal anomalies detected by theoriginal MODVOLC algorithmover the Cantarell Oil Field,binned by month. b Increase innumber of thermal anomaliesdetected by the hybrid algorithmover MODVOLC (note changedy-axis) using ALICE thresholdsof 2.0, 2.6, and 3.0

1 >25612864 192Number of Anomalies

50 km

Fig. 11 Two-dimensional histogram of thermal anomalies identifiedover the Cantarell oil field by the hybrid time series algorithm usingan ALICE threshold of 3.0. Arrows indicate thermal anomaliesdetected by the algorithm that were undetected by MODVOLC

588 Bull Volcanol (2011) 73:577–593

power calculated from GOES data that overestimated theexcess radiance observed in MODIS (e.g., November2006). The cause of the lack of correlation at Cantarell isnot immediately apparent, but it may be a combination ofthe small size of the gas flares at Cantarell relative to theGOES pixel size (4 km) and the presence of many badpixels in the GOES images over this region. The excessradiance spikes do not appear to be seasonal, as might bethe case if the hybrid time series algorithm had incorrectlymade reference and/or variability images.

The number of images required bythe hybrid time seriesapproach

The time series component of the hybrid algorithm requiresthat we characterize the thermal behavior of a surfacebefore comparing individual images to the reference.However, it is not straightforward to estimate the minimumamount of data that are required to create the necessaryreference and variability images. Obviously, using moredata to create the references can lead to a better character-ization of the surface, but how long must we collect images

before a time series analysis can be employed on a new dataset?

To estimate the number of MODIS images necessary forrelatively good results, we used a Monte Carlo approach toanalyze MODIS data over Anatahan for the months ofJanuary and July. We created 10,000 variability imagesover Anatahan using a randomized assortment of anincreasing number of MODIS images, from 2 to the totalnumber of images available in the given month. Forexample, in January we created 100 different variabilityimages by using 2 random MODIS images from theJanuary data, 100 variability images using 3 randomMODIS images from January, etc. until all of the availableJanuary data were used to make the variability images (149images). The average value of each variability image wasplotted (14,900 values in January, 16,400 values in July) toshow the decreasing envelope of standard deviations as weapproached using all of the available data (Fig. 15).

Unfortunately, randomizing the data is not necessarilyrealistic. For example, in reality radiances are typicallygrouped by year, and it is quite possible that a particularmonth of one year is significantly more or less radiant thanall of the same months in multiple previous years (e.g.,“unseasonably warm weather”). Nevertheless, Fig. 15suggests that by the time ~80 MODIS frames are used tomake the reference and variability image, the envelope hasstabilized enough to generate reliable statistics. Presumably,this number is most likely to be valid if the 80 images comefrom multiple years of observations.

The frequency at which GOES data are obtainedprovided another opportunity to test using additionalimages to make the reference and variability image requiredby the hybrid time series approach. We calculated anotherseries of variability images using GOES data found within30-, 60-, 120-, and 240-minute windows of the MODIS

Mean Excess Power per Month

0

10

20

30

40

50

2000 2001 2002 2003 2004 2005 2006 2007 2008

Month

MW

µm

-1 s

r-1

MODIS

GOES

Fig. 12 Mean excess power permonth (MWμm-1 sr-1)calculated for Kīlauea byMODIS (black) and GOES (red)

0

10

20

30

0 10 20 30 40 50

MODIS MW m-1 sr-1

GO

ES

MW

m-1 s

r-1

y = -0.008x2+ 0.8x

R2= 0.9

Fig. 13 MODIS vs. GOES mean excess power per month at Kīlauea

Bull Volcanol (2011) 73:577–593 589

acquisition times. This provided us with variability imagesthat incorporated between ~100 and 3500 images for asingle month. However, we found that increasing thenumber of GOES images used to make the variabilityimages had virtually no effect on decreasing the overallstandard deviations present in the variability images. Thiswas not wholly unexpected because increasing the timewindow of GOES data collection also increases thevariability present in the data set. However, it didcorroborate our estimation that there is little to gain fromusing more than ~80 images per month in each set ofreference and variability images.

Discussion

MODVOLC analyzes single images on a pixel-by-pixelbasis and uses a simple threshold to flag index values basedon two (or three, in the case of saturated pixels) MODIS

bands at 4 and 12 μm. MODVOLC is fast, requires limitedcomputer resources and storage capacities, and has beenshown to be effective in monitoring Anatahan, Bezy-mianny, Colima, Erebus, Etna, Karymsky, Merapi, Nyir-agongo, Popocatépetl, Soufriere Hills, and many othervolcanoes worldwide (e.g., Flynn et al. 2002; Rothery etal. 2003; Wright et al. 2004, 2005). However, becauseMODVOLC uses a single threshold to analyze MODISdata regardless of where they were collected, the thresholdhad to be set high enough to accommodate a global rangeof situations without incurring false detections. Computerresources are not the constraint that they were when theMODVOLC system for automated volcano monitoringusing MODIS was initially proposed (Mouginis-Mark etal. 1991), and a more sophisticated algorithm can now beglobally employed.

Tramutoli (1998) proposed that a time series approachcould be used to analyze TIR satellite data, and Pergola etal. (2004) suggested that a time series analysis had the

Mean Excess Power per Month

2000 2001 2002 2003 2004 2005 2006 2007 2008

Month

MW

µm

-1 s

r-1

MODIS

GOES

0

5

10

15

20

25

Fig. 14 Mean excess power permonth (MWμm-1 sr-1) calculat-ed for Cantarell by MODIS(black) and GOES (red)

0 15010050Number of MODIS Frames

0.1

0.2

0.3

Sta

ndar

d D

evia

tion July

January

Fig. 15 Average standard devi-ation for the Anatahan studyarea as a function of the numberof MODIS frames used to makethe variability image. MODISimages from January (dark grey)and July (light grey) were col-lected, and random images ofthe monthly data set were usedto make variability images. Theplot shows that at least ~20images (from multiple years) arenecessary to avoid the worststandard deviations, and by ~80images the envelope is stable.Black lines represent the meanvalue of all iterations for a givennumber of MODIS frames

590 Bull Volcanol (2011) 73:577–593

potential to identify much lower temperature thermalanomalies than MODVOLC. Their approach (embodiedby the RST algorithm) aims to characterize the typicalradiant behavior of image pixels and identify statisticallyanomalous behavior, i.e., change detection. To that end,RST has been used to successfully identify the location andextent of both lava flows and volcanic ash clouds on Mt.Etna and Stromboli (Di Bello et al. 2004; Pergola et al.2004; Pergola et al. 2008; Pergola et al. 2009).

The hybrid time series approach we describe first uses afixed NTI threshold to identify major thermal anomaliessimilar to MODVOLC, and then it applies a time seriesanalysis similar to RST to identify additional thermalanomalies. This is different than RST in that our purposeis not only to monitor changes but also to monitor locationsthat show consistently high thermal signatures. At manyvolcanoes, it is useful to monitoring agencies if a pixel isflagged every time there is a lava flow present, not onlywhen the lava flow shows a change in thermal activity.

When calculating the reference and variability imagesnecessary for the analysis, the time series component of thehybrid algorithm also takes a slightly different approach fromRST as documented by Tramutoli (1998) and Pergola et al.(2004). The RST methodology employs both an events maskand a cloud mask to attain reference and variability imagesthat statistically characterize the surface-only radiance valuesfor a location. The hybrid approach’s use of the NTIthreshold (i.e., the MODVOLC algorithm) prior to the timeseries analysis is analogous to the events mask of RST.However, we found no reliable way to incorporate a MODIScloud mask into the hybrid approach. Tramutoli (1998)describes the k-σ clipping filter as one method of removingclouds from the time series; however, this method is onlyapplicable in cases where cloudy values are not a significantportion of the data, which was a problem at all of our testlocales. We also found the derived MODIS cloud maskunreliable for 4 μm nighttime MODIS data over Hawai‘i,and it may be suspect over any topographic feature that hashigh elevations and a lack of vegetation, which arecharacteristics of many active volcanoes. Instead, the hybridalgorithm uses all of the data to build a time series containingall of the natural variability, surface and clouds, over a givenlocation. This necessarily reduces the sensitivity of thealgorithm to surface thermal anomalies because the standarddeviations for each cloud-containing location are higher.However we show here that including cloud values in thetime series does not result in false detections, and stillprovides an improvement in detection of thermal anomaliesover MODVOLC.

This study shows that an ALICE threshold of 2.0 is toolow to be reliably applied globally using the hybrid timeseries algorithm, and an ALICE threshold of 2.6 stillproduced a small number of false detections in our study

over the Big Island of Hawai‘i. An ALICE threshold of 3.0was necessary to avoid any known false detections in allthree of our test locales.

Comparing the hybrid algorithm to the existing MOD-VOLC algorithm, we determined an average improvementof ~15% more thermal anomalies detected at Anatahan andKīlauea volcanoes, with very few (if any) false detections.Most of the newly identified pixels correspond to knownvolcanic activity, though some, which were manuallyinspected and validated in individual MODIS images, areassumed to be fires or other transient events. Although a15% improvement may seem small, it expands thecapabilities of our global monitoring system to specificallyincorporate lower temperature thermal events than found byMODVOLC alone, and it could be useful in identifyinglower temperature thermal precursors to larger eruptions(Di Bello et al. 2004; Pergola et al. 2004).

The hybrid time series algorithm could be globallyapplied without any user interaction, and we havedemonstrated an improved ability over MODVOLC toflag thermal anomalies for three test sites with markedlydifferent thermal emission characteristics. Accurate mon-itoring of the most stable and persistent anomalies,present at Kīlauea and represented in the extreme bythe Cantarell gas flares, requires correction of thereference and variability images prior to running thetime series component of the hybrid algorithm. However,this correction is easily applied via the method weoutline in this work and unnecessary in order to obtaingood results from less persistent anomalies includingfires and most volcanoes. Furthermore, any globalreference and variability images created from MODISdata need only to be corrected once for the algorithm toachieve optimal results for all events worldwide.

Conclusions

We developed and tested a hybrid time series algorithm,designed to be employed as an autonomous, near-real-timeglobal thermal monitoring system. The hybrid algorithmenhances the capabilities of the current thermal detectionalgorithm, MODVOLC, by added a time series analysis. Thehybrid algorithm uses the MODVOLC methodology toidentify thermal anomalies, as well as preprocess the timeseries data to remove high-temperature outliers prior thecalculation of a statistical envelope of “normal” values for agiven surface. Then, a time series analysis is performedwherein pixels that fall outside the statistical envelope areflagged and added to those identified by MODVOLC. Weshow that the hybrid approach detects ~15% more thermalanomalies than MODVOLC, the currently applied thermalmonitoring system, at Anatahan and Kīlauea volcanoes.

Bull Volcanol (2011) 73:577–593 591

Though the hybrid algorithm will work in a completelyautomated fashion for most applications, the monitoring ofsome very persistent thermal events (e.g., Kīlauea volcano)would require an additional, semi-automated 12-μm correc-tion that must be applied in order to correctly estimate both thenumber of events in the area and the total excess radiancebeing emitted by them.

Although ideally all of the available data should be includedto make the best possible reference and variability images forthe hybrid algorithm, we estimated that at least 80 images permonth, preferably from multiple years, were necessary togenerate good statistics from which to run the algorithm.

We found that the hybrid time series approach achievedrelatively good results on data from multiple sensors(MODIS and GOES). This provides confidence that thisalgorithm could be run on instruments flown in the future,regardless of their exact spatial and temporal resolutions.

Acknowledgements This work was funded by NASA grantNNX08AF08G and the MASINT Consortium to RW. The authorsthank Harold Garbeil for writing the georeferencing software used inthis work. The manuscript was improved by thoughtful reviews fromMatthew Patrick, Martin Wooster, and Andrew Harris. This is HIGPpublication 1876 and SOEST publication 8057.

References

BarnesWL, Pagano TS, Salomonson VV (1998) Prelaunch characteristicsof theModerate Resolution Imaging Spectroradiometer (MODIS) onEOS-AM1. IEEE Trans Geosci Remote Sens 36(4):1088–1100

Bonneville A, Gouze P (1992) Thermal survey of Mount Etna volcanofrom space. Geophys Res Lett 19(7):725–728

Cuomo V, Lasaponara R, Tramutoli V (2001) Evaluation of a newsatellite-based method for forest fire detection. Int J Remote Sens22(9):1799–1826

Dehn J, Dean K, Engle K (2000) Thermal monitoring of North Pacificvolcanoes from space. Geology 28(8):755

Dehn J, Dean KG, Engle K, Izbekov P (2002) Thermal precursors insatellite images of the 1999 eruption of Shishaldin Volcano. BullVolcanol 64(8):525–534

Di Bello G, Filizzola C, Lacava T, Marchese F, Pergola N,Pietrapertosa C, Piscitelli S, Scaffidi I, Tramutoli V (2004)Robust satellite techniques for volcanic and seismic hazardsmonitoring. Ann Geophys 47(1):49–64

Flynn L, Wright R, Garbeil H, Harris A, Pilger E (2002) A globalthermal alert system using MODIS: initial results from 2000–2001. Adv Environ Monit Model 1(1):37–69

Giglio L, Descloitres J, Justice CO, Kaufman YJ (2003) An enhancedcontextual fire detection algorithm for MODIS. Remote SensEnviron 87(2–3):273–282

Harris AJL, Stevenson DS (1997) Thermal observations of degassingopen conduits and fumaroles at Stromboli and Vulcano usingremotely sensed data. J Volcanol Geotherm Res 76:175–198

Harris AJL, Vaughan RA, Rothery DA (1995) Volcano detection andmonitoring using AVHRR data: the Krafla eruption, 1984. Int JRemote Sens 16:1001–1020

Harris AJL, Pilger E, Flynn LP, Garbeil H, Mouginis-Mark PJ,Kauahikaua J, Thornber C (2001) Automated, high temporal

resolution thermal analysis of Kilauea volcano, Hawai‘i, usingGOES satellite data. Int J Remote Sens 22(6):945–967

Heliker C, Swanson DA, Takahashi TJ (2003) The Pu‘u ‘Ö‘ö-Küpaianaha eruption of Kïlauea volcano, Hawai‘i: the first20 Years. USGS Prof Paper 1676

Higgins J, Harris AJL (1997) VAST: a program to locate and analysevolcanic thermal anomalies automatically from remotely senseddata. Comput Geosci 23:627–645

KaufmanYJ, Justice CO, Flynn LP, Kendall JD, Prins EM, Giglio L,WardDE, Menzel WP, Setzer AW (1998) Potential global fire monitoringfrom EOS-MODIS. J Geophys Res 103(D24):32215–32238

Lagios E, Vassilopoulou S, Sakkas V, Dietrich V, Damiata BN, GanasA (2007) Testing satellite and ground thermal imaging of low-temperature fumarolic fields: the dormant Nisyros Volcano(Greece). ISPRS J Photogramm Remote Sens 62(6):447–460

Mouginis-Mark PJ, Francis PW, Friedman T, Garbeil H, Gradie J, SelfS, Wilson L, Crisp JA, Glaze L, Jones K, Kahle AB, Pieri DC,Zebker H, Krueger A, Walter L, Wood C, Rose W, Adams J,Wolff R (1991) Analysis of active volcanoes from the earthobserving system. Remote Sens Environ 36:1–12

Patrick M, Dean K, Dehn J (2004) Active mud volcanism observed withLandsat 7 ETM+. J Volcanol Geotherm Res 131(3–4):307–320

Pergola N, Pietrapertosa C, Lacava T, Tramutoli V (2001) Robustsatellite techniques for volcanic eruptions monitoring. AnnGeophys 44(2):167–177

Pergola N, Marchese F, Tramutoli V (2004) Automated detection ofthermal features of active volcanoes by means of infraredAVHRR records. Remote Sens Environ 93:311–327

Pergola N, Marchese F, Tramutoli V, Filizzola C, Ciampa M (2008)Advanced satellite technique for volcanic activity monitoring andearly warning. Ann Geophys 51(1):287–301

Pergola N, Giuseppe DA, Lisi M, Marchese F, Mazzeo G, TramutoliV (2009) Time domain analysis of robust satellite techniques(RST) for near real-time monitoring of active volcanoes andthermal precursor identification. Phys Chem Earth 34:380–385

new paradigm for volcanological observations from orbit. JVolcanol Geotherm Res 135(1–2):13–28

Pieri D, Abrams M (2005) ASTER observations of thermal anomaliespreceding the April 2003 eruption of Chikurachki volcano,Kurile Islands, Russia. Remote Sens Environ 99(1–2):84–94

Platnick S, King MD, Ackerman SA, Menzel WP, Baum BA, RiédiJC, Frey RA (2003) The MODIS cloud products: algorithms andexamples from Terra. IEEE Trans Geosci Remote Sens 41(2):459–473

Prins E, Feltz J, Menzel W, Ward D (1998) An overview of GOES-8 diurnal fire and smoke results for SCAR-B and 1995 fire season inSouth America. J Geophys Res 103(D24):31821–31835

Roberts G, Wooster M (2008) Fire detection and fire characterizationover Africa using Meteosat SEVIRI. IEEE Trans Geosci RemoteSens 46(4 Part 2):1200–1218

Rothery DA, Thorne MT, Flynn L (2003) MODIS thermal alerts inBritain and the North Sea during the first half of 2001. Int JRemote Sens 24(4):817–826

Tramutoli V (1998) Robust AVHRR Techniques (RAT) for environ-mental monitoring theory and applications. In: Checchi G, ZilioliE (eds) Earth surf remote sens II. SPIE, Barcelona, Spain, pp101–113

Tramutoli V, Di Bello G, Pergola N, Piscitelli S (2001) Robust satellitetechniques for remote sensing of seismically active areas. AnnGeophys 44(2):295–312

Trunk L, Bernard A (2008) Investigating crater lake warming usingASTER thermal imagery: case studies at Ruapehu, Po·s, KawahIjen, and CopahuÈ Volcanoes. J Volcanol Geotherm Res 178(2):259–270

592 Bull Volcanol (2011) 73:577–593

Pieri D, Abrams M (2004) ASTER watches the world's volcanoes: a

Trusdell FA, Moore RB, Sako M, White RA, Koyanagi SK, Chong R,Camacho JT (2005) The 2003 eruption of Anatahan volcano,Commonwealth of the Northern Mariana Islands: chronology,volcanology, and deformation. J Volcanol Geotherm Res 146(1–3):184–207

Webley P, Wooster M, Strauch W, Saballos J, Dill K, Stephenson P,Stephenson J, Wolf E, Matias O (2008) Experiences from near-real-time satellite-based volcano monitoring in Central America:case studies at Fuego, Guatemala. Int J Remote Sens 29(22):6621–6646

Wooster MJ (2001) Long-term infrared surveillance of LascarVolcano: contrasting activity cycles and cooling pyroclastics.Geophys Res Lett 28(5):847–850

Wooster MJ, Wright R, Blake S, Rothery DA (1997) Coolingmechanisms and an approximate thermal budget for the 1991–1993 Mount Etna lava flow. Geophys Res Lett 24:3277–3280

Wright R, De La Cruz-Reyna S, Harris AJL, Flynn LP, Gomez-Palacios JJ (2002a) Infrared satellite monitoring at Popocatepetl:

explosions, exhalations, and cycles of dome growth. J GeophysRes 107(B8):2153

Wright R, Flynn L, Garbeil H, Harris A, Pilger E (2002b) Automatedvolcanic eruption detection using MODIS. Remote Sens Environ82:135–155

Wright R, Flynn L, Garbeil H, Harris A, Pilger E (2004) MODVOLC:near-real-time thermal monitoring of global volcanism. J Volca-nol Geotherm Res 153:29–49

Wright R, Carn SA, Flynn LP (2005) A satellite chronology of theMay-June 2003 eruption of Anatahan volcano. J VolcanolGeotherm Res 146:102–116

Xiong X, Wu A, Caos C (2008) On-orbit calibration and inter-comparison of Terra and Aqua MODIS surface temperaturespectral bands. Int J Remote Sens 29(17–18):5347–5359

Xu W, Wooster M, Roberts G, Freeborn P (2010) New GOES imageralgorithms for cloud and active fire detection and fire radiativepower assessment across North, South and Central America.Remote Sens Environ 114(9):1876–1895

Bull Volcanol (2011) 73:577–593 593