DETECTING TERRESTRIAL CAVES BY APPLYING TOPOGRAPHIC ANALYSIS TECHNIQUES TO
THERMAL IMAGERY. J.J. Wynne1, J. Jenness2, M.D. Jhabvala3, T.N. Titus4 and D. Billings5. 1The SETI Institute, Carl
Sagan Center, Mountain View, CA and Department of Biological Sciences, Northern Arizona University, Flagstaff, AZ,
email@example.com; 2Jenness Enterprises, GIS Analysis & Application Design, Flagstaff, AZ; 3NASA Goddard Space Flight
Center, Instrument Systems and Technology Division, Code 550, Greenbelt, MD; 4U.S. Geological Survey, Astrogeology Sci-
ence Center, Flagstaff, AZ; and, 5National Speleological Society, Desert Dog Troglodytes, Los Angeles, CA.
Introduction: Since Rinker’s  groundbreaking
work on terrestrial cave detection nearly 50 years ago,
our ability to find caves using airborne and spacecraft
acquired imagery has improved considerably. Due to
superior analytical techniques, improved instrument
optics, and high resolution imagery, we have furthered
terrestrial cave detection capabilities [2,3,4] and con-
firmed cave-like features on the Moon [5,6,7] and
Mars [8,9,10]. Lunar caves may serve as the best loca-
tions for human habitation [2,4,6], while Martian caves
are of great interest as astrobiological targets, access-
ing potential water-ice reserves, as well as astronaut
bases [2,4]. Further, geothermal vents associated with
vapor plumes identified on Saturnian , Jovian
[12,13] and Neptunian moons  hint at additional
planetary subterranean access points, and represent
high priority targets for future habitability studies.
Cave detection is typically most successful when
multiple thermal images are acquired during both the
warmest (mid-afternoon) and coolest (predawn) times
of day . Although data acquisition is logistically
easiest on Earth, repeat thermal imagery over short
temporal periods is lacking for most terrestrial loca-
tions. When searching for caves on other planetary
bodies, obtaining multiple images for regions of inter-
est within a limited window of time is challenging.
Accordingly, researchers must rely on imagery sporad-
ically captured from spacecraft platforms where acqui-
sition is dictated by fly-by or orbiter mission schedules
and tempered by other mission objectives.
For extraterrestrial cave detection, we explored two
targeting categories using terrestrial analogs. (1) Deep
caves, which have sufficient linear length and/or depth
to adequately buffer interior environments from harsh
surface conditions. On Mars, these caves would be
among the best candidates to search for evidence of
life. (2) Shallow caves, which extend tens of meters in
length, may represent suitable sites for establishing
astronaut bases on the Moon and Mars. Within these
less protected sites, habitat pods may be constructed or
inserted a small distance within the cave entrance.
Hypotheses: H1: Thermal imagery may be exam-
ined using techniques for analyzing digital elevation
models. All cave entrances (deep and shallow) will
appear as distinct “topographic” features on a 3-D rep-
resentation of a thermal image. We tested this idea
using: (1) topographic position index (TPI), (2) ther-
mal gradient (or slope), and (3) curvature (Fig. 1A).
TPI is the relative value of a specific pixel compared to
the average thermal value of a neighborhood of pixels.
Thermal gradient is the rate of change of thermal val-
ues indicating where temperature values change quick-
ly. Curvature represents the rate of change in slope on
the “thermal surface” depicted as regions of concavity
(dips or valleys) and convexity (hills and peaks).
Figure 1. (A) Example of image layers developed for topo-
graphic analysis. (B) Method for extracting pixels values for
deep cave, shallow cave and control points (random surface
locations) used to test H1 and H2. Resolution ~0.5 m2.
H2: Thermal signatures associated with deep cave
entrances are discernable from shallow caves (e.g.,
alcoves and caves ≤ 50 m in length) and control points
(i.e., random areas of surface) when imagery is cap-
tured at the appropriate time for a single acquisitional
period (either predawn or midday).
9029.pdf2nd International Planetary Caves Conference (2015)
Data and Analysis: We collected thermal imagery
of Pisgah Lava Beds, located ~175 miles northeast of
Los Angeles, CA. Using a Quantum Well Infrared
Photodetector thermal imaging camera  mounted
in the forward camera port of a Beechcraft King Air
B200 (operated by NASA Armstrong Flight Research
Center), we acquired imagery at nadir. Two missions
were flown – one from 1222 to 1457h on 11 April
2011 (4,445 midday images), another from 0426 to
0649h on 13 April 2011 (3,675 predawn images; ).
For our study sites (nine deep caves, four shallow
caves and 13 control sites), a seven point grid was cen-
tered over each feature and the pixel values (i.e., digi-
tal number, DN) extracted (Fig. 1B). We analyzed
DNs of these three groups using Welch’s t-tests 
for all paired comparisons of mean actual thermal, TPI,
gradient and curvature values.
Results: H1: Barring shallow caves vs. control sites
for curvature in the predawn imagery (p = 0.07), deep
and shallow cave entrances were significantly different
from control sites for TPI, gradient, and curvature (p ≤
0.02). Deep caves were not discernable from shallow
caves for all topographic techniques (p ≥ 0.35).
Figure 2. Scatter plot of diurnal difference and diurnal
mean for midday and predawn of actual thermal DN values
of deep cave (red), shallow cave (green) and surface control
sites (blue). DNs = digital numbers.
H2: Most paired tests of actual thermal imagery were
significantly different (p ≤ 0.042), except for deep cave
vs. control sites in the predawn imagery (p = 0.118)
and deep vs. shallow caves in the midday imagery (p =
0.117). In most cases, topographic analyses techniques
differentiated deep and shallow caves from control
sites (H1 above).
Discussion: Applying topographic analytical tech-
niques shows promise as a valid method for analyzing
thermal imagery. Using these techniques, we were able
to differentiate deep and shallow caves from control
sites except for one case (shallow cave vs. control sites
for curvature of predawn imagery). In one case, we
were able to discern deep from shallow caves using the
actual thermal of predawn imagery. Because this result
occurred only once, we interpret it cautiously. Our
limited ability to detect differences between deep and
shallow features may be due to a small sample of shal-
low caves (n = 4). Based upon this limited sample,
these features exhibit thermal behaviors similar to deep
caves (Fig. 2). Further, this result may be due to limita-
tions in our seven point DN extraction method. While
this systematic approach appears successful overall, it
may be affecting the results by limiting the DN values
analyzed for each feature. Most of the cave entrances
were either crescent (lateral entrances viewed at nadir)
or oblong in shape. Thus, our method of selecting
points in a circle may be extracting DN values in loca-
tions where surface temperatures are interacting with
cave temperatures resulting in subpixel mixing .
Our research is proceeding in the following areas: (1)
examining different pixel extraction methods used for
dataset development; (2) increasing sample size by
obtaining representative samples for analysis of differ-
ent entrance shapes (crescent, oblong and oval) for
both deep and shallow caves through fieldwork verifi-
cation; (3) analyzing signal strength of deep cave en-
trances by comparing deep cave entrance size and
length against thermal pixel values of these deep cave
entrances; and, (4) identifying regions on the landscape
most similar to known cave locations using Mahalano-
bis distances  (which describe the distance in mul-
tivariate space of the mean actual thermal, TPI, gradi-
ent, and curvature values) and confirm their identity
(deep or shallow caves) through a field effort.
Acknowledgments: This project was supported by
NASA Exobiology Program grant #EXOB07-0040 and
the NASA Goddard Space Flight Center. Glen Cush-
ing, Jasmeet Dahliwal, Laszlo Kestay and Cecilia
Leung provided valuable comments that improved this
References:  Rinker J. N. (1975), Photogram. Eng.
Remote Sens. 41, 1391–1400.  Wynne J.J. et al. (2009),
Abstract 2451, 40th LPSC, Houston, TX;  Titus, T.N. et al.
(2011), Abstract 8024; 1st Internatl. Planet. Cave Workshop,
Carlesbad, NM;  Wynne, J.J. et al. (2008), Earth Planet.
Sci. Lett. 272: 240–250;  Haruyama, J. et al. (2010), Ab-
stract 1285, Lunar Planet. Sci. 41;  Ashley et al. (2012), J.
Geophys. Res. 117, E00H29;  Wagner, R.V. & M.S. Rob-
inson (2014), Icarus 237: 52–60;  Cushing, G.E. et al.
(2007), Geophys. Res. Lett. 34, L17201;  Cushing, G.E.
(2012), J. Cave Karst Stud. 74: 33–47;  Cushing, G.E. et
al. (2015), J. Geophys Res-Planet 120: 1023–1043; 
Hansen, C.J. et al. (2011), Geophys. Res. Lett. 38, L11202;
 Geissler, P.E. & M.T. McMillan (2008), Icarus 197:
505–518;  Roth et al. (2014), Science 343: 171–174; 
Duxbury, N.S. & R.H. Brown (1997), Icarus 125: 83–93;
 Jhabvala, M.D. et al. (2012), Proc. SPIE 8268, 82682P.
 Welch, B. L. (1947), Biometrika 34: 28–35;  Foody,
G.M. & D.P. Cox (1994), Int. J. Remote Sens. 15: 619–631.
 Mahalanobis, P. (1930), J. Proc. Asiatic Soc. Bengal 36:
9029.pdf2nd International Planetary Caves Conference (2015)