Multispectral Classification With Bias-Tunable Quantum Dots-in-a-Well Focal Plane Arrays

Abstract

Mid-wave and long-wave infrared (IR) quantum-dots-in-a-well (DWELL) focal plane arrays (FPAs) are promising technology for multispectral (MS) imaging and sensing. The DWELL structure design provides the detector with a unique property that allows the spectral response of the detector to be continuously, albeit coarsely, tuned with the applied bias. In this paper, a MS classification capability of the DWELL FPA is demonstrated. The approach is based upon: 1) imaging an object repeatedly using a sequence of bias voltages in the tuning range of the FPA and then 2) applying a classification algorithm to the totality of readouts, over multiple biases, at each pixel to identify the “class” of the material. The approach is validated for two classification problems: separation among different combinations of three IR filters and discrimination between rocks. This work is the first demonstration of the MS classification capability of the DWELL FPA.

Full text

Page 1

1342IEEE SENSORS JOURNAL, VOL. 11, NO. 6, JUNE 2011
Multispectral Classification With Bias-Tunable
Quantum Dots-in-a-Well Focal Plane Arrays
Biliana S. Paskaleva, Member, IEEE, Woo-Yong Jang, Student Member, IEEE, Steven C. Bender,
Yagya D. Sharma, Senior Member, IEEE, Sanjay Krishna, Senior Member, IEEE, and
Majeed M. Hayat, Senior Member, IEEE
Abstract—Mid-wave and long-wave infrared (IR) quantum-
dots-in-a-well (DWELL) focal plane arrays (FPAs) are promising
technology for multispectral (MS) imaging and sensing. The
DWELL structure design provides the detector with a unique
property that allows the spectral response of the detector to be
continuously, albeit coarsely, tuned with the applied bias. In
this paper, a MS classification capability of the DWELL FPA is
demonstrated. The approach is based upon: 1) imaging an object
repeatedly using a sequence of bias voltages in the tuning range
of the FPA and then 2) applying a classification algorithm to the
totality of readouts, over multiple biases, at each pixel to identify
the “class” of the material. The approach is validated for two
classification problems: separation among different combinations
of three IR filters and discrimination between rocks. This work is
the first demonstration of the MS classification capability of the
DWELL FPA.
IndexTerms—Biastunability,dots-in-a-well(DWELL),infrared
(IR) detector, multispectral (MS) classification, quantum-dots,
rock classification.
I. INTRODUCTION
T
multiple detectors, each sensing at a specific range of wave-
lengths.Anotheralternativeistouseasinglebroadbanddetector
combined with a dispersive optical system, such as a bank of IR
optical filters, each tuned to a specific wavelength band. In ei-
ther case, the sensor represents a highly complex optomechan-
YPICALinfrared (IR) multispectral(MS) and hyperspec-
tral (HS) systems are usually implemented by deploying
Manuscript received March 24, 2010; revised October 30, 2010; accepted
November 17, 2010. Date of publication December 03, 2010; date of current
versionApril 20,2011.This workwassupportedinpartbytheNationalScience
FoundationunderAwardIIS-0434102andAwardECS-401154.Additionalsup-
port wasprovided by theNationalConsortium for MASINTResearch through a
Partnership Project and the National Science Foundation led by the Los Alamos
National Laboratory and the University of New Mexico. (Sandia National Lab-
oratories is a multi-program laboratory managed and operated by Sandia Cor-
poration, a wholly owned subsidiary of Lockheed Martin Corporation, for the
U.S. Department of Energy’s National Nuclear Security Administration under
Contract DE-AC04-94AL85000.) The associate editor coordinating the review
of this paper and approving it for publication was Dr. M. Abedin.
B. S. Paskaleva was with the Center for High Technology Materials and
the Department of Electrical and Computer Engineering, University of New
Mexico, Albuquerque, NM 87106 USA. She is now with Sandia National
Laboratories, Albuquerque, NM 87123 USA (e-mail: bspaska@sandia.gov).
W.-Y. Jang, Y. D. Sharma, S. Krishna, and M. M. Hayat are with the
Center for High Technology Materials and the Department of Electrical
and Computer Engineering, University of New Mexico, Albuquerque,
NM 87106 USA (e-mail: wjang@ece.unm.edu; yagya@chtm.unm.edu;
skrishna@chtm.unm.edu; hayat@ece.unm.edu).
S. C. Bender is with the Los Alamos National Laboratory, Los Alamos, NM
87545 USA (e-mail: sbender@lanl.gov).
Digital Object Identifier 10.1109/JSEN.2010.2095456
ical system that requires precision alignment and calibration.
Once the calibration is completed and the sensor is deployed,
thesensor’s functionality cannot be easily modified.As a result,
thesensor cannot be easily adapted to take advantage of specific
sensing situations. In the absence of an agile software/hardware
architecture, one is typically forced to acquire all available im-
agery data before its relevance becomes clear. This leads to the
acquisition of maximum and often massive amounts of data that
has to be stored for subsequent processing in applications such
asclassification,abundanceestimation,imagesegmentationand
analysis, etc. Besides the large storage demands, the analysis of
this MS and HS imagery requires powerful hardware systems
and efficient processing algorithms.
The inflexibility of the functionality of most present-day
MS/HS sensors to adapt to specific applications has prompted
the development of a new modality for MS and HS sensing,
providing greater flexibility in terms of compressive data ac-
quisition and offering reduced complexity and cost. Spectrally
adaptive sensing methods include those that are based on mi-
croelectromechanical systems (MEMS) [1] and acousto-optic
tunable filters (AOTFs) [2]. However, a new emerging tech-
nology for adaptable and compressive sensing is the quantum
dots-in-a-well (DWELL)-based IR focal plane array (FPA)
[3]–[5]. Owing to the quantum-confined Stark effect (QCSE)
[6], [7], the DWELL sensor exhibits a unique feature that
enables continuous spectral tuning in the mid-wave infrared
(MWIR) and long-wave infrared (LWIR) spectral regions by
means of changing the applied bias voltage. Such tunable
sensors provide greater optical simplicity because their spec-
tral response is controlled electrically rather than optically or
mechanically.
Fig. 1(a) shows the spectral responses of a single-pixel
DWELL device. As seen in the figure, the central wavelength
and the shape of the detector’s responsivity change continu-
ously with the applied bias voltage. Furthermore, the spectral
diversity offered by the DWELL can be further enhanced by
design via optimizing the well width and the asymmetric band
structure. As a result, in the context of MS and HS sensing, a
single DWELL detector can be exploited as a MS IR sensor; the
photocurrents measured at different operational biases can be
viewed as outputs of different spectrally broad and overlapping
bands. Clearly, this capability is an excellent fit to compressive
sensing once the sensor is combined with algorithms and
reconfigurable readout integrated circuits (ROICs).
The flexibility offered by the DWELL FPA is not without a
price, however. For instance, the DWELL’s spectral response
is relatively broad (
). As a result, the spectral bands
1530-437X/$26.00 © 2010 IEEE

Page 2

PASKALEVA et al.: MULTISPECTRAL CLASSIFICATION WITH BIAS-TUNABLE QUANTUM DOTS-IN-A-WELL FOCAL PLANE ARRAYS1343
Fig. 1. (a)SpectralresponsesofaDWELLdetectorasafunctionofthe applied
bias which demonstrate the tuning capability. (b) Series of synthesized filters
(solid lines) in the projection stage of algorithmic spectrometer with full width
half maximum (FWHM) of 0.3 ??.
corresponding to different bias voltages exhibit significant
overlap. Another complication is the bias-dependence of the
noise (darkcurrent) inthephotocurrents. In ourpreviousworks,
we targeted addressing both of these challenges. In particular,
the DWELL-based algorithmic spectrometer (DAS), proposed
in [8] and demonstrated in [9] and [10], banks on using linear
superposition of bias-tunable bands of the individual DWELL
detector to minimize the effect of high correlation in the
DWELL’s bands in the presence of noise in achieving target
spectrum reconstruction. Fig. 1(b) shows series of synthe-
sized filters (solid lines) which approximate ideal tuning filters
(dashed lines) in the projection stage of the DAS. The full width
half maximum (FWHM) of the synthesized filters is 0.3
Another example is the canonical correlation feature selec-
tion(CCFS)algorithm,reportedin[11]and[12].Thisalgorithm
addresses the problem of linear superposition of bias-tunable
DWELL bands to perform spectral feature selection (for mate-
rial classification) based on spectral matched filtering. Our pre-
vious work [11] also includes successful demonstration of MS
classification of the DWELL detector, at a single-pixel level,
in terms of rock-type classification. The study was conducted
using laboratory spectral data for the rock types and spectral re-
sponsivities measurements of the DWELL detector.
In this paper, we demonstrate for the first time MS classifica-
tion using imagery obtained from the DWELL FPA. In our set-
ting, multiple DWELL-FPA images are taken at different bias
voltages. As such, the sequence of all images can be viewed as
a MS imagery. We performed two studies. In the first study, we
compare the classification error among different materials for
two general classification problems using few combinations of
biasvoltages.Thefirstclassificationproblem,termedfilterclas-
sificationproblem,isthatofclassifyingcombinationsofMWIR
and LWIR optical filters with different bandwidths and center
.
wavelengths. The second classification problem, termed rock
classification problem, focuses on classification between pairs
of rocks drawn from a set of three distinct rock types: granite,
hornfels, and limestone.
The second study is a separability analysis and optimal bias
selection for classification. The separability analysis focuses on
investigation of the separation between pairs of materials as a
function of the DWELL’s bias voltage. For the granite-lime-
stone and granite-hornfels classification problems, we carry out
exhaustive search over all possible combinations of the bias
voltages in the context of optimal bias selection based on mini-
mization of the classification error.
The organization of this paper is as follows. In Section II,
we briefly describe the development and the operation principle
of the DWELL FPA, followed by the description of the bias-
dependent spectral tunability capability of the DWELL FPA.
Section III describes the MS classification approach for the two
classification problems described above. In Section IV, the MS
classification results are discussed. The separability and clas-
sification analysis for optimal bias selection are also presented
and discussed in Section IV. Our conclusions are presented in
Section V.
II. THE DWELL FOCAL PLANE ARRAY
In this section, we briefly describe the operation principle,
characterization and bias-dependent spectral tunability of the
DWELL FPA.
A. Operation Principle and Spectral Characterization of the
DWELL FPA
The DWELL photodetector, pioneered by Krishna [13], is a
hybrid version of quantum dot (QD) and quantum well (QW)
photodetectors. The DWELL photodetector reported in [9] has
already been shown to exhibit bias tunability in the range of
MWIR (3–5
) to the LWIR (8–12
trum. In general, the MWIR response is driven by a bound-to-
continuum transition, while the LWIR is driven by the bound
state in the dot-to-a-bound state in the well transition, as shown
in Fig. 2. In addition, the asymmetry of the electronic poten-
tial controlled by the shape of the dot and the different thick-
nesses of QW above and below the dot, results in variation of
the local potential as a function of the applied bias. As a result,
byadjustingtheappliedbiasvoltageonthedevice,spectralshift
(calledalso“redshift”)andoverlapsareobtained,particularlyin
LWIR (8–12
) region.
The spectral responses of the single-pixel DWELL shown in
Fig. 3 demonstrate bias-dependent spectral tunability for var-
ious device operating temperatures. The details of the device
characterizationhadbeenreportedin[9].TheDWELLhasbeen
fabricatedintoa320by256detectorarrayformatandisusedfor
this study. The fabrication process is described in great detail in
[14] and [15]. The DWELL FPA responses have been charac-
terized by using CamIRa demonstration system.1Recently, an
optimized DWELL FPA was reported in [14] demonstrating an
increase in the operating temperature (up to 80 K) and smaller
) portions of the spec-
1Manufactured by SE-IR Corporation, Goleta, CA. Use of the trade name is
for descriptive purposes only and does not constitute endorsement.

Page 3

1344IEEE SENSORS JOURNAL, VOL. 11, NO. 6, JUNE 2011
Fig. 2. Top: schematic of the energy transition levels in the conduction band
of DWELL. Bottom: growth schematic of single-pixel DWELL (adapted from
[9].).
Fig. 3. Bias-tunable spectral responses of the single-pixel DWELL photode-
tector for various operating temperatures (adapted from [12]).
noise equivalent difference in temperature (min. NEDT2around
78mK).Thehigheroperatingtemperaturehasbeenachievedby
astrainreductionandanincreasednumberofstacksintheactive
region, improving the responsivity and the absorption quantum
efficiency.
2The NEDT is a sensitivity measure of a detector, indicating the smallest dif-
ference in uniform scene temperature that a detector can detect, so a small value
of NEDT is desired.
Fig. 4. Columns one, three, and five, a-d show DWELL FPA raw imagery ac-
quired at 0.3, 0.7, and 1.2 V, respectively. Columns two, four, and six, a-d, show
the normalized imagery at 0.3, 0.7, and 1.2 V, respectively. For more details
on the normalization, please refer to Section III. Objects in scene; scene in row
(a): filters ?? (left) and ?? (right); scene in raw (b): filters ?? (left)
and ?? (right); scene in row (c): filters ?? (left), ?? (center) and ??
(right); scene in row (d): filter ?? (top), limestone (left) and granite (right).
B. Bias Tunability of the DWELL FPA
DWELL FPA imagery shown in Fig. 4(a)–(d) are used to
demonstrate the DWELL FPA bias tunability. For all imagery,
the operating temperature of the DWELL FPA was set to 60 K
and the integration (exposure) time was 11.5 ms. The images
shownincolumnsone,threeandfiveinFig.4(a)–(d)aretakenat
0.3, 0.7,and 1.2 V, respectively. Normalized images at 0.3, 0.7,
and 1.2 V are shown in columns two, four, and six, respectively,
in Fig. 4(a)–(d). The DWELL FPA data is normalized at each
pixel by the approximate area of the multibias pixel response in
order to eliminate the intensity effect in the calculations. More
details about the normalization are given in Section III.
Fig. 4(a)–(c) contains images of different configurations of
three IR optical filters, manufactured by Northumbria Optical
Coatings, Ltd. The spectral responses of the filters are shown
in Fig. 5 (left). The first scene, shown in Fig. 4(a) includes
two IR filters: filter at 3–4
termed, metal filter holders, and a blackbody background
at 150 C. A 150 C temperature was used since such a high-
temperature blackbody offered a good transmittance for objects
in a scene. The blackbody is manufactured by MIKRON com-
pany (model M315) providing a temperature between ambient
5 C and 350 C, a control to within 0.2 C and an emissivity
of
.
The second scene shown in Fig. 4(b) consists of two filters:
and filter at 8.5termed
holders and the uniform background at the same temperature.
The third scene in Fig. 4(c) consists of all three filters
, andand the background. The scene in Fig. 4(d)
includes two rocks: granite and limestone, and the
Granite is a common and widely occurring type of intrusive,
felsic igneous rock. Granites usually have a medium to coarse
grained texture. Limestone is a sedimentary rock composed
largely of the mineralscalcite and aragonite, which are different
crystal forms of calcium carbonate. Hornfels is a fine-grained
nonfoliated metamorphic rock with no specific composition. It
is produced by contact metamorphism. Normalized reflectance
termed, filter at 4–5
, the same metal filter
,
filter.

Page 4

PASKALEVA et al.: MULTISPECTRAL CLASSIFICATION WITH BIAS-TUNABLE QUANTUM DOTS-IN-A-WELL FOCAL PLANE ARRAYS 1345
Fig. 5. Left: spectral responses of the three IR optical filters: ?? , ?? ,
and ?? . Right: normalized (by the peak value) reflectance spectra of granite,
hornfels, and limestone.
Fig. 6. Left: normalized multibias FPA responses for background, metal
holder, ?? , and ?? as a function of applied DWELL FPA bias. Right:
normalized multibias FPA responses for background, ?? , limestone, and
granite as a function of applied DWELL FPA bias.
measurements of granite, limestone, and hornfels using a
broadband single-pixel HgCdTe device cooled to 77 K are
shown in Fig. 5 (right).
Fig. 6, left and right, shows plots of the normalized DWELL
FPA multibias responses for every object from the scenes in
Fig. 4(a) and (d), respectively. The multibias response for
every object is averaged over spatially uniform regions that are
visually associated with that object. From the plots in Fig. 6
(left), we can observe that at bias voltages 0.7 and 0.8 V the
responses of the DWELL FPA for all objects exhibit significant
overlap. Higher separation between all objects for this problem
is observed in the bias range of 0.3–0.6 V and in the range of
1.0–1.2 V.
The normalized multibias FPA responses for granite, lime-
stone, background and filter comprising the scene in Fig. 4(d)
overlap at bias voltages 0.7 and 0.8 V, as observed in Fig. 6
(right).Thepairwiseseparabilitybetweenthenormalizedmulti-
bias FPA responses for granite and limestone has the lowest
value at 0.6 V. At 0.9 V, the normalized multibias FPA re-
sponses for the black-body background and the filter are very
closetoeachother.Therest ofthebiasesprovideagoodsepara-
bility between the black-body background and the filter, and the
black-bodybackgroundandeachoneofthetworocks.Thepair-
wise separability between the multibias responses for the two
rocks, granite and limestone, is low across the entire bias range
making this problem very challenging for MS classification.
Fig. 7, left and right, shows plots of the spectral ratios for
pairs of sensed materials in as a function of the applied bias.
Fig. 7 (left) shows the spectral ratios calculated for the pair of
objects (filters, background and metal holder) from the scene
in Fig. 4(b). The spectral ratios vary between 0.4 to almost 1.4
when the applied bias changes in the range from 0.3 to 1.2 V
with a step of 0.1 V. Note that for bias voltages 0.7 and 0.8 V,
Fig. 7. Left: ratio of pixel values for various pairs of the objects ?? , ?? ,
metal holder, and the background as a function of applied DWELL FPA bias.
Right: ratio of pixel values between objects granite and limestone, granite and
background, and limestone and background as a function of applied DWELL
FPA bias.
theratiosbetweenpairofobjectsareclosetoone,indicatinglow
spectral separability between the materials at these particular
biases.
The fact that the ratio values change from one bias to another
indicates that the DWELL FPA can sense different spectral con-
tents of the targets observed in a scene simply by changing the
applied bias. Note that for the conventional (nontunable) de-
tector the spectral ratios would remain fixed as a function of
the applied bias.
Fig. 7 (right) shows spectral ratio plots for two rocks: granite
and limestone, and the background. As observed from the plots
in Fig. 7 (right), the ratio between granite and limestone does
not exhibit wide range as, for example, the granite-background
ratio or limestone-background ratio. Note that for bias voltages
0.6, 0.7, and 0.8 V, the ratios between granite and limestone are
close to one, indicating low spectral separability at these biases.
The classification results presented in Section IV demonstrate
that the spectral contrast captured by the bias-tunable DWELL
FPAissufficienttodiscriminatebetweenthetwotypesofrocks.
III. MULTISPECTRAL CLASSIFICATION USING
BIAS TUNABLE DWELL FPA
In this section, we provide a brief overview of the mathemat-
ical model for bias-tunable MS sensing and discuss the classifi-
cation approach.
A. Bias-Tunable MS Sensing
Mathematically, the DWELL spectral bands can be viewed
as a family of functions
bias voltages
[11]. In what follows, we denote the spectrum
of an object by
. For example,
mittance, bidirectional reflectance measurement or hemispher-
ical reflectance data. The photocurrent for the th band of the
DWELL detector sensing an object with a given spectrum
can be written as
, parameterized by the applied
may represent trans-
(1)
Here,
with the th band, and the interval
common spectral support for all bands and objects. Next, for a
denotes additive, scene-independent noise associated
represents the

Page 5

1346IEEE SENSORS JOURNAL, VOL. 11, NO. 6, JUNE 2011
TABLE I
SUMMARY OF IDENTIFIED CLASSES FOR THE FILTERS AND
ROCK CLASSIFICATION PROBLEMS
givensetofappliedbiasvoltages
DWELL detector is a set of photocurrents at these bias voltages
,theoutputofthe
(2)
This set represents the multibias or MS signature of the object
as “seen” by the DWELL dectector.
Because the DWELL bands are wide and overlapping, the
photocurrents in
are highly correlated. The redundancy in the
information content of the photocurrents can be reduced by a
suitable postprocessing algorithm, which, in turn, can be used
to improve the efficiency of the classification process. Here, we
shall use the CCFS [11] algorithm to replace the -dimensional
multibias signature in (2) by a single feature that is optimized
with respect to a given class of objects.
For a given class of objects represented by a mean spectrum
, the output from the CCFS algorithm is a single trans-
formed feature
combination of all features in (2). The weights
bytheCCFSforeveryclassofobjectsrepresentedbytheirmean
spectrum
.
The transformed feature
can be viewed as the current gen-
erated by a “virtual” superposition band,
the optimal selection rule of the weights is derived rigorously
in [11]. Consequently, the problem of determining the optimal
current, , for a given class representative or class mean spec-
trum
, is equivalent to finding a superposition band
provides the best approximation of
be interpreted as an approximation of
by
, which minimizes the distance and at the same time
maximizes the signal-to-noise ratio [11].
, which is a weighted linear
are optimized
;
that
can . Mathematically,
in the space spanned
B. Classification Problems
Thefirstclassificationproblemconsideredinthispaperisthat
of separating multiple combinations of MW and LWIR spectral
filters with different bandwidths and center wavelengths. For
this problem, we used the three scenes shown in Fig. 4(a)–(c).
The second classification problem is to discriminate between
pairs of rocks drawn from a set of three distinct rock types:
granite, hornfels, and limestone. The scene configurations for
thisproblemareshowninFigs.4(d)and9,left.Theclassesiden-
tifiedforbothclassificationproblemsaresummarizedinTableI.
Two types of normalization techniques are applied to the
raw digital numbers (DNs) that are retrieved directly from the
DWELL FPA. First, at each bias voltage, pixel’s DN values
are radiometrically corrected by a two-point nonuniformity
correction (NUC) algorithm. The NUC compensates for the
spatially nonuniform response of the detectors within the FPA
[16] and is an integrated part of the image acquisition process.
The two-point NUC is performed using temperatures at 22 C
and 150 C. The lower temperature of 22 C corresponds to
the lens-cap’s room temperature, which was used to yield the
lower-temperature uniform field.
Next,foreveryradiometricallycorrectedpixelanditsreplicas
at each bias voltage, the pixel’s value is normalized as follows:
(3)
where
FPA’s bias. Equation (3) is equivalent to normalization by the
area enclosed under the multibias response of each pixel in the
DWELLFPA.The normalized multibiasresponse ofa pixelcan
then be written as
isthevoltagestepsizeusedtoincrementtheDWELL
(4)
This normalization minimizes the role of broadband emissivity
in the discrimination process and emphasizes the spectral con-
trast.Thenormalizedimagesat0.3,0.7,and 1.2Vfor bothclas-
sification problems are shown in columns two, four, and six in
Fig. 4 (a)–(d) and in Fig. 9 (left), respectively.
We perform a supervised classification comprising of
training and testing steps for both classification problems. To
determine representative multibias signatures for each class
listed in Table I, we follow the same approach as used in [11].
Specifically, for each class we compute statistical mean and
covariance matrix using spatially uniform regions that are
visually associated with that class. Subsequently, Euclidean-
and Mahalanobis-distance classifiers are trained by the classes’
mean multibias signatures and the covariance matrices [17].
At the testing step, the trained classifiers are used to classify
the objects in Table I from a set of testing scenes. These scenes
capturethesameimagesasthetrainingscenesbutwereacquired
at different times. As a result, the testing scenes carry inherent
variability in the data due to the difference in the measurement
conditions from day-to-day and the presence of ambient and
system noise. The testing images are normalized in the same
fashion as the training images. The size of training and testing
data set for the filter and rock classification problems are listed
in Table II.
IV. DISCUSSION OF THE RESULTS
A. Classification Results
The thematic maps for the filter and rock classification
problems using Euclidean-distance classifier are presented
in Figs. 8(a)–(d) and 9, respectively. These maps show the
distribution of the derived classes over the spatial area captured
by the DWELL FPA. Each map defines a partitioning of the
area into sets, each including the points with identical class
labels. In order to investigate the effect of the bias selection on
the classification accuracy, the classification is performed for
multiple combinations of biases.
The results for the filter classification problem, specified in
TableI,areshowninFig.8(a)–(c),andTableIIIshowsthecalcu-
lated classification errors per various class. The thematic maps