You are on page 1of 4

LAND COVER MAPPING CAPABILITY OF MULTISPECTRAL THERMAL DATA: THE

TASI-600 CASE STUDY

M. F. Carfora*, A. Palombo**, S. Pascucci**, S. Pignatti** and F. Santini**

*National Research Council, Istituto per le Applicazioni del Calcolo "Mauro Picone" (CNR IAC),
80131 Napoli, Italy (a.carfora@iac.cnr.it)
**National Research Council, Institute of Methodologies for Environmental Analysis (CNR IMAA),
85050 Tito Scalo (PZ), Italy (federico.santini@imaa.cnr.it)

ABSTRACT applications [7], the advances made in the sensor technology


in the TIR domain have allowed the increasingly use of IR
This study shows the land cover mapping accuracy cooled detectors for the deployment of new airborne IR
retrievable by the TASI-600 thermal airborne multispectral hyperspectral spectrometers. Among them, the AHI and the
sensor and describes some of the classification results tested SEBASS and recently the TASI-600 (ITRES), the TELOPS
on the thermal preprocessed data for a rural area. (FIRST), and the AisaOWL (SPECIM) sensors are all
In the paper is provided an overview of the principal operating in the LWIR range [3-5, 8].
TASI-600 characteristics, i.e. 32 spectral bands in the 8.0- Within this context, the paper describes the TASI-600 [9]
11.5 μm spectral range, and land cover classification sensor characteristics’ and land cover classification results
performances. A full assessment of the TASI-600 spectral tested on the thermal preprocessed data in a rural area in
bands has been also obtained by ranking them in order to Belgium.
understanding their role in land cover classification. Concerning the classifiers tested in this study, at present
Results accuracies have been validated using available there are several methodologies for classifying vegetation
ground truth. The study highlights that the new generation multispectral images and even if standard univariate (i.e.,
of multi/hyperspectral thermal sensors opens up interesting single spectral band) classification techniques can be
opportunities for accurate land cover classification. adapted for application to hyperspectral data, their
effectiveness is often reduced by the presence of redundant
Index Terms — TASI-600, multispectral thermal data, or irrelevant information in the multivariate data set
land cover mapping, classification accuracies. procedures. To cope with performance and accuracy
problems associated with high dimensionality, it is
1. INTRODUCTION commonplace to apply a pre-processing step that transforms
the data to a space of significantly lower dimension with
In the last decades optical hyperspectral remote sensing limited loss of the information present in the original data
has been applied to understand and solve problems related [10]. Therefore, this paper applies a feature extraction and
to Earth surface processes in the framework of different dimension reduction algorithm based on discriminant
applications such as agriculture, forest, geology, analysis for supervised classification. The features produced
archaeology, land degradation, coastal zones, crisis by our procedure in the transformed space are uncorrelated
information [1, 2]. or independent, so that the multivariate density estimates
Moreover, there is a raising interest in applying remote characterizing each class are replaced by univariate product
sensing data in the thermal infrared spectral domain (LWIR estimators; moreover, since parametric densities do not
range; 8.0–14.0 μm). The use of remote sensing data in the adequately represent land categories, we estimate these
LWIR spectral range alone or joint to the VSWIR domain is univariate densities nonparametrically [10, 11].
producing new powerful results for detecting and
identifying surface materials, soil and plant properties, 2. TASI SENSOR DATA
pollutants and gases detection [3-5].
LWIR multi/hyperspectral data has been applied in the TASI-600 is a 32 bands multispectral sensor in the 8.0-
past in remote sensing research studies almost to map 11.5 μm spectral range, with a swath of 600 pixels for a
variations in surface and mineral composition [6]. Starting FOV of 40° and an IFOV of 1.2 mrad. TASI-600 sensor is
from TIMS, MODIS or ASTER multispectral satellite designed with custom athermal optics which removes the
thermal data that have been applied for different studies and need for internal environmental control systems that can fail
and produce electronic noise thus assuring a diffraction- normalizing the radiance of each pixel to the Planck’s curve
limited sensor. The TASI-600 system is composed by the that is generated from the pixel with the maximum
Scan Head Unit (SHU) provided with a custom Stirling brightness temperature with an emissivity value set to 0.96,
Cycle cooled MCT detector (HgCdTe) and the Instrument i.e., a reasonable hypothesis for the study area land cover.
Control Unit (ICU) and also of licensed pre-processing Last, TASI radiances and emissivities images were
software to radiometrically correct (RADCORR) and orthorectified by using the GEOCORR software obtaining
geometrically correct (GEOCOR) TASI data. RADCORR an accuracy of 2 pixels (i.e. 2 m).
removes system offsets (i.e. dark current) and data can be As regard the classification methods applied to the
output as spectral mSRU radiance units (1 mSRU is equal to corrected radiance and emissivity imagery, it is well known
1 nW cm-2 sr-2 nm-1) or apparent temperatures; while, the that classification of TASI-600 imagery aims at
GEOCOR software incorporates photogrammetric bundle discriminating different objects (in our case different land
adjustment solutions to produce orthorectified images using cover types) using multivariate information coming from
precision IMU/GPS and terrain height data. The instrument several spectral bands. Thus the problem is to assign an
was mounted on a Partenavia P68 Observer airborne (Fig. 1) unknown subject to one of k classes on the basis of a
and the imagery was acquired June, 2010 on a rural flat area multivariate observation. In discriminant analysis the
of an airfield in Belgium where different land covers occur. distribution of the observations in each class is characterized
The land cover classes considered in this study are: #1 by the class conditional probability density function fk. Once
dry grass, #2 apple trees, #3 sugar beet 1, #4 grass, #5 assigned the a priori probability πk of observing an
potatoes, #6 sugar beet 2, #7 wheat, #8 new asphalt, #9 old individual from any of the k populations (as the relative
asphalt, #10 urban, and #11 soil (the classes are reported in frequency of the classes in the labeled data, to fully exploit
Table 1 as class label). These classes were used for the information in the training dataset, or as uniform, when
retrieving the land cover capability mapping of the TASI- size or representativeness of the labeled information is not
600 radiances and emissivities. appropriate), the Bayes decision rule suggests to allocate
each observation x to the population which maximizes its
3. METHODS posterior probability πk fk(x). If we assume the densities fk to
be Gaussian, we obtain parametric rules such as linear
TASI-600 data preprocessing included first radiometric (LDA) and quadratic (QDA) discriminant analysis, for
correction of the data by using the RADCORR software common variance among classes and general covariance
provided by ITRES (Canada) and the application of a new structure, respectively. On the other hand, without any prior
correction tool for blinking pixel correction, developed by knowledge about the distribution of the data we can estimate
CNR IMAA (Italy). More in detail, TASI-600 data has been non-parametrically the unknown class conditional densities
re-processed to refine the radiometric accuracy and to apply from the training dataset. Moreover, to reduce such a
a new procedure to solve the blinking pixels correction challenging task to the product of univariate density
problem affecting thermal MCT sensors. estimation we apply PCA or ICA transforms to the original
Standard preprocessing procedures (e.g., those multivariate data before kernel estimation [10].
implemented in the ENVI 5 software) for thermal For this study, two nonparametric and two parametric
atmospheric correction and temperature-emissivity discriminant analysis methods for classifying TASI
separation were then applied. Regarding the atmospheric radiances and emissivities images have been considered: (a)
correction of the spectral radiances, it was performed by LDA (Linear Discriminant Analysis), assuming Gaussian
applying the ISAC (In-Scene Atmospheric Compensation) functions for the class conditional densities with common
algorithm [12]. This algorithm is a well established and variance among classes; (b) QDA (Quadratic Discriminant
commonly used for in scene atmospheric thermal data Analysis), assuming again Gaussian density functions but
correction and assumes two pixels of the scene to be with general covariance; (c) PCDA (Principal Component
blackbodies on which neither locations nor temperatures are Discriminant Analysis), where the original components are
known. The ISAC algorithm is suitable when the transformed into principal components prior to
atmospheric radiative conditions are not available during the nonparametric density estimation; (d) ICDA (Independent
acquisition time as in our case. For this study, we have component discriminant analysis), where the original
applied the “most hits” method as described by [12] and components are transformed into independent components
pixels whose emissivity was equal to 1 at the wavelength prior to nonparametric density estimation.
were used as a marker. Finally, to realize a good trade-off between classification
Once the images were atmospherically corrected, the accuracy and computational effort, reducing data
TES method used for this study to separate temperature and redundancy, we have investigated a progressive inclusion of
emissivity is a widely used method in the literature, i.e. the the bands in the classification analysis and estimate their
“emissivity normalization routine” [13]. This routine derives effectiveness in classifying the data, with the main objective
first the brightness temperature of each pixel from the pixel of quantitatively assessing their role in the overall accuracy.
radiance. Next, the apparent emissivity image is obtained by In particular, we consider a “simple forward” strategy,
where the first band is selected by an exhaustive search over considering that the temperature, in homogeneous solar
all N bands as the one that gives the best classification illumination conditions as for this study, depends on the
performance on the training dataset and each of the other interactions light-materials also in the VNIR spectral range,
bands is recursively selected by the same criterion among all thus facilitating to discriminate different materials occurring
the remaining bands. in a scene. Of course this method has to be applied
independently for each thermal image and depends on the
4. RESULTS AND DISCUSSION environmental condition thus limiting the use of spectral
libraries for classification.
We reported results of two experiments. In the first one, In particular, the last three columns of Table 1 show
we use as a training dataset the entire validation dataset and classification results in this case, with significant accuracy
evaluate the classification performance on the same dataset, improvement on every class. The worst performance on
while in the second experiment we randomly split the class #5 is due not only to the small size of the training
validation dataset in two parts that we use as training and dataset but mainly to the high spectral similarity of this class
validation dataset, respectively. In particular, two different to the more populated class #7; in this case the more robust
numerical experiments have been conducted in order to QDA procedure gave the best results.
evaluate the accuracy of the classification methods and Finally we tested the robustness of the proposed
robustness with respect to the training dataset: (a) a training procedures by considering different training and test
dataset has been defined starting from the ground validation datasets, as described in the setup of experiment (b).
and classification has been performed on the whole dataset Results confirm the general trend of the procedures, with
through the band selection procedure previously discussed a slight loss in accuracy.
and performance indicators evaluated for the same training
set; (b) to assess the performance of the proposed 5. CONCLUSIONS
procedures on external datasets, the experimental validation
dataset has been randomly splitted into two parts, and, then, The TASI-600 imagery results to be an advanced and
the first subset is used as a training dataset, whereas suitable remote sensing tool providing excellent quality
performance indicators are evaluated on the second one. In airborne thermal multispectral imagery to the science
order to be not sensitive to the particular random realization community involved in the Earth’s environment.
of the training dataset, the experiment is repeated 20 times Preliminary results of application of TASI-600 data for
with different random choices of the partitioning, and the land cover classification, here reported, are very promising
results are averaged. and suggest a more thorough investigation.
Results of the assessment of the role of the different These high spectral and spatial resolution airborne
spectral bands in reaching classification accuracy showed thermal data can be also applied for properly identifying
for both the radiance and the emissivity datasets that the science/application data products to be derived from next
three most informative bands for all methods were identified generation multi/hyperspectral satellite data such as
as 23-25 (10463 nm – 10682 nm), 12 (9259 nm), 1-2 (8054 HyspIRI (launch of the Hyperspectral Infrared Imager
nm-8164 nm). However, in contrast with our previous feasible for NASA on 2015) or LDCM (NASA's Landsat
results on classification of VNIR images [10], where a few Data Continuity Mission successfully launched on 11
spectral bands sufficed to reach good classification February, 2013) in view of a correct and updated Earth’s
accuracy, on this dataset almost every band bring a not surface composition mapping for the sustainable
negligible improvement to the overall result. Then we management of natural resources.
proceeded to classify the entire training set (Experiment
(a)). Table 1 shows classification accuracies as the observed
percentage of correctly classified pixels for each class and
for the whole volume of data for the considered methods
(ICDA results are not shown since they do not improve on
PCDA results, mainly because of the different class sizes).
It is evident that PCDA outperforms the parametric
methods, both on the radiance and emissivity data; however,
it is also clear that, while radiance data lead to a higher
success percentage in general, emissivity data allow to a
better discrimination of certain classes. This is the reason
why we decided to apply the classification algorithms on a
merged dataset, comprising the best spectral bands of both
datasets, as chosen by the band selection procedure.
Furthermore, a reasonable explanation of this is that we used Table 1. Size of the training set and performance indicators for
both radiance and emissivity datasets for classification the LDA, QDA, PCDA methods applied to the radiance, emissivity
and merged data set: agreement percentage S for the eleven [9] Pignatti, S., Lapenna, V., Palombo, A., Pascucci, S., Pergola,
classes and for the entire training set. N., Cuomo, V. “An advanced tool of the CNR IMAA EO facilities:
Overview of the TASI-600 hyperspectral thermal spectrometer”.
Workshop on Hyperspectral Image and Signal Processing,
Evolution in Remote Sensing, 2011.
[10] Amato, U., Antoniadis, A. and Gregoire, G. (2003).
Independent Component Discriminant Analysis. International
Journal of Mathematics, 3: 735-753.
[11] U. Amato, A. Antoniadis, M.F. Carfora, P. Colandrea, V.
Cuomo, M. Franzese, S. Pignatti, C. Serio (2013), Statistical
classification for assessing PRISMA hyperspectral potential for
agricultural land use, IEEE Journal of Selected Topics in Applied
Earth Observations and Remote Sensing [accepted in press]
[12] Johnson, B. R. and S. J. Young, 1998. In-Scene Atmospheric
Compensation: Application to SEBASS Data Collected at the
ARM Site. Technical Report, Space and Environment Technology
Center, The Aerospace Corporation, May 1998.
[13] Z.L. Li, F. Becker, M.P Stoll and Z. Wan. 1999. Evaluation of
six methods for extracting relative emissivity spectra from thermal
Figure 1. Example of TASI-600 orthorectified radiances (R-11230 infrared images. Remote Sensing of Environment, vol. 69, 197-214.
nm; G-9587 nm; B-8493 nm) and emissivities (R-8493 nm; G-9587
nm; B-11230 nm) imagery acquired over a rural area in Belgium
on June 2010 with the sensor mounted on the Partenavia P68
Observer airborne.

6. REFERENCES

[1] Goetz, A.F.H. Three decades of hyperspectral remote sensing


of the Earth: A personal view. Remote Sensing of Environment,
Volume 113, Supplement 1, September 2009, Pages S5-S16.
[2] Bassani, C., Cavalli, R.M., Goffredo, R., Palombo, A.,
Pascucci, S., Pignatti, S. Specific spectral bands for different land
cover contexts to improve the efficiency of remote sensing
archaeological prospection: The Arpi case study. Journal of
Cultural Heritage, 2009 10 (SUPPL. 1), pp. e41-e48.
[3] Pascucci, S., Bassani, C., Palombo, A., Poscolieri, M., Cavalli,
R. Road asphalt pavements analyzed by airborne thermal remote
sensing: Preliminary results of the Venice highway (2008) Sensors,
8 (2), pp. 1278-1296.
[4] Ribeiro da Luz, B., J.K. Crowley. Identification of plant species
by using high spatial and spectral resolution thermal infrared (8.0–
13.5 μm) imagery (2010) Remote Sensing of Environment, Volume
114, Issue 2, pp 404-413.
[5] Schlerf, M., Rock, G., Lagueux, P., Ronellenfitsch, F.,
Gerhards, M., Hoffmann, L., Udelhoven, T. A Hyperspectral
Thermal Infrared Imaging Instrument for Natural Resources
Applications. Remote Sensing. 2012; 4(12):3995-4009.
[6] Hook, S. J. Abbott, E.A. Grove, C. Kahle A.B and Palluconi,
F.D. 1999. Use of multispectral thermal infrared data in geological
studies. In Manual of Remote Sensing. V.3: Remote Sens. for the
Earth Sciences, Rencz A.N. Eds.; John Wiley & Sons Inc., 59-110.
[7] Murphya, S.W. R. Wrightb, C. Oppenheimerc, C.R. Souza
Filhoa. MODIS and ASTER synergy for characterizing thermal
volcanic activity. Remote Sensing of Environment. Volume 131, 15
April 2013, Pages 195–205.
[8] Sobrino, J.A.. Thermal remote sensing from Airborne
Hyperspectral Scanner data in the framework of the SPARC and
SEN2FLEX projects: an overview. Hydrol. Earth Syst. Sci., 13,
2031–2037, 2009.

You might also like