IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO.

6, NOVEMBER 2012 1051

Probabilistic Graphical Models for Flood State

Detection of Roads Combining Imagery and DEM
Daniel Frey, Matthias Butenuth, and Daniel Straub

Abstract—A new system for estimating the state of roads during flood state of infrastructure (referred to as DEM simulation).
flooding based on probabilistic graphical models is presented. However, simulations based only on DEM to detect the flood-
The location of the roads is given by a geographic information plain lead to correct assignments between only 60% and 80%
system, whereas the up-to-date information for the assessment
depending on the resolution and the model [2]. The goal of
of flood state is delivered by remote sensing data. Furthermore,
the height information from a digital elevation model (DEM) is this letter is the development of a system which combines the
combined with image data to improve the accuracy of the results. approaches by exploiting the information of DEM at locations
The presented system is based on factor graphs, which are used to where imagery does not provide accurate information due to
model the statistical dependence between random variables. Three occlusions.
different models are presented: a 1-D pixel-based model, a 2-D The image analysis method has to fulfill several require-
topology-based model considering the dependences of neighboring ments. First, the method has to comply with high demands on
pixels, and a 3-D multitemporal-based model, which can deal
with sequential remote sensing imagery at several points in time.
the semantic correctness. Second, the computation time has to
The proposed models are compared to a flood simulation based be near real time. Third, the method has to be flexible with
only on the DEM and a maximum likelihood classification based respect to input data. In this letter, we present a method to
only on the image data. A numerical evaluation demonstrates the detect the flood state of roads based on probabilistic graphical
improved performance of the three proposed models. models, which fulfills these requirements and facilitates the
Index Terms—Bayesian network (BN), detection, factor graph, combination of remote sensing data with height information
flooding, probabilistic graphical model. from DEM.
In Section II, existing assessment systems and their
methodologies are discussed. Section III describes the ba-
I. I NTRODUCTION sics of different types of probabilistic graphical models.
The graphical model proposed in this letter is presented in
N ATURAL disasters are responsible for more than
2 275 000 fatalities worldwide between 1980 and 2010
[1]. One possibility to reduce the amount of fatalities and
Section IV. Step by step, the graphical model is built up from
a pixel-based approach to a multitemporal approach incorpo-
economical damages is to provide up-to-date information about rating the topology of neighboring pixels. Finally, results are
the caused infrastructure damage to support rescue teams. Up- shown for the assessment of roads during flooding, and an
to-date crisis information derived from remote sensing data of evaluation is conducted comparing quantitatively the proposed
the affected areas is available in near real time and is widely graphical models with a flood simulation based only on the
used to support rescue teams in emergency actions. DEM (DEM simulation) and a maximum likelihood classifi-
The manual interpretation of imagery during natural disasters cation based only on the image data (image classification).
is difficult and, most importantly, time consuming. There-
fore, near-real-time automatic or semiautomatic image analysis
II. DAMAGE A SSESSMENT S YSTEMS
methods have to be developed, which enable the interpretation
of infrastructure objects that are relevant for rescue teams. The rapid assessment of damages during natural disasters
Derived floodplains from remote sensing images demonstrate is of crucial importance to support the disaster management.
good performance if the flooded areas are not occluded by The increasing amount of remote sensing data from different
vegetation, clouds, or shadows (in the remainder, we refer to sensors and the development of methods exploiting the image
this approach as image classification). In case of flooding, also information automatically lead to a large variety of different
the digital elevation model (DEM) can be used to detect the damage assessment systems. Both optical and radar data can be
used to infer the flooded areas. Although radar data have the
strong advantage of cloud penetration, the results are limited
in urban areas due to the effects of shadowing, layover, and
Manuscript received December 12, 2011; revised January 25, 2012 and foreshortening [3]. Aside from using optical and radar sensors,
February 15, 2012; accepted February 17, 2012. This work was supported additional information can be used to improve the assessment.
by the International Graduate School of Science and Engineering, Technische
Universität München.
In case of flooding, DEMs can deliver additional information
D. Frey and M. Butenuth are with the Chair of Remote Sensing Tech- that leads to a better performance in detecting flooded areas.
nology, Technische Universität München, 80333 München, Germany (e-mail: Wang et al. [4] generate a flood mask of each individual
[email protected]; [email protected]). data source, the final flood mask consisting of the set union.
D. Straub is with the Engineering Risk Analysis Group, Technische Univer-
sität München, 80333 München, Germany (e-mail: [email protected]). Brivio et al. [5] combine the extracted flood mask with the least
Digital Object Identifier 10.1109/LGRS.2012.2188881 accumulative cost distance matrix generated from the DEM.

1545-598X/$31.00 © 2012 IEEE

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on February 28,2024 at 09:22:12 UTC from IEEE Xplore. Restrictions apply.
 1052 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 6, NOVEMBER 2012

Due to the increasing number of acquired images in short

time frames, more and more methods analyze the dynamics
of floods [6]. A multitemporal and multisensorial approach
for the damage assessment of linear infrastructure objects is
shown in Butenuth et al. [7]. However, the fusion of the
different data sources is carried out after processing the input
information individually and not within a consistent statistical
framework.
A general statistical framework used for damage assessment
during natural disasters is probabilistic graphical models [8].
Fig. 1. (Above) One-dimensional pixel-based BN and (below) corresponding
Straub [9] presents a rockfall hazard rating system based on factor graph with (circles) variable nodes and (rectangles) factor nodes.
Bayesian networks (BNs). An area-wide risk assessment sys-
tem for avalanches linking BNs to a geographic information tion from remote sensing data with a DEM. Second, the 2-D
system (GIS) is presented by Grêt-Regamey and Straub [10]. topology-based model takes the statistical dependence among
These approaches incorporate the experience of previous dis- neighboring pixels into account. Third, the time domain is
asters but do not consider observations representing the current modeled in the 3-D multitemporal-based model. Although the
situation. Martinis et al. [11] use Markov random fields (MRFs) following section focuses on the assessment of roads during
to detect flooded areas in synthetic aperture radar data. flooding, the model is applicable to all kinds of linear infras-
In this letter, a probabilistic graphical model is presented, tructure objects, such as railroads or pipelines.
which assesses roads during flooding. Its three main charac-
teristics are as follows. First, the model is not restricted to a A. One-Dimensional Pixel-Based Model
specific sensor and, in particular, enables to include the DEM
Fig. 1 shows the proposed BN, which combines the image
into the analysis. Thus, it is directly transferable to various
information with height information from the DEM. The BN
scenarios. Second, the model is embedded into a consistent
is designed as a causal network. Therefore, the arrows in the
statistical framework which quantifies the associated uncer-
graphical model in Fig. 1 describe causal relations between
tainty. Third, the model is a multitemporal approach exploiting
random variables. The variable H represents the elevation
the information from previous imagery and monitoring the
at a specific location and is modeled as a normal distribu-
dynamic of flooding.
tion p(H) ∼ N (h|μDEM , σDEM 2
), where μDEM is the observed
2
value given from the DEM and σDEM is the variance, which
III. P ROBABILISTIC G RAPHICAL M ODELS can be derived from the accuracy of the DEM in the z-direction.
The water gauge G describes the elevation of the water surface
Probabilistic graphical models combine probability theory with the same reference as μDEM . It is modeled as a normal
and graph theory [8]. A graph-based representation is used distribution p(G) ∼ N (g|μg , σg2 ). The distribution p(G) can be
as the basis for compactly encoding the dependence structure derived by intersecting the extracted floodplain with the DEM.
of a complex distribution. In general, probabilistic graphical The height values at the borders between flooded and non-
models can be divided into two different types: directed and flooded areas are used to calculate μg and σg2 . For investigations
undirected graphical models. Both types consist of nodes, in small areas with low flow dynamics, the same distribution
which correspond to random variables, and links, which p(G) can be used as a rough approximation for the whole
represent dependence among the random variables. The most scene. For larger regions, the distribution p(G) is obtained by
established directed graphical models are causal BNs, in which including only height values smaller than a fixed distance. The
the causal links are represented by conditional probabilities. distance is used as a weighting factor. Several interpolation
BNs show good performance through combining different methods are proposed in the literature, extracting the water
types of data. Theoretically well-founded probabilities, as well gauge via a fusion of the DEM and the extracted floodplain
as subjective estimates, can be used for the same network. based on cross sections, linear regression, polynomial regres-
The second major class of probabilistic graphical models is sion, or composed bicubic splines [12]. In this letter, the same
undirected graphical models, also known as Markov Random distribution p(G) is used for the whole area, but interpolation
Fields (MRFs), which are widely used in image processing. methods will be investigated in future work. Furthermore, the
MRF theory provides a convenient and consistent way of estimation of p(G) is highly dependent on the accuracy of
modeling context-dependent entities such as image pixels. the DEM. Instead of the estimation of p(G) by intersecting
Every directed and undirected graph can be converted to factor the DEM with the extracted floodplain, in situ measurements
graphs, which are a superclass of MRF and BN. Factor graphs can be used if available. The water level W can be calculated
consist of two types of nodes: variable nodes corresponding to deterministically from H and G as W = G − H. Therefore,
the random variables and factor nodes describing the statistical W has the normal distribution p(W |μg , μDEM ) ∼ N (w|μw =
relation between the random variables. μg − μDEM , σw 2 2
= σDEM + σg2 ). The variable g corresponds to
the n-dimensional gray value vector obtained from the imagery
(n: number of bands). The water level W and the damage
IV. G RAPHICAL M ODEL FOR F LOOD S TATE D ETECTION
D are hypothesis variables that we want to estimate. The
In the following sections, the probabilistic graphical model random variable damage D is discrete and consists of only
for flood state detection is built up step by step. First, a 1-D two states: trafficable and flooded. This is a simplification
pixel-based model is presented, which combines the observa- since slightly flooded roads could be still trafficable. Since
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on February 28,2024 at 09:22:12 UTC from IEEE Xplore. Restrictions apply.
 FREY et al.: PROBABILISTIC GRAPHICAL MODELS FOR FLOOD STATE DETECTION OF ROADS 1053

TABLE I
C ONDITIONAL P ROBABILITY TABLE p(C|D)

the damage state of a road often cannot be directly observed

from the imagery due to occlusions by vegetation or clouds,
the intermediate variable class C is introduced. The variable
class consists of the four possible states: road r, water w,
cloud c, and vegetation v. The causal relations in the BN
Fig. 2. Factor graph of 2-D topology-based model. Numbered messages refer
are modeled with conditional probabilities. The conditional to the numbering of the equations.
probability p(D = f looded|W ) is the Heaviside step function
with argument W . The complementary conditional probabil- ideally suited for describing equivalent interactions such as
ity p(D = traf f icable|W ) is p(D = traf f icable|W ) = 1 − the interdependences of neighboring pixels. The modeling of
p(D = f looded|W ). The probability p(C|D) is given by the symmetrical interactions between random variables is usually
probability table in Table I, where p(C = c) and p(C = v) are done by MRFs. Factor graphs, a superclass of BN and MRF,
the prior probabilities of the existing cloud coverage and the are used to map the causal relations and the interdependences
vegetation occurring in the imagery. Finally, the conditional between neighboring pixels into a single graphical model. In
probability p(g |C) is given as an n-dimensional probability Fig. 1, the 1-D pixel-based model is converted into a factor
function generated by mixtures of Gaussian distributions. For graph. The factors correspond to the conditional probabilities,
a given class C, this probability function is learned from a set e.g., fCg = p(g |C). The extension of this factor graph to
of K training areas. The training areas are selected manually include the interactions of neighboring pixels is described in
from the current imagery. Exemplarily, the probability function the following section.
for C = r is

K
B. Two-Dimensional Topology-Based Model
p(g |C = r) = πr,k N (g |
μr,k , Σr,k ). (1)
k=1 The 2-D topology-based model accounts for dependence
The number K corresponds to the number of training sam- between neighboring pixels due to correlation in the water
ples for the class road, which should cover all different appear- gauge G. In Fig. 2, a detail of the factor graph of the 2-D
ances of roads in the image [7]. The mean values μ  r,k and the topology-based model is shown. The neighborhood is modeled
covariance matrix Σr,k are learned for each training area. The as links between the random variables G. These can be inter-
mixing coefficients πr,k have to fulfill the following equations: preted as representing the smoothness of the water surface and
are modeled via the factor fGi Gj (Gi , Gj ) describing a normal

K
distribution. Since the factor graph represents a tree structure,
πr,k = 1 0 ≤ πr,k ≤ 1. (2)
exact inference is possible. The calculation of marginal
k=1
probabilities in a factor graph can be carried out using the
After determining all conditional probabilities, the marginal sum–product rule. In general, the procedure of the sum–product
probabilities p(W |μg , μDEM , g ) and p(D|μg , μDEM , g ) can be rule can be divided into four steps: 1) selection of a root node;
calculated. The formulas to compute the marginals can be 2) forward message passing; 3) backward message passing; and
directly inferred from the graphical structure in Fig. 1. For ease 4) calculation of marginal probability. The root can be selected
of notation, let U = {μg , μDEM , g } be the set of all observed arbitrarily; in Fig. 2, the node Gn is selected as the root node.
variables; then, the marginal probabilities are The next step is the propagation of the information from all
1  
p(W |U ) = p(W |μg , μDEM ) p(D|W ) p(C|D)p(g |C) leave nodes to the root node. This propagation of information
Z is done via message passing. The message passing in the
D C
(3) direction to the root node is called forward message passing.
1  Once all messages arrived at the root node, the information
p(D|U ) = p(W |μg , μDEM )p(D|W ) p(C|D)p(g |C) propagates back from the root to all leave nodes, which is called
Z
W C backward message passing. After calculating all messages, the
(4) marginal probability for a random variable can be calculated
where Z is the normalizing factor given by by multiplying all incoming messages at its node. In general,
  the propagated messages can be divided into two distinct kinds,
Z= p(W |μg , μDEM )p(D|W ) p(C|D)p(g |C). (5) namely, those going from a factor node f to a variable node X
D W C denoted as μf →X (X) and those going from a variable node to
Applying (3)–(5) for every pixel, the random variables W a factor node denoted as μX→f (X). The message μX→f (X)
and D can be estimated. However, in the 1-D pixel-based from a variable node to a factor node is simply the product
model, dependences among neighboring pixels are neglected, of all incoming messages at the variable node. The message
e.g., the fact that water gauges at two neighboring pixels are μf →X (X) from a factor node to a variable node is the product
highly correlated. BNs can model causal relations but are not of all incoming messages into the factor node multiplied by the
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on February 28,2024 at 09:22:12 UTC from IEEE Xplore. Restrictions apply.
 1054 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 6, NOVEMBER 2012

factor associated with the factor node and then marginalized

over all of the variables associated with the incoming messages.
Exemplarily, some messages for the 2-D topology-based
model are shown. The numbers in Fig. 2 correspond to the
numbering of the formulas. The forward message from factor
node fW D to the variable node Wj is
 
μfW D →Wj (Wj ) = p(Dj |Wj ) p(Cj |Dj )p(gj |Cj )
D Cj
j

= p(Dj |Wj ) · μDj →fW D (Dj ) (6)

Dj

wherein μDj →fW D (Dj ) is the forward message from Dj to

fW D . The forward message from the variable node Gj to the
factor node fGj Gj+1 is
Fig. 3. Three cases of different cloud coverages and original Namibia scenario
μGj →fGj Gj+1 (Gj ) = μfGi Gj →Gj (Gj ) · μfHGW →Gj (Gj ). (7) (near-infrared).

The reference was generated manually from the original scene

Both forward messages are shown as arrows in the direction
shown in Fig. 3 (original case) and by means of the GIS layer.
to the root node in Fig. 2. Equally, the backward massages
However, the focus of the proposed models is the assessment
are calculated. Equation (8) shows exemplarily the backward
of roads even when the roads are partly occluded by vegetation
massage μGj →fGi Gj (Gj ) from the variable node Gj to the
or clouds by combining the imagery with the DEM. For that
factor node fGi Gj reason, artificial clouds with the shape of ellipses are introduced
to investigate the performance of the different models
μGj →fGi Gj (Gj ) = μfGi Gj+1 →Gj (Gj ) · μfHGW →Gj (Gj ). (8)
[see Fig. 3 (cases 1–3)]. Even if no information from the
imagery is given, the roads occluded by the artificial clouds can
Finally, the unnormalized marginal probabilities p̃(Wj ) and
be assessed using the information from the DEM. In the test
p̃(Dj ) can be computed by multiplying all incoming messages
scenario, an ASTER DEM with an averaged standard deviation
to the variable nodes
σDEM = 5.88 m of elevation and a spatial resolution of 15 m. In
p̃(Wj ) = μfHGW →Wj (Wj ) · μfW D →Wj (Wj ) the 3-D multitemporal-based model, an additional TerraSAR-X
p̃(Dj ) = μfW D →Dj (Dj ) · μfDC →Dj (Dj ). (9) image in SpotLight mode, acquired on April 6th with a pixel
spacing of 1.75 m, is used. The 3-D multitemporal-based model
is evaluated at the time of acquiring the RapidEye image. The
C. Three-Dimensional Multitemporal-Based Model processing time of the algorithm for the described scenario is
approximately 60 s on an AMD Athlon X2 Dual Core computer
The 1-D pixel-based and 2-D topology-based models con- with 2.21 GHz and therefore fulfills the requirement for near-
sider only one image at a specific point in time. However, real-time applicability.
imagery acquired at a previous time T − t can add further in- The 1-D pixel-based model (Fig. 1), the 2-D topology-based
formation, particularly if the imagery at the current time T has a model (Fig. 2), and the 3-D multitemporal-based model are
lower spatial or spectral resolution than that at the previous time compared by means of the receiver operating characteristics
or if the image at time T is partly occluded by clouds. The only (ROCs). The computation of the ROC curves is obtained via
difference to the 2-D topology-based model is the integration a threshold, which corresponds to the probability that has to
of one additional link with one factor between the successive be exceeded to assign a road pixel to flooded. By shifting
points in time. This factor fGj ,T −t,T links the random variables this threshold from 0% (assigns all road pixels to flooded) to
Gj,(T −t) and Gj,T which are the water gauges at times T − t 100% (assigns that no road pixel is flooded), the true positive
and T , respectively. The factor fGj ,T −t,T describes the changes rate (TPR) and the false positive rate (FPR) for each threshold
of water gauge between the points in time. Inference is again can be calculated. In addition, the results of the proposed
performed using the sum–product algorithm in the factor graph probabilistic graphical models are compared on the one hand
as shown in the 2-D topology-based model. with a flood simulation using only the DEM and on the other
hand with a pixel-based multispectral maximum likelihood
classification using only the imagery. The ROC curves for all
V. R ESULTS
case scenarios are plotted in Fig. 4. The ROC curves show the
The presented models are evaluated at a test scenario in better performance of the proposed models compared to those
Namibia, a flooding in March 2009 at the Chobe River. The of the simulation and classification. The simulation is based
used remote sensing data are a RapidEye image, acquired on on a simple superposition of the DEM with the water plane
April 8 with a resolution of 6.5 m [see Fig. 3 (original case)]. generated from water gauge estimations. The simple model and
The RapidEye image consists of five bands (blue, green, red, the large standard deviation σDEM of the ASTER DEM result
red edge, and near infrared), which are all used as features in a poor performance of the simulation. However, in case of
for the classification. The evaluation compares the flooded emergency actions, often, no other data are available. In the
roads detected using the proposed models with a reference. original case, the maximum likelihood classification has nearly
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on February 28,2024 at 09:22:12 UTC from IEEE Xplore. Restrictions apply.
 FREY et al.: PROBABILISTIC GRAPHICAL MODELS FOR FLOOD STATE DETECTION OF ROADS 1055

Fig. 5. Result of the 3-D multitemporal-based model for case 1. (Green) Not
flooded. (Yellow) Possibly flooded. (Red) Flooded.

roads during flooding. The graphical model forms a statistical

framework that consistently combines imagery and DEM. The
evaluation of the three presented models shows the benefit of
Fig. 4. Evaluation: ROCs of (red dotted/dashed) 1-D pixel-based model, (blue modeling dependence among neighboring pixels and embed-
dashed) 2-D topology-based model, (green continuous) 3-D multitemporal-
based model, (black continuous) the image classification, and (black dotted)
ding information from previous points in time. In future work,
the simulation based on the DEM. (TPR) True positive rate. (FPR) False additional dependence among random variables should be in-
positive rate. vestigated. Furthermore, the simple 1-D hydraulic approach
estimating the water gauge G could be substituted by more
the same performance as the 1-D pixel-based approach. This sophisticated interpolation methods. Furthermore, additional
is reasonable since, in the original case, no occlusions exist information can be exploited in the 3-D multitemporal-based
and, therefore, the DEM adds little information. However, in model, e.g., the development of the flood, the time difference
cases 1–3, the better performance of the proposed models is of the acquired images, and information about the employed
evident. The influence of DEM information is shown by the sensors.
comparison of the 1-D pixel-based model, which uses image
and DEM information (see Fig. 4, red dotted), with the image R EFERENCES
classification without using DEM information (see Fig. 4, black
[1] MunichRe, “Natural catastrophes worldwide 1980–2010,” Geo Risk Re-
line). In the original case scenario, the performances of the search, NatCatService, 2011.
2-D topology-based model and the 3-D multitemporal-based [2] P. Bates and A. De Roo, “A simple raster-based model for flood inundation
model are similar and both better than that of the 1-D pixel- simulation,” J. Hydrol., vol. 236, no. 1/2, pp. 54–77, Sep. 2000.
[3] D. Mason, R. Speck, B. Devereux, J. Guy, J. Neal, and P. Bates, “Flood
based model. The similar performance of the 2-D topology- detection in urban areas using TerraSAR-X,” IEEE Trans. Geosci. Remote
based model and the 3-D multitemporal-based model is obvious Sens., vol. 48, no. 2, pp. 882–894, Feb. 2010.
since the information of the image at a previous point in time [4] Y. Wang, J. Colby, and K. Mulcahy, “An efficient method for mapping
T − t is limited if no occlusions at time T exist. In all other flood extent in a coastal floodplain using Landsat TM and DEM data,”
Int. J. Remote Sens., vol. 23, no. 18, pp. 3681–3696, 2002.
case scenarios with artificial clouds, the 3-D multitemporal- [5] P. Brivio, R. Colombo, M. Maggi, and R. Tomasoni, “Integration of
based model outperforms the 2-D topology-based model which remote sensing data and GIS for accurate mapping of flooded areas,” Int.
is better than the 1-D pixel-based model. In particular, in case 2, J. Remote Sens., vol. 23, no. 3, pp. 429–441, 2002.
the 3-D multitemporal-based model leads to better results, since [6] G. Schumann, J. Neal, D. Mason, and P. Bates, “The accuracy of sequen-
tial aerial photography and SAR data for observing urban flood dynamics,
the artificial clouds occlude a lot of areas near to the water line, a case study of the UK summer 2007 floods,” Remote Sens. Environ.,
which are difficult to assess. vol. 115, no. 10, pp. 2536–2546, Oct. 2011.
Finally, in Fig. 5, the assessed roads for the 3-D [7] M. Butenuth, D. Frey, A. A. Nielsen, and H. Skriver, “Assessment system
of infrastructure objects using multi-sensorial and multi-temporal imagery
multitemporal-based model are shown for case 1. The roads for disaster management,” Int. J. Remote Sens., vol. 32, no. 23, pp. 8575–
are divided into three categories: flooded roads (red), trafficable 8594, 2011.
roads (green), and roads which are not assigned to flooded [8] D. Koller and N. Friedman, Probabilistic Graphical Models: Principles
or trafficable due to a marginal probability p̃(Dj ) less than and Techniques. Cambridge, MA: MIT Press, 2009.
[9] D. Straub, “Natural hazards risk assessment using Bayesian networks,” in
a threshold pt (yellow). The threshold describes the balance Proc. 9th ICOSSAR, 2005, pp. 2535–2542.
between wrong assignments and no assignments and was man- [10] A. Grêt-Regamey and D. Straub, “Spatially explicit avalanche risk assess-
ually chosen. In future work, the estimation of an optimal ment linking Bayesian networks to a GIS,” Nat. Hazards Earth Syst. Sci.,
threshold will be considered. vol. 6, no. 6, pp. 911–926, Oct. 2006.
[11] S. Martinis, A. Twele, and S. Voigt, “Unsupervised extraction of flood-
induced backscatter changes in SAR data using Markov image modeling
on irregular graphs,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 1,
VI. C ONCLUSION pp. 251–263, Jan. 2011.
[12] M. Gianinetto and P. Villa, “Rapid response flood assessment using min-
This letter has presented a probabilistic graphical model imum noise fraction and composed spline interpolation,” IEEE Trans.
based on factor graphs, which assesses the damage state of Geosci. Remote Sens., vol. 45, no. 10, pp. 3204–3211, Oct. 2007.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on February 28,2024 at 09:22:12 UTC from IEEE Xplore. Restrictions apply.