Professional Documents
Culture Documents
Proceedings of Spie: A Novel Framework For Change Detection in Bi-Temporal Polarimetric SAR Images
Proceedings of Spie: A Novel Framework For Change Detection in Bi-Temporal Polarimetric SAR Images
Proceedings of Spie: A Novel Framework For Change Detection in Bi-Temporal Polarimetric SAR Images
SPIEDigitalLibrary.org/conference-proceedings-of-spie
Davide Pirrone, Francesca Bovolo, Lorenzo Bruzzone, "A novel framework for
change detection in bi-temporal polarimetric SAR images," Proc. SPIE 10004,
Image and Signal Processing for Remote Sensing XXII, 100040Z (18 October
2016); doi: 10.1117/12.2241636
ABSTRACT
Last years have seen relevant increase of polarimetric Synthetic Aperture Radar (SAR) data availability, thanks to satellite
sensors like Sentinel-1 or ALOS-2 PALSAR-2. The augmented information lying in the additional polarimetric channels
represents a possibility for better discriminate different classes of changes in change detection (CD) applications. This
work aims at proposing a framework for CD in multi-temporal multi-polarization SAR data. The framework includes both
a tool for an effective visual representation of the change information and a method for extracting the multiple-change
information. Both components are designed to effectively handle the multi-dimensionality of polarimetric data. In the
novel representation, multi-temporal intensity SAR data are employed to compute a polarimetric log-ratio. The multi-
temporal information of the polarimetric log-ratio image is represented in a multi-dimensional features space, where
changes are highlighted in terms of magnitude and direction. This representation is employed to design a novel
unsupervised multi-class CD approach. This approach considers a sequential two-step analysis of the magnitude and the
direction information for separating non-changed and changed samples. The proposed approach has been validated on a
pair of Sentinel-1 data acquired before and after the flood in Tamil-Nadu in 2015. Preliminary results demonstrate that the
representation tool is effective and that the use of polarimetric SAR data is promising in multi-class change detection
applications.
Keywords: Synthetic Aperture Radar (SAR), Polarimetric SAR, Multi-temporal analysis, Change detection, automatic
approach, polarimetric change vector
1. INTRODUCTION
Spaceborne Synthetic Aperture Radar (SAR) systems are microwave active systems which represent a very important tool
in Earth Observation. This is because their imaging capabilities have low dependence on the cloud coverage and
illumination. Thus, SAR images are widely exploited in multi-temporal analysis for change detection applications such as
damage assessment, crop monitoring, etc. Most of the SAR systems acquire data with a single polarization channel,
transmitting and receiving signals with the same polarization (e.g., VV for vertical transmit and vertical receive). Among
SAR systems, Polarimetric ones (PolSAR) are able to acquire data with different combinations for the transmitted and
received wave polarization, enhancing the information extraction capabilities. In the recent years, data volume from
PolSAR sensors has gradually increased, because of both the new constellations with polarimetric sensors for continuous
global monitoring (i.e., Sentinel-1 mission) and the increased performance in terms of polarimetric channels and/or spatial
resolution (i.e., ALOS-2 PALSAR-2 and the future Cosmo-SkyMed Second Generation sensors). With respect to earlier
single-polarization data, the multi-polarization ones have enhanced information lying on the multiple polarimetric
channels, in terms of both amplitude and relative phase.
In the framework of the multi-temporal analysis of remote sensing images, change detection (CD) has a great relevance
because of the several civilian applications involving it (i.e., damage assessment, glaciers monitoring, crops monitoring).
In the recent years, the application capabilities have risen with the rising amount of data coming from both passive sensors
(i.e., multi-spectral and hyper-spectral) and active ones (i.e., SARs), and their capabilities in terms of spatial resolution
and number of channels (i.e., spectral channels for passive sensors, polarimetric channels for active ones). For this reason,
the hype on efficient strategies for unsupervised CD analysis is still high. In the literature, some of the approaches for CD
consider a direct comparison of corresponding pixels in multi-temporal images (i.e., pixel-based approaches)1-4, 7-9, 11-13,
while others exploit the spatial correlation of homogeneous regions for mitigating the noise effects and improving global
Image and Signal Processing for Remote Sensing XXII, edited by Lorenzo Bruzzone,
Francesca Bovolo, Proc. of SPIE Vol. 10004, 100040Z · © 2016 SPIE
CCC code: 0277-786X/16/$18 · doi: 10.1117/12.2241636
In the first step, a multi-temporal comparison of the two PolSAR images is computed by defining a multi-dimensional
polarimetric log-ratio image XLR.
𝑋𝐿𝑅 = [𝑋𝐿𝑅,𝑉𝑉 , 𝑋𝐿𝑅,𝑉𝐻 , 𝑋𝐿𝑅,𝐻𝐻 ] (2)
𝑋𝐿𝑅 has the same number of polarimetric channels as the input images. Each of them, namely 𝑋𝐿𝑅,𝑝𝑞 , is defined as the log-
ratio image of the corresponding polarimetric channels of the two multi-temporal polarimetric images (3).
𝑋2,𝑝𝑞
𝑋𝐿𝑅,𝑝𝑞 = log ( ) = log 𝑋2,𝑝𝑞 − log 𝑋1,𝑝𝑞 , 𝑝𝑞 ∈ [𝑉𝑉, 𝑉𝐻, 𝐻𝐻] (3)
𝑋1,𝑝𝑞
The use of the log-ratio mitigates the effect of the speckle noise on the multi-temporal comparison and reduces the
skewness for the different modes on the related histogram. Each sample on XLR can be represented as a Polarimetric Change
Vector (PCV) lying in a P-dimensional orthonormal feature space. This representation can be alternatively expressed in
Cartesian coordinates (i.e., the components 𝑋𝐿𝑅,𝑝𝑞 ) or in polar/spherical coordinates, by defining the magnitude 𝜌 and P-1
directions, namely 𝜗 and 𝜑, according to (4).
2 2 2
𝜌 = √𝑋𝐿𝑅,𝑉𝑉 + 𝑋𝐿𝑅,𝑉𝐻 + 𝑋𝐿𝑅,𝐻𝐻
𝑋𝐿𝑅,𝑉𝐻
𝜑 = tan−1 [ ] (4)
𝑋𝐿𝑅,𝑉𝑉
𝑋𝐿𝑅,𝐻𝐻
𝜃 = cos −1 [ ]
{ 𝜌
In the polar/spherical coordinate system, PCVs are bounded into a sphere S having radius 𝜌𝑚𝑎𝑥 , that corresponds to the
maximum magnitude of the PCVs for the considered image pair (Fig.2a).
𝑆 = {𝜌, 𝜗, 𝜑 : 0 ≤ 𝜌 ≤ 𝜌𝑚𝑎𝑥 , 0 ≤ 𝜗 ≤ 𝜋, 0 ≤ 𝜑 < 2𝜋} (5)
In S, PCVs related to unchanged pixels are expected to fall close to the origin, as the two multi-temporal images have
similar values on all the polarimetric channels and thus result in small 𝜌 values. These pixels tend to fall in an inner sphere
Cn centered in the origin (dark sphere in Fig.2b). The radius T of the sphere Cn is set based on minimum error decision rule
applied to the magnitude variable.
𝐶𝑛 = {𝜌, 𝜗, 𝜑 : 0 ≤ 𝜌 < 𝑇, 0 ≤ 𝜗 ≤ 𝜋, 0 ≤ 𝜑 < 2𝜋} (6)
On the other hand, PCVs associated to changed pixels are expected to be located in the spherical annulus Ac complementary
to Cn, defined as the region of S with magnitude greater than T (dark volume in Fig.3a).
𝐴𝑐 = {𝜌, 𝜗, 𝜑 : 𝜌 ≥ 𝑇, 0 ≤ 𝜗 ≤ 𝜋, 0 ≤ 𝜑 < 2𝜋} (7)
In Ac, the PCVs are clustered along preferred directions, based on the kind of change. Each of these directions can be
(a) (b)
Figure 2 – PCVs spherical domain S (a) and the inner sphere Cn containing PCVs of unchanged pixels (b).
(a) (b)
Figure 3 – Spherical annulus Ac containing PCVs for all the changed pixels (a) and sector Sk containing PCVs referring to
the k-th change (b).
The angular threshold pairs (𝜑𝑘1 , 𝜑𝑘2 ) and (𝜃𝑘1 , 𝜃𝑘2 ) of the sector 𝑆𝑘 are set based on minimum error decision rule applied
on the statistical distribution of classes along the direction variables. In the following section, an unsupervised multi-class
CD approach based on this framework is presented.
Masking
L
Change -class
densities estimation [ Detection
Figure 4 - Block scheme of the unsupervised CD approach based on the Polarimetric CVA framework.
In the first step of the algorithm, we model a two-class problem along the magnitude variable that aims at separating 𝛺𝑐
and 𝜔𝑛𝑐 . While 𝜔𝑛𝑐 is the only class associated to the non-change, 𝛺𝑐 is the union of all the K change classes 𝜔𝑐𝑘 :
𝐾
𝛺𝑐 = ⋃ 𝜔𝑐𝑘 (9)
𝑘=1
Thus, the distribution of the magnitude variable can be modeled as a mixture of the two components.
𝑝(𝜌) = 𝑃(𝛺𝑐 )𝑝(𝜌|𝛺𝑐 ) + 𝑃(𝜔𝑛𝑐 )𝑝(𝜌|𝜔𝑛𝑐 ) (10)
where 𝑃(𝛺𝑐 ), 𝑃(𝜔𝑛𝑐 )are the the prior probabilities for the classes 𝛺𝑐 and 𝜔𝑛𝑐 respectively and 𝑝(𝜌|𝛺𝑐 ), 𝑝(𝜌|𝜔𝑛𝑐 ) are the
conditional distributions for the same classes. An appoximation of 𝑝(𝜌) is given by the histogram of the magnitude image.
In the literature, different statistical distributions have been proposed for building mixture models for SAR data. Rayeigh-
Rice distributions have been widely used for modelling SAR single-date amplitude, while multi-temporal log-ratio has
been studied through mixtures of Gaussians or Generalized Gaussians. The magnitude of the Polarimetric Change Vector
is not easily characterizable in terms of SAR statistics and an exhaustive study of this is out of the scope of the paper.
Therefore, we approximate the conditional probability density function (pdf) of the magnitude variable as a mixture of
Generalized Gaussians, with parameters αi, βi, μi, as in (11).
𝛽
|𝜌−𝜇𝑖 | 𝑖
𝛽𝑖 −[ ]
𝑝(𝜌|𝜔𝑖 ) = 1
𝑒 𝛼𝑖
; 𝜔𝑖 ∈ {𝜔𝑛𝑐 , 𝛺𝑐 } (11)
2𝛼𝑖 𝛤 ( )
𝛽𝑖
The choice of the Generalized Gaussian mixture represents a good tradeoff between the model complexity (i.e., the number
of parameters regulating the distribution) and the likelihood overall performance.
Let us define a parameters vector Θ𝜌 containing both the unknown distribution parameters and the prior probabilities for
both 𝛺𝑐 and 𝜔𝑛𝑐 :
where 𝜌𝑧 is an element of 𝜌 . We can substitute the Generalized Gaussian model into (13). At the general (t+1)-th iteration,
the EM algorithm provide an estimation of the parameter vector Θ𝜌 of the single Generalized Gaussian in terms of the
same parameters estimated at the t-th iteration, according to (14). The estimation is obtained by maximizing 𝐿(Θ𝜌 |𝜌)
computed with the parameter vector at the t-th iteration.
(𝑡+1) ∑𝜌𝑧 ∈ 𝜌 𝑝(𝑡) (𝜔𝑖 |𝜌𝑧 )
𝑃 (𝜔𝑖 ) =
𝑀𝑁
1
(𝑡) (𝑡)
(𝑡) (𝑡) 𝛽𝑖 𝛽
𝑖
(𝑡+1) ∑𝜌𝑧 ∈ 𝜌 𝑝 (𝜔𝑖 |𝜌𝑧 )|𝜌𝑧 −𝜇𝑖 |
𝛼𝑖 =[ (𝑡)
]
∑𝑧𝜌𝑧 ∈ 𝜌 𝑝 (𝜔𝑖 |𝜌𝑧 )
The estimated parameters are employed in the detection step. Under the assumption of inter-pixel independence, we
consider Bayesian Minimum Error decision rule for assigning each pixel to the class that maximizes the posterior
probability:
𝜌 ∈ 𝜔𝑖 ⇔ 𝑃(𝜔𝑖 )𝑝(𝜌|𝜔𝑖 ) > 𝑃(𝜔𝑗 )𝑝(𝜌|𝜔𝑗 ); 𝑗 ≠ 𝑖; 𝜔𝑖 , 𝜔𝑗 ∈ {𝜔𝑛𝑐 , 𝛺𝑐 } (15)
The magnitude threshold T is the value of 𝜌 that makes the posterior probabilities for the two classes equal. Dashed circle
in Fig.5 represents the decision threshold along the magnitude separating the class (𝜔𝑛𝑐 ) (checked cluster) from the macro-
class 𝛺𝑐 (striped clusters).
'MGM
An accurate statistical model for this variable is complex and out of the goal of this work. Therefore, we consider the
conditional distribution for each 𝜔𝑐𝑘 being a Gaussian with parameters 𝑚𝑘 , 𝜎𝑘 .
2
(𝜑−𝑚𝑘 )
1 − 2
2𝜎𝑘
𝑝(𝜑|𝜔𝑐𝑘 ) = 𝑒 (17)
√2𝜋𝜎𝑘2
Similarly to what described for the magnitude, the EM algorithm is used for the estimation of the single contributions of
the mixture. In order to do that, a new parameter vector Θ𝜑 is defined as (18).
Θ𝜑 = {𝑃(𝜔𝑐1 ), ⋯ , 𝑃(𝜔𝑐𝑘 ), 𝜇𝑐1 , ⋯ , 𝜇𝑐𝑘 , 𝜎𝑐1 , ⋯ , 𝜎𝑐𝑘 } (18)
The log-likelihood function for the direction distribution is computed and maximized with respect to the vector Θ𝜑 , until
convergence. This provides the estimation of the parameters at the (t+1)-th iteration, as described in (10).
}
∑𝑐𝑎𝑟𝑑{𝛺
𝑧=1
𝑐
𝑝(𝑡) (𝜔𝑐𝑘 |𝜑𝑧 )
𝑃 (𝑡+1) (𝜔𝑐𝑘 ) =
𝑐𝑎𝑟𝑑{𝛺𝑐 }
}
(𝑡∗1) ∑𝑐𝑎𝑟𝑑{𝛺
𝑧=1
𝑐
𝑝 (𝑡) (𝜔𝑐𝑘 |𝜑𝑧 )𝜑𝑧
𝑚𝑐𝑘 = } ; ∀ 𝜑𝑧 |𝜌𝑧 > 𝑇 (19)
∑𝑐𝑎𝑟𝑑{𝛺
𝑧=1
𝑐
𝑝(𝑡) (𝜔𝑐𝑘 |𝜑𝑧 )
} (𝑡)
2 )(𝑡+1)
∑𝑐𝑎𝑟𝑑{𝛺
𝑧=1
𝑐
𝑝(𝑡) (𝜔𝑐𝑘 |𝜑𝑧 )(𝜑𝑧 − 𝑚𝑐𝑘 )2
(𝜎𝑐𝑘 = }
{ ∑𝑐𝑎𝑟𝑑{𝛺
𝑧=1
𝑐
𝑝(𝑡) (𝜔𝑐𝑘 |𝜑𝑧 )
Once the Gaussian distributions for the change classes are available, the Bayesian Minimum Error decision rule is
employed for assigning the pixel class, as it follows:
𝜑 ∈ 𝜔𝑐𝑘 ⇔ 𝑃(𝜔𝑐𝑘 )𝑝(𝜑|𝜔𝑐𝑘 ) > 𝑃(𝜔𝑐𝑗 )𝑝(𝜑|𝜔𝑐𝑗 ); ∀ 𝑗 ≠ 𝑘; 𝜔𝑐𝑗 , 𝜔𝑐𝑘 ∈ 𝛺𝑐 (20)
The angular thresholds can be derived from the decision rule by making the posterior probabilities equal for pairs of
adjacent classes, as described in (21). This provides the values for K thresholds.
𝑝(𝜔𝑐𝑗 )𝑃(𝜑|𝜔𝑐𝑗 ) = 𝑝(𝜑|𝜔𝑐𝑘 )𝑃(𝜔𝑐𝑘 ); 𝜔𝑐𝑗 , 𝜔𝑐𝑘 ∈ 𝛺𝑐 (21)
At the end of the Step 2, the different classes of change are separated and the multiple change detection problem solved,
as illustrated in Fig.6.
(a) (b)
(c) (d)
zo -
15
Dry out \
area
10 /.. ...
-0
o
-
ange
Flooded
-15
-20
-25
.
-15 -10
..
-5
area
0 5
W-channel Log-ratio (dB)
10 15 20
(e)
Figure 7 – Optical image of the test area (source: Google Earth©) (a); ground truth map (b); multi-temporal false color
composition for VV (c) and VH (d) channels. R: Oct. 2015, G: Nov. 2015; B: Oct. 2015; scatter-plot of the polarimetric
log-ratio components XLR,VV, XLR,VH (e).
0.25
T=10,02
0.2
0.15
4
4 0.1
0.05
-10 0 10 20 30p
(a)
(b) (c)
Figure 8 – Histogram of the magnitude variable and decision threshold T (a); binary CD Map (b); reference map (d)
In order to apply the proposed unsupervised CD strategy, magnitude and direction features have been derived from the
polarimetric log-ratio. In the first step, binary CD analysis was conducted on the magnitude feature. In our preliminary
experiment, magnitude threshold was tuned with optimal manual trial-and-error procedure. In this way, we limit the impact
of errors in Step 1 on the performance of the multiple change detection conducted in Step 2. The manual trial-and-error
threshold value was set to 10.02. set the value to T = 10.0 in order to proceed with the multi-class analysis of 𝛺𝑐 . Fig.8c
and Fig.8d show the related binary CD map and the reference map respectively. As we can see, the two maps look very
𝜔̂
𝑛𝑐
̂𝑐
𝛺
𝜔𝑛𝑐 109656 3209
𝛺𝑐 8997 6238
In the second step, multi-class CD was conducted in 𝛺𝑐 . The unchanged pixels were filtered out of the analysis. In the
multi-class CD, two change classes were considered, corresponding to the changes described above. The direction of the
selected pixels was thus modeled as a mixture model of K = 2 contributions, each of them modeled as Gaussian distributed.
The parameters of the mixture were estimated with EM algorithm. A summary of the parameters for the two mixture
components is reported in Table 2.
P μ σ2
𝜔𝑐1 0.7809 -149.3553 308.1824
𝜔𝑐2 0.2191 35.4502 137.4606
By applying the Bayesian decision rule for minimum error on the estimated distributions, the direction threshold based on
the Bayesian approach is set to 𝜑1 = -34.79°. It is worth noting that the direction variable is periodic on intervals
[2(𝑘 − 1)𝜋; 2𝑘𝜋], 𝑘 ∈ 𝑍, while the analysis considered the wrapped interval [0; 2𝜋]. The second threshold can be derived
from practical considerations, given by the periodicity of the direction variable. For a two-class problem, the second
angular threshold is the opposite direction with the first one, therefore 𝜑2 = -214.79°. Fig.9b shows the final multi-class
CD map. From a visual comparison with the reference map in Fig.7b, we clearly see that most of the changes are correctly
identified and separated.
(a) (b)
Figure 9 – Direction distribution its Gaussian mixture model (a); multi-class CD map (b).
From the numerical point of view, we have an Overall Accuracy of 90.43%. The confusion matrix, given in Table 3, shows
that very few pixels are misclassified between the two classes of changes. This confirms that the direction variable allows
to effectively separating among different kinds of change.
𝜔̂
𝑐2 𝜔̂
𝑛𝑐 𝜔̂
𝑐1
5. CONCLUSIONS
In this work, we proposed a novel framework for the visual representation of the change information from polarimetric
SAR images acquired at two different times. The framework is based on the definition of the polarimetric log-ratio. A
Polarimetric Change Vector has been introduced as a representation of the polarimetric log-ratio in a novel feature space.
Different loci associated to both no-change and multiple change classes have been described and an efficient representation
in polar/spherical coordinates introduced. Magnitude and direction features were derived in the log-ratio image domain.
Based on them, a novel automatic multi-class CD strategy has been presented, based on two steps. The first step is a binary
CD operating along the magnitude variable which detect the non-change class and a change macro-class. In the second
step, the macro-class is analyzed along the direction in order to discriminate multiple kinds of change. A preliminary
validation analysis of the CD strategy points out the effectiveness of both the proposed representation and the CD strategy.
As future developments, we aim at: i) associating a physical meaning to the backscattering variations on the different
polarimetric channels; performing a rigorous analysis of the statistical distribution model for the Polarimetric Change
Vectors; iii) exploiting additional polarimetric features for better discrimination of the change classes.
REFERENCES
[1] [1] Hussain, M., et al. "Change detection from remotely sensed images: From pixel-based to object-based
approaches.," ISPRS Journal of Photogrammetry and Remote Sensing, 80, 91-106 (2013).
[2] Bruzzone, L., and Prieto, D.F., "Automatic analysis of the difference image for unsupervised change detection." IEEE
Transactions on Geoscience and Remote sensing, 38(3), 1171-1182 (2000).
[3] Bovolo, F., and Bruzzone, L., "A theoretical framework for unsupervised change detection based on change vector
analysis in the polar domain." IEEE Transactions on Geoscience and Remote Sensing, 45(1), 218-236 (2007).
[4] Zanetti, M., Bovolo, F., and Bruzzone, L., "Rayleigh-Rice Mixture Parameter Estimation via EM Algorithm for
Change Detection in Multispectral Images." IEEE Transactions on Image Processing, 24(12), 5004-5016 (2015).
[5] Celik, T., "Unsupervised change detection in satellite images using principal component analysis and-means
clustering." IEEE Geoscience and Remote Sensing Letters, 6(4), 772-776 (2009).
[6] Bovolo, F., and Bruzzone, L., "A detail-preserving scale-driven approach to change detection in multitemporal SAR
images." IEEE Transactions on Geoscience and Remote Sensing, 43(12), 2963-2972 (2005).
[7] Bazi, Y., Bruzzone,L., and Melgani, F., "An unsupervised approach based on the generalized Gaussian model to
automatic change detection in multitemporal SAR images." IEEE Transactions on Geoscience and Remote
Sensing, 43(4), 874-887 (2005).
[8] Aiazzi, B., et al. "Nonparametric change detection in multitemporal SAR images based on mean-shift
clustering." IEEE Transactions on Geoscience and Remote Sensing, 51(4), 2022-2031 (2013).
[9] Carincotte, C., Derrode, S., and Bourennane, S., "Unsupervised change detection on SAR images using fuzzy hidden
Markov chains." IEEE Transactions on Geoscience and Remote Sensing, 44(2), 432-441 (2006).
[10] Inglada, J., and Mercier, G., "A new statistical similarity measure for change detection in multitemporal SAR images
and its extension to multiscale change analysis." IEEE transactions on geoscience and remote sensing, 45(5), 1432-
1445 (2007).
[11] Kasetkasem, T., and Varshney, P. K., "An image change detection algorithm based on Markov random field
models." IEEE Transactions on Geoscience and Remote Sensing, 40(8), 1815-1823 (2002).
[12] Ban, Y., and Yousif, O. A., "Multitemporal spaceborne SAR data for urban change detection in China." IEEE Journal
of Selected Topics in Applied Earth Observations and Remote Sensing 5(4), 1087-1094 (2012).