- Email: [email protected]
Evaluation of SAR speckle filter technique for inundation mapping
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Contents lists available at ScienceDirect
Remote Sensing Applications: Society and Environment journal homepage: http://www.elsevier.com/locate/rsase
Evaluation of SAR speckle filter technique for inundation mapping Vikas Kumar Rana *, T.M.V Suryanarayana Water Resources Engineering and Management Institute, Faculty of Technology & Engineering, The Maharaja Sayajirao University of Baroda, Gujarat, India
A R T I C L E I N F O
A B S T R A C T
Keywords: Inundation SAR Machine learning Sentinel-1
The presence of speckle in visual images makes the automated digital image classification a challenging task. Therefore, reduction of speckles is an important pre-processing step. The choice of speckle filter depends on the requirements of an application and the characteristics of the dataset. In this study, some most preferred speckle filters are assessed for the data from Sentinel-1 to map flood extent. The Sentinel-1 (VV-vertical transmit, vertical receive and VH- vertical transmit, horizontal receive) polarizing filter data were used to evaluate machine learning algorithms, namely, random forest (RF) and support vector machine (SVM), to classify an inundated area. The accuracies of the classifications were assessed by kappa coefficient, overall accuracies, and producer’s and user’s accuracies. The present study suggests an approach to monitor damage and provide basic information to help local communities manage water-related risk, land planning, water management, and flood control programs.
1. Introduction In recent years, severe rainfall events have afflicted the state of Kerala in southern India causing damage to houses and infrastructures. Remotely sensed data can provide significant mapping capabilities during such severe rainfall events. However, obtaining remotely sensed data with an ideal combination of fine spatial and temporal resolution with the ability to see through clouds and discriminate flooding under forest cover is a difficult task. The extent of inundation, caused by river flooding and/or coastal storm surges, is required quickly to expedite relief and repair services. The precipitation over Kerala amid June, July and August (1–19 August 2018) were 15%, 18% and 164% above normal, respectively. Due to intense rainfall, all the major reservoirs were full by the end of July 2018 and had no buffer storage to accom modate the inflows from 10th of August 2018 (Central Water Commis sion, 2018). Serious spell of precipitation from the 14 August 2018 to 19 August 2018 brought appalling flood in 13 out of 14 districts. The perpetuated exceptional rainfall in August (170% above normal) in the catchment areas compelled the authorities to resort to hefty downstream release into the rivers (India Meteorological Department, 2018). Remote sensing promises exceptional capacity in catastrophe control owing to its regular acquisition function over a large spatial extent (Serpico et al., 2012; Nirupama and Simonovic, 2007; Gitas et al., 2008; Khan, 2005). Flooding is a complex phenomenon due to its heteroge neity and spectral diversity. The analysis of flood mapping require high
spatial and temporal resolution images to track the rapidly retreating flood process (Zhang et al., 2014). Drastic variability in climate has accelerated the incidence of cata strophic flood events in the last decade (Chunming et al., 2005). In any flood-related study, identification of the flood extent and susceptible areas is a prerequisite to assess the disaster impact. Flood mapping can best be achieved with the help of remote sensing due to the inaccessi bility to the flood-affected regions. However, cloudy conditions reduce flood mapping accuracy below the acceptable levels in optical remote sensing. Synthetic Aperture Radar (SAR) imaging is an efficient remote sensing technique offering well-developed, consistent, efficient, and reliable means of collecting information to extract earth’s surface dielectric properties (Lee and Pottier, 2009). The ability of SAR to penetrate clouds is extremely useful in flood-related studies. Synthetic aperture radar uses microwave radiation to illuminate the earth’s sur face for recording the amplitude and phase of the back-scattered radi ation, which makes the imaging process coherent. The active sensor of Sentinel-1 forms a SAR image by coherently processing the returning signals from successive radar pulses. Stronger or weaker final signals (output) are generated by the out-of-the-phase waves by constructively or destructively interfering with each other. These interferences produce a seemingly random pattern of brighter and darker pixels giving the radar images a distinctly grainy appearance known as ‘Speckle’ (Goodman, 1976; Lee et al., 1994). Speckle noise changes the spatial
* Corresponding author. WREMI, The Maharaja Sayajirao University of Baroda, India. E-mail addresses: [email protected] (V.K. Rana), [email protected] (T.M.V Suryanarayana). https://doi.org/10.1016/j.rsase.2019.100271 Received 16 June 2019; Received in revised form 3 October 2019; Accepted 22 October 2019 Available online 22 October 2019 2352-9385/© 2019 Elsevier B.V. All rights reserved.
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
retain the polarimetric characteristics of certain pixels. However, it over filters the point targets, creates a combination of heterogeneous pixels and degrades the overall spatial information. The filter is simple and fast, however, it is not isotropic (i.e. circularly symmetric), but smooths further along diagonals than along rows and columns. Also, disconti nuities are found in the smoothed image due to an abrupt cut-off of weights rather than decline gradually to null. 1.1.2. Gamma map filter This filter is based on the Bayesian analysis of image statistics. The scene reflectivity of the underlying image in Gamma-Map algorithm is assumed to be Gamma distributed rather than normally distributed, and speckle is noise within it. Thus this filter works best for geospatial im ages containing homogenous areas such as oceans, forests, fields, etc. (Lopes et al., 1990). It is given by following cubic equation (Frost et al., 1982).
Fig. 1. Selected Homogeneous area, linear feature and Edge.
statistics of the underlying scene backscatter making the classification of imageries a difficult task (Durand et al., 1987). A brief introduction of some well-known despeckling methods is presented below.
bI 3
1.1. DE-NOISING methods
2 IbI þ σðbI
DNÞ ¼ 0
bI ¼ required value
The presence of speckle is the major challenge in the SAR image processing. A speckle reduces the resolution of an image and the detectability of the ground targets. It also distorts the spatial patterns of surface characteristics and reduces the accuracy of image classification (Wang and Ge, 2012). Speckles are signal-dependent and, therefore, act like multiplicative noise (LeeSen, 1981).
I ¼ local mean DN ¼ input value σ ¼ original image variance Gamma-Map approach has several advantages compared to the other filters, as it can simultaneously take into account realistic first and second order statistical models for both speckle and underlying scene reflectivity, and combine them through Bayesian inference. Thus this filter works best for geospatial images containing homogenous areas such as oceans, forests and fields.
1.1.1. Boxcar filter A simple averaging filter that replaces the center pixel in a 3 � 3 or a larger sized moving kernel with the mean value of kernel pixels. It has good performance in reducing speckles in a homogeneous area; how ever, it degrades spatial resolution due to indiscriminately averaging pixels from the inhomogeneous area and destroys the polarimetric properties (Lee and Pottier, 2009). This easy operation can very well
Fig. 2. Study area Kerala. 2
– 7624 – Descending 22-Aug18 INS-NOBS
Fig. 3. Study area Asssam
1.1.3. Frost filter This filter uses local image statistics to remove high-frequency noise (speckles) while preserving features (edges) by averaging less in the edge areas. It replaces the pixel of interest with a weighted sum of the values within an n � n moving kernel (Qiu et al., 2004). The despeckled pixel value is estimated using a sub-window of the processing window. The size of the sub-window varies as a function of the target local het erogeneity measured with a coefficient of variation. X Digital number ðDNÞ ¼ kαe αjtj
C-SAR
MultiSpectral Instrument 10 days
nXn
�
S1A_IW_GRDH_1SDV_20180821T004109_20180821T004134_023337_0289D5_B2B2
S2B_MSIL1C_20180822T050649_N0206_R019_T43PFL_20180822T085140 SENTINEL2B
α¼
SENTINEL1A
Product name
Table 1 Specifications of Sentinel-1 and Sentinel-2 products.
Repeat cycle 12 days
Instrument
Product type Ground Range Detected S2MSI1C
Interferometric Wide swath (IW)
Sensing date 21-Aug18
Descending
165
23337
DV (dual VV þ VH polarization)
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Acquisition mode
Pass
Track
Orbit
Polarization
V.K. Rana and T.M.V Suryanarayana
4 nσ 2
�� � σ2 I
2
k ¼ normalized constant I ¼ local mean σ ¼ local variance σ ¼ image coefficient of variation value jtj ¼ |X X0| þ | Y Y0|are present. n ¼ moving kernel size When uniform regions are filtered, the Frost filter acts as a mean filter. When high contrast regions are filtered, the filter acts as a highpass filter with rapid decay of elements away from the filter center. Thus, large uniform areas will tend to be smoothed out and speckle removed, whilst high contrast edges and other objects will retain their signal values and not be smoothed. After application of the Frost filter, the denoised images show better sharpness at the edges. 1.1.4. Lee filter Lee filter is based on the assumption that the filtered or output pixel value is a weighted sum of the reference pixel value and the mean of the values within the kernel (LeeSen, 1981). The filter removes the noise by minimizing either the mean square error or the weighted least square estimation (Qiu et al., 2004). The Lee filter utilizes the statistical dis tribution of the digital number values within the moving kernel to es timate the value of the pixel of interest. This filter assumes the normal distribution for the noise in image data. Iout ¼ ½mean� þ K½Uin
3
mean�
V.K. Rana and T.M.V Suryanarayana
– 12314
VarðxÞ mean2 σ 2 þ VarðxÞ
Variance of � is defined as:
Descending 16-Jul19 INS-NOBS
–
DV (dual VV þ VH polarization) 28113 Ascending 14-Jul19
10 days S2B_MSIL2A_20190716T042709_N0213_R133_T46RDQ_20190716T083530 SENTINEL2B
½variance within kernel� þ ½mean within kernel�2 σ2 þ 1
½mean within kernel�2 Lee’s smoothing filter is adaptive to the local statistics in an image, however, it is an isotropic adaptive filter which cannot remove noise in the edge region effectively. Lee filter is reportedly superior in its ability to preserve prominent edges, linear features, point target, and texture information. 1.1.5. Lee sigma filter This filter is based on sigma probability of the Normal distribution. The sigma (Standard Deviation) of the entire scene is first computed and then each central pixel in a moving window is replaced with the average of only those neighborhood pixels that have intensities within a fixed sigma range of the center pixel. It is well known that, in the normal distribution, the two-sigma likelihood is 0.955. The pixels outside the two-sigma range are considered outliers and ignored. pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi variation S tan dard deviation of an image ¼ mean ¼ Coefficient of Variation ¼ Sigma ðσÞ Due to the use of a fixed sigma computed for the entire scene de Leeuw and de Carvalho, 2009, found that the Lee sigma filter blurred some of the low-contrast edges and linear features.
MultiSpectral Instrument
12 days S1A_IW_GRDH_1SDV_20190714T115653_20190714T115718_028113_032CCD_F972 SENTINEL1A
C-SAR
Ground Range Detected S2MSI1C
Interferometric Wide swath (IW)
41
Polarization Orbit Pass Acquisition mode Product type Repeat cycle
Instrument
K¼
VarðxÞ ¼
Sensing date
Track
Iout ¼ filtered output Uin ¼ unfiltered input mean ¼ average of pixels in a moving kernel
Product name
Table 2 Specifications of Sentinel-1 and Sentinel-2 products.
Remote Sensing Applications: Society and Environment 16 (2019) 100271
1.1.6. Median This filter is not an adaptive filter as it does not account for the particular speckle properties of the image. Destructive and constructive interferences in SAR information are represented by extreme values (low-value and high-value pixels), which are efficiently suppressed by the Median filter (Sheng and Xia, 1996; Qiu et al., 2004). The median filter is successful at removing pulse and spike noise while retaining step and ramp functions. Therefore, the median filter is better than the mean filter in terms of preserving the edges between two different features, but it does not preserve single pixel-wide features, which will be altered if speckle noise is present. Median filter preserves the texture informa tion very well for small window size (3 � 3) but does not retain the mean value at an acceptable level. Since the median is less sensitive than the mean to extreme values (outliers), those extreme values are more effectively removed. 1.2. Measuring performance efficiency of SAR speckle filters A speckle suppression filter is expected to filter the homogeneous areas with reasonable speckle reduction. A good SAR despeckling technique should have the following characteristics (i) scene feature preservation (such as texture, linear features, and point features) (ii) radiometric preservation (iii) speckle-noise reduction, smoothing, blur reduction, and edge preservation. The evaluation of the performance of the filters in de-speckling the SAR image is, therefore, necessary. Selected Homogeneous area, linear feature and Edge from Kerala SAR data is shown in Fig. 1. Following are the parameters to evaluate the performances of a despeckling filter:
4
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 4. Methodology.
Fig. 6. Speckle suppression index.
Fig. 5. Mean square error.
1.2.1. Mean square error (MSE) Mean square error (MSE) is the measurement of the difference be tween the output image and the input image. Higher the value of MSE, higher is the dissimilarity between the unfiltered image and the filtered image. A lower MSE value represents better image quality of the filtered image (Senthilnath et al., 2013). MSE based measurements, however, yield little information about the preservation of specific features as it assesses the whole image. MSE ¼
n 1 hX Iu K i¼1
If
Iu ¼ unfiltered image If ¼ filtered image K ¼ total number of pixels 1.2.2. Speckle suppression index (SSI) The ability of a filter to suppress speckles is measured in terms of the standard deviation of the image to its mean intensity. For homogeneous � � pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� areas, the ratio VarðIu Þ meanðI Þ is regarded as the measurement u
�2 i
of speckle strength. The speckle suppression index (SSI) is the coefficient of variance of the filtered image standardized by that of the unfiltered
5
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 7. Speckle mean preservation index.
Fig. 10. Percent change in standard deviation.
1.2.3. Speckle mean preservation index (SMPI) SSI is not accurate when the mean value is overestimated due to the existence of extreme values in a relatively lower region of the image. In addition to SSI, therefore, SMPI (Speckle Suppression and Mean Pres ervation Index) is used to evaluate the filter efficiency (Wang and Ge, 2012). In terms of mean conservation and noise removal, lower SMPI values show better filter efficiency (Shamsoddini et al., 2010). ! qffiffiffiffiffiffiffiffiffiffiffiffiffi�ffiffi. ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi p SMPI ¼ Q � Var If VarðIu Þ where Q is calculated as under: � �� �Q ¼ 1 þ meanðIu Þ mean If � 1.2.4. Equivalent number of looks (ENL) Another commonly used evaluation criterion is the equivalent number of looks (ENL), also known as measure of the signal-to-noise ratio. This index is calculated using the following equation: � � �2 mean If � ENL ¼ standard deviation If
Fig. 8. Equivalent number of looks.
Higher ENL value for a filter represents higher efficiency in smoothing speckle-noise over homogeneous areas (Bruniquel and Lopes, 1997). The performance of a filter method is evaluated by considering changes in mean and standard deviation. Ideally, the implementation of filters should not result in any change in the mean of the target image, while it should reduce the standard deviation. To date, different classification algorithms, including the Support Vector Machine (SVM), Maximum Likelihood (ML), Classification and Regression Tree (CART), K-Nearest Neighbor (KNN), and Random Forest (RF), have been applied in various studies. The RF classifier is one of the most effective approaches for classification (Breiman, 2001). Various studies have been conducted using pixel-based RF algorithm for wetland vegetation mapping using high spatial resolution SAR data (Amani et al., 2017; Fu et al., 2017; Mahdianpari et al., 2017). Dumitru et al. (2015) applied the SVM classifier for the rapid mapping of damage assessment for flood in Germany in 2013 and the tsunami in Japan in 2011 using TerraSAR-X pre- and post-flood data. The key element of a quantitative accuracy assessment is the creation of a confusion matrix (Congalton, 2001; Janssen and Vanderwel, 1994; Story and Congalton, 1986). The confusion matrix is represented by a table that shows correspondence between the classification result and a reference image assigned to a particular category, which is relative to the actual category as indicated by the reference data. Senthilnath et al. (2013) used quantitative accu racy assessment for flood mapping, which was estimated from the error or confusion matrix. The overall accuracy was calculated by summing the number of correctly classified values and dividing it by the total
Fig. 9. Percent change in mean.
image, which is defined as: ! � � qffiffiffiffiffiffiffiffiffiffiffiffiffi�ffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffi. � VarðIu Þ meanðI Þ SSI ¼ Var If mean I � � f u VarðIf Þ ¼ variance of filtered image VarðIu Þ ¼ variance of the unfiltered image SSI has an inverse relationship with the suppression ability of the filter. The filtered image has lower variance because of speckle sup pression SSI smaller than 1.0 indicates efficient speckle suppression (Sheng and Xia, 1996).
6
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Table 3 Quantitative evaluation of the filters over Kerala. Kerala Percent Change – Mean
Percent Change – Standard Deviation
Filters
3� 3
MSE 5� 5
SSI 3� 3
5� 5
SMPI 3� 3
5� 5
ENL 3� 3
5� 5
3� 3
5� 5
3� 3
5� 5
Frostrowhead Gamma map Lee sigma Lee Boxcar Median
0.005 0.007 0.007 0.094 0.095 0.124
0.004 0.006 0.113 0.124 0.129 0.148
0.937 0.503 0.986 0.505 0.507 0.248
0.983 0.361 0.979 0.386 0.350 0.165
0.930 0.503 0.963 0.505 0.507 0.221
0.979 0.359 0.974 0.386 0.351 0.142
0.120 0.417 0.109 0.414 0.411 1.712
0.109 0.812 0.110 0.708 0.860 3.883
0.825 0.100 2.636 0.055 0.004 12.58
0.396 0.609 0.554 0.142 0.009 15.77
7.05 49.71 3.99 49.51 49.31 78.28
2.138 64.14 2.651 61.41 64.95 86.11
Table 4 Quantitative evaluation of the filters over Assam. Assam Percent Change – Mean
Percent Change – Standard Deviation
Filters
3� 3
MSE 5� 5
SSI 3� 3
5� 5
SMPI 3� 3
5� 5
ENL 3� 3
5� 5
3� 3
5� 5
3� 3
5� 5
Frost Gamma map Lee sigma Lee Boxcar Median
0.007 0.011 0.004 0.011 0.011 0.015
0.004 0.030 0.003 0.030 0.031 0.043
0.769 0.757 0.964 0.757 0.757 0.668
0.886 0.593 0.953 0.594 0.593 0.431
0.763 0.756 0.968 0.757 0.757 0.625
0.874 0.590 0.976 0.592 0.593 0.379
0.297 0.306 0.189 0.306 0.307 0.394
0.224 0.499 0.193 0.498 0.499 0.945
0.924 0.122 0.292 0.062 0.000 7.085
1.579 0.539 2.155 0.270 0.001 13.26
23.82 24.38 3.28 24.34 24.30 37.95
12.75 41.00 2.63 40.79 40.70 62.60
number of values. Producer’s accuracy is the probability that value in a given class was correctly classified.
the subtle structures of the image is essential, the efficiency of noise suppression must be balanced with the effectiveness of the filter in order to keep fine detail. The most preferred speckle filters are, therefore, assessed in the current study over the data from Sentinel-1, intended for Flooded area properly identified in a classification method Producer’s accuracy ¼ flood mapping applications. The Sentinel-1 (VV-vertical transmit, ver Flooded area in the reference ground truth tical receive and VH-vertical transmit, horizontal receive) polarised filtered data were later used for performance evaluation of machine User’s accuracy is the probability that a value predicted to be in a learning algorithms, namely, random forest (RF) and support vector certain class is really in that class. machine (SVM), to classify an inundated area. The accuracies of the Flooded area properly identified in a classification method classifications were assessed by the confusion matrix parameters, which User’s accuracy ¼ Total flooded area calculated from the method include kappa coefficient, overall accuracies, producer’s and user’s accuracies. The kappa coefficient measures the agreement between classification and truth-values. A kappa value of 1 represents perfect agreement, while 2. Study area and datasets a value of 0 represents no agreement Kappa coefficient ¼
2.1. a. Kerala
Observed accuracy Expected agreement 1 Expected agreement
Kerala is a small, elongated coastal state in peninsular India’s southwestern tip. It is surrounded by the Western Ghats in the east and the Arabian Sea in the west. A part of the state of Kerala was considered in this study. The state faces severe and varied damages due to floods and heavy rainfall. Monsoon circulation dominates the climate of India and Kerala in particular. The wind blows from the oceans to the south of the Asian land masses during the half of the year, while a seasonal wind blows from the Asian land masses to the oceans in the south during the other half of the year causing a spectacular reversal of pressure and wind patterns between the two six-month periods. South-west monsoon (June–September) and post-monsoon (October–November) are the main rainy seasons in Kerala. The state witnessed heavy floods in the year 1924 and 1961. The IMD recorded rainfalls for 15 to 17 August 2018 were found to be comparable to the rigorous storm that occurred in 1924 (Central Water Commission, 2018). Heavy rainfall resulted in high surface runoff in Kerala’s major river basins, filling all dams and sub sequent opening of these dams, causing widespread flooding in down stream areas, low-lying coastal areas, and Kerala’s backwaters. Fig. 2 shows the area covered under the study. The National Aeronautics and Space Administration Alaska Satellite Facility (NASA/ASF) houses a complete archive of Sentinel-1 SAR data processed by the European Space Agency (ESA). The Sentinel-1 Level-1
where Expected agreement ¼
Ct * ft þ nct * nft ðAÞ2
Ct ¼ actual total flooded area ft ¼ actual total flooded area nct ¼ total non-flooded area from a classified method nft ¼ actual non-flooded total area A ¼ total area under the study The objective of this study was to develop an approach for opera tional flood extent mapping. Despeckling is an important preprocessing step as speckle complicates the interpretation of visual images making automated digital image classification a challenging task. The choice of filter relies on the particular application requirements and the data set characteristics used. Despeckle filters with excellent noise extraction capacities often appear to degrade the spatial and radiometric accuracy of the actual image and trigger image reflection deterioration. This may be acceptable for applications involving large-scale interpretation or mapping of images. However, in many cases where the preservation of 7
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 11. Visual comparison of de-noising methods on VV polarization over Kerala test data (a) Boxcar 3 � 3 (b) Frost 3 � 3 (c) Gamma map 3 � 3 (d) Lee 3 � 3 (e) Lee sigma 3 � 3 (f) Median 3 � 3 (g) Boxcar 5 � 5 (h) Frost 5 � 5 (i) Gamma map 5 � 5 (j) Lee 5 � 5 (k) Lee sigma 5 � 5 (l) Median 5 � 5 and (m) Unfiltered test image in VV polarization.
ground range detected (GRD) data acquired in interferometric wide swath (IW) mode, which is the predefined mode over land with VV and VH polarizations, were downloaded via the ASF application program ming interface (API). Sentinel-1 and Sentinel-2 data that are available closest to event date were acquired on 21 August 2018 at 00:40:44 and 22 August 2018 at 05:06:49, respectively. The specific parameters of the Sentinel-1 and Sentinel-2 products are given in Table 1.
owing to the increase in water concentrations of the Brahmaputra river and its related tributaries likely led from high continuous rainfall in the upper catchment regions of the Brahmaputra Basin. Fig. 3 shows the area covered under the study. Sentinel-1 and Sentinel-2 data that are available closest to event date were acquired on 14 July 2019 at 11:57:18 and 16 July 2019 at 04:27:09, respectively. The specific pa rameters of the Sentinel-1 and Sentinel-2 products are given in Table 2. 3. Methodology
2.2. b. Assam
3.1. Pre-processing
Assam is a state in northeast India, situated south of the eastern Himalayas along the Brahmaputra and Barak River valley. The state has recently witnessed heavy flood in July 2019. The Brahmaputra basin falls within the monsoon rainfall regime, getting an average rainfall of about 230 cm. The heavy floods in the Brahmaputra river in Assam
A schematic of the Sentinel-1-based processing chain is outlined in Fig. 4. The downloaded Sentinel-1 Leve-1 GRD data acquired in IW with VV and VH polarizations were loaded onto Sentinel Application 8
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 12. Visual comparison of de-noising methods on VV polarization over Assam test data (a) Boxcar 3 � 3 (b) Frost 3 � 3 (c) Gamma map 3 � 3 (d) Lee 3 � 3 (e) Lee sigma 3 � 3 (f) Median 3 � 3 (g) Boxcar 5 � 5 (h) Frost 5 � 5 (i) Gamma map 5 � 5 (j) Lee 5 � 5 (k) Lee sigma 5 � 5 (l) Median 5 � 5 and (m) Unfiltered test image in VV polarization.
Fig. 13. Point target scattering on (a) Boxcar 3 � 3 (b) Lee 3 � 3 (c) Boxcar 5 � 5 and (d) Lee 5 � 5.
Platform (SNAP). SNAP offers a wide range of tools and features for Sentinel-1 imagery processing and analysis. Due to the large swath width of the Sentinel-1 data, the image was first divided into a subset for
the study sites to reduce the processing time. Multi-looking was then performed to reduce the standard deviation of the noise. The number of Azimuth looks and the number range of looks (2 � 2) with mean GR 9
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 14. Kerala (a) Raw VH amplitude data; (b) Raw VV amplitude data; (c) Multi-looked, calibrated, Filtered (Lee) VH data in dB; (d) Multi-looked, calibrated, Filtered (Lee) VV data in dB; (e) The Range-Doppler terrain corrected VH data; (f) The Range-Doppler terrain corrected VV data; (g) Histogram for Sigma0 VH in dB, and (h) Histogram for Sigma0 VV in dB.
mean pixel of 20 m were applied to a 1 m � 5 m (single look). The multilooked data were then calibrated to transform the pixel values from the digital values recorded by the sensor into backscatter coefficient values or Sigma0 (σ0). This was achieved using the following equation:
σ0 ¼
jDNi j Ai 2
DNi ¼ pixel’s digital number Ai ¼ absolute calibration constant 3.2. Application of filters
2
Speckles inherently corrupt all radar images, degrading the image quality, and making it more difficult to interpret features. Thus, it is 10
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 15. Assam (a) Raw VH amplitude data; (b) Raw VV amplitude data; (c) Multi-looked, calibrated, Filtered (Lee) VH data in dB; (d) Multi-looked, calibrated, Filtered (Lee) VV data in dB; (e) The Range-Doppler terrain corrected VH data; (f) The Range-Doppler terrain corrected VV data; (g) Histogram for Sigma0 VH in dB, and (h) Histogram for Sigma0 VV in dB. Table 5 Comparison of user’s accuracy (UA), producer’s accuracy (PA), overall accuracy (%), and kappa coefficient using random forest tree and support vector machine algorithms for VV and VH polarization over Kerala region. VV Polarization RF Inundation Rest Kappa coefficient Overall Accuracy (%)
VH Polarization SVM
RF
SVM
PA
UA
PA
UA
PA
UA
PA
UA
0.89 0.88 0.72
0.96 0.72
0.79 0.98 0.64
0.99 0.60
0.78 0.98 0.61
0.99 0.58
0.79 0.98 0.63
0.99 0.60
88.80%
Table 6 Inundated area statistics of RF and SVM over cloud-free optical data for Kerala region.
83.80%
82.60%
Inundated area (Km2) Rest (Km2)
VV Polarization
VH Polarization
RF
SVM
RF
SVM
NDWI –
40.25 16.29
34.82 21.72
34.17 22.38
35.19 21.36
41.78 14.77
kernel size, used in the study were available in SNAP and applied using default system parameters.
83.60%
3.3. Evaluating the performance efficiency of filters
Note: UA - User’s accuracy, PA - Producer’s accuracy VV - Vertical-Vertical, VH Vertical-Horizontal, RF - Random Forest, SVM -Support Vector Machine.
Several techniques are available to quantitatively assess the effi ciency of a speckle filter in distinct ways, for example, edge conservation and conservation of features. The findings of the various measurements may be contradictory. Therefore, distinct techniques of evaluation should be used to discover the optimum tradeoff between the various
often necessary to enhance the image by filtering speckles before data can be used in different applications. All of the filters, namely, Boxcar, Median, Frost, Gamma map, Lee and Lee sigma with 3 � 3 and 5 � 5 11
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
producer’s accuracy, user’s accuracy, and the overall accuracy of the classifiers. The overall accuracy gives the correctly classified regions for the image and is calculated by the proportion of the correctly classified pixels to the total number of pixels in the confusion matrix. To calculate inundation for entire scene, thresholding was done in SNAP by using the conditional function given below after carefully analysing the histogram:
Table 7 Comparison of user’s accuracy (UA), producer’s accuracy (PA), overall accuracy (%), and kappa coefficient using random forest tree and support vector machine algorithms for VV and VH polarization over Assam region. VV Polarization RF Inundation Rest Kappa coefficient Overall Accuracy (%)
VH Polarization SVM
RF
SVM
PA
UA
PA
UA
PA
UA
PA
UA
0.77 1.0 0.65
1.0 0.62
0.82 1 0.72
1.0 0.69
0.81 0.99 0.70
0.99 0.67
0.89 0.99 0.81
0.99 0.77
83.60%
87.60%
86.60%
If Sigma0 VV/VH db < X then 1 else 0 X ¼ threshold value 1 ¼ inundated pixels 0 ¼ Rest
92.00%
Note: UA - User’s accuracy, PA - Producer’s accuracy VV - Vertical-Vertical, VH Vertical-Horizontal, RF - Random Forest, SVM -Support Vector Machine.
The resulted output was later used to calculate the difference in inundated areas calculated by thresholding technique and images clas sified using the random forest and support vector machine algorithms having highest accuracy for both the study areas.
elements of the image. The sigma0 values of VV and VH polarised data with the applied filter was terrain corrected using SNAP’s ‘Range Doppler Terrain Correction’ algorithm with an SRTM 1 ArcSecond digital elevation model. The bilinear interpolation method was used for DEM and Image resampling with a pixel spacing of 20 � 20 m. Terrain correction helps in improving the geometric representation of the real-world surface. This is needed because, during image capture, topographical variations and off-nadir distortion unsettle the image.
4. Results and discussion 4.1. Speckle filtering The performance efficiency of filters in speckle suppression, feature preservation, and preventing the loss of meaningful data was evaluated using MSE, ENL, SSI, SMPI, examination of mean, standard deviation, and also by close visual assessment. The MSE value indicates the amount of error present in a filtered image. It allows the comparison of the pixel values of a filtered image to the degraded image before filtering. It is observed that for study area Kerala (window size 3 � 3) Frost, Gamma map, and Lee sigma filters showed very low MSE, indicating their effectiveness in feature preser vation. Median filter showed the highest MSE values indicating poor performance in terms of feature preservation. Boxcar and Lee filters performed moderately. For Kerala (window size 5 � 5) and Assam (window size 3 � 3 and 5 � 5), Frost and Lee sigma showed low MSE, Median filter showed the highest MSE and moderate performance was observed for Boxcar, Gamma map and Lee (Fig. 5). The high value of MSE depicts a greater difference between the original test image and despeckled image which concludes the significant speckle reduction but at the cost of feature loss. Lower SSI and SMPI values were obtained for Median filter (3 � 3 and 5 � 5) in both the study areas indicating high efficiency in speckle suppression. Moderate values were obtained for Gamma map, Lee and Boxcar indicating medium proficiency in the reduction of speckle. None of the filters produced value 0 and 1. A value of 0, indicates complete mean preservation and 100% noise reduction. Whereas, a value of 1 indicates 0% speckle reduction. It was observed that as the window size increased from 3 � 3 to 5 � 5, the SSI values reduced for all the filters (Figs. 6 and 7), indicating superior speckle suppression with an increase in the size of the moving window. The lowest SSI values, as well as SMPI values, corresponds to Median filter (3 � 3 and 5 � 5) indicating high efficiency in speckle suppression. For both the study areas Median filter (3 � 3 and 5 � 5) had high ENL values indicating a higher efficiency in smoothing speckle noise over homogeneous areas, which shows enhanced capacity to distinguish the distinct gray levels within the image. For constant flat areas where the sample variance is null, ENL become ∞, this will repute highly blurred data as excellent. According to ENL, SSI and SMPI values, Gamma map (3 � 3 and 5 � 5), Lee (3 � 3 and 5 � 5) and Boxcar (3 � 3 and 5 � 5) filters showed moderate performance for both the study areas (Fig. 8). The application of filters should ideally not bring about any change in the mean of target image while it should reduce the standard devia tion. The filter, Median (3 � 3 and 5 � 5) for both the study areas was most effective in increasing the standard deviation but also changed the mean value considerably thereby implying that the filters reduced
3.4. Machine learning algorithms for classification The terrain corrected images were classified using the random forest and support vector machine algorithms as a next step. For both the classifiers, the same number of training samples was used. The training inundated pixels covered 5.2 Km2 and the rest of the training pixels covered 3.1 Km2 of the study area Kerala. Similarly, for the study area Assam, the training inundated pixels covered 11 Km2 and the rest of the training pixels covered 54 Km2. During the southwest monsoon season, it is nearly impossible to obtain 100% cloud-free data, however, a small extent of the cloud-free data can be used for validation. The normalized difference water index (NDWI) is defined for Sentinel–2 data as ((B03) (B08)/(B03) þ (B08)), where B03 is a green band and B08 is the near-infrared band (Mcfeeters, 1996). When NDWI is applied over a multispectral image, the water feature has positive values, while soil and terrestrial vegeta tion features have zero or negative values. This is because NIR is absorbed strongly by water but reflected strongly by terrestrial vegeta tion and dry soil, while in a green light, water has high reflectance than terrestrial vegetation and soil. Therefore, the NDWI was applied to extract water from the optical data. A cloud-free part of satellite optical image was collected by Sentinel-2 at 05:06:49 on 22 August 2018, 28 h after the Sentinel-1 pass over the study area Kerala. Similarly, a cloud-free part of satellite optical image was collected by Sentinel-2 at 04:27:09 on 15 July 2019, 40 h after the Sentinel-1 pass over the study area Assam. Sentinel-2 data were converted to reflectance and dark object subtraction atmospheric correction (DOS1) was applied. The corrected Sentinel-2 image was used to validate the extent of the flood. The normalized difference water index (NDWI), established earlier (Mcfeeters, 1996) to extract the water from the optical data, was calculated as: NDWI ¼
ρGreen ρNIR ρGreen þ ρNIR
Moreover, stratified random sampling was used with 500 points for the accuracy assessment. The classification maps were evaluated in terms of their overall accuracy (OA), producer’s accuracy (PA), user’s accuracy (UA), and the kappa index of agreement (k) or kappa coeffi cient. Confusion matrix was created to compare the kappa coefficient, 12
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 16. Kerala (a) Random forest tree classification on filtered VH; (b) Support vector machine classification on VH; (c) Random forest tree classification on filtered VV; (d) Support vector machine classification on VV.
13
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 17. Assam (a) Random forest tree classification on filtered VH; (b) Support vector machine classification on VH; (c) Random forest tree classification on filtered VV; (d) Support vector machine classification on VV.
speckle but also caused considerable loss of meaningful data. The me dian filter is better than the Boxcar filter in terms of preserving the edges between two different features, but it does not preserve single pixel-wide features, which will be altered if speckle noise is present. Median filter preserves the texture information very well for small window size (3 � 3) but does not retain the mean value at an acceptable level. Since the median is less sensitive than the mean to extreme values (outliers), those extreme values are more effectively removed. On the contrary, Boxcar (3 � 3 and 5 � 5) for both the study areas, made the least change in the mean, while reducing the standard deviation moderately. How ever, Lee filter (3 � 3 and 5 � 5) provided a fair balance by reducing the standard deviation without drastically affecting the mean (Figs. 9 and 10, and Tables 3 and 4). The Median filter performed well in terms of ENL, SSI, and SMPI values; however, its performance in terms of speckle reduction and feature preservation was far inferior compared to the Boxcar and Lee filter. However, since the objective of the speckle filtering was to use these SAR images for inundation mapping, the performance of the filters on a water body in terms of reduction of standard deviation while pre serving the mean of the original image was the most important. The
performance of Boxcar and Lee filters was far better in feature preser vation in the filtered images followed by Gamma map. Although quan titative measures are often employed to compare different speckle suppression filters, it has been noted by Raouf and Lichtenegger (1997) and others that visual inspection probably provides the best assessment of the performance of a speckle filter. Visual assessment is an easy and efficient way to investigate both the capability of a filter to suppress speckles and its effectiveness in preserving image details. Lee et al. (1994) stated that, in general, filters using small windows (such as 3 � 3) preserve texture information better. Visual examination was, therefore, carried out and it was observed that the filters which reduced speckles effectively also resulted in considerable loss of meaningful data (Figs. 11 (a)-(m) and 12 (a)-(m)). Lee sigma, and Gamma map clearly resulted in the loss of edges and details. It was difficult to grade Boxcar, Frost and Lee filter visually, since the variation was not perceivable. However, it was observed that point scatters were over filtered, transformed to spread targets and sharp edges were generally blurred in Boxcar filter (Fig. 13). We chose Lee filter for further analysis as it had low MSE, SSI and SMPI values, and a higher percentage change in standard deviation compared to the Boxcar filter for most of the cases. 14
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 18. (a) Natural colour composite R–B04 G-B03 B–B02, (b) Green band (c) Near-infrared (d) Calculated NDWI over cloud-free extent for Kerala region.
The raw SAR data in VH and VV polarization acquired during the crisis events on 21 August 2018 and 14 Jul 2019 for study areas Kerala and Assam are shown in Fig. 14 (a) and (b), and Fig. 15 (a) and (b), respectively. The multi-looked, calibrated, filtered (Lee) data of the VH and VV are shown in Fig. 14 (c) and (d), and Fig. 15 (c) and (d), respectively. Multi-looked, calibrated, filtered (Lee) SAR data were not projected on the map coordinates of each pixel. The pixel was in the original coordinate position of data (rows/columns) in the field of ground range. In the orthorectified imageries, each of the pixels that were corrected and projected using the Range-Doppler terrain correc tion appeared at the actual position. The Range-Doppler terrain cor rected pixels and their respective histograms for Kerala and Assam regions are shown in Fig. 14 (e)–(h), and Fig. 15 (e)- (h), respectively. Tables 5 and 7, and Figs. 16 and 17 show the comparison of classi fication results of random forest classifier and support vector machine classifier for VV and VH polarization. The training data were kept the same for both the classifiers to avoid optimistic bias in the classification. For the study area Kerala, the random forest classifier exhibited maximum overall accuracy of 88.80% with the kappa coefficient value of 0.72. Both the classifiers obtained better accuracy results in VV po larization compared to the VH polarization. The least overall accuracy of 82.60% and a kappa coefficient value of 0.63 were observed with random forest in VH polarization, which was followed by the support vector machine in VV polarization. RF achieved higher classification accuracy than SVM by about 5% in VV polarization. However, both the
classifiers produced comparable overall accuracies in VH polarization (SVM achieved higher classification accuracy than RF by about 1%). The NDWI calculated for the cloud-free extent is shown in Fig. 18 (d). The inundated area in the calculated NDWI over the cloud-free extent is 73.88%, which is 41.78 km2. However, it has also been observed (Table 6) that the inundated area using random forest classification on filtered VV data over the cloud-free extent is 71.18%, which is 40.25 Km2. For the study area Assam, the SVM classifier exhibited maximum overall accuracy of 92% with the kappa coefficient value of 0.81. Both the classifiers obtained better accuracy results in VH polari zation compared to the VV polarization. The least overall accuracy of 83.60% and a kappa coefficient value of 0.65 were observed with random forest in VV polarization, which was followed by the RF in VH polarization. SVM achieved higher classification accuracy than RF by about 5.38% in VH polarization. The NDWI calculated for the cloud-free extent is shown in Fig. 19 (d). The inundated area in the calculated NDWI over the cloud-free extent is 74.09%, which is 491.47 km2. However, it has also been observed (Table 8) that the inundated area using SVM classification on filtered VH data over the cloud-free extent is 62.76%, which is 416.99 Km2. To calculate inundation for the entire scene, threshold value of 10.96 and 19.58 was selected for Kerala and Assam region respectively after analysing the histograms. The calculated inundated area (Fig. 20) with thresholding technique for Kerala was found 204 km2 (2% more than classified VV polarised data using RF algorithm). Similarly, the calculated inundated area with 15
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 19. (a) Natural colour composite R–B04 G-B03 B–B02, (b) Green band (c) Near-infrared (d) Calculated NDWI over cloud-free extent for Assam region.
balance in feature preservation as well in despeckling compared to the other filters used in the study. The accuracy assessment of machine learning algorithms for flood classification over Kerala shows that random forest classifier has higher classification accuracy than the support vector machine (about 5% higher in VV polarization). However, both the classifiers performed better in VV polarization than VH polar ization. For study area Assam, SVM in VH polarization achieved higher accuracy and least performance was observed by RF in VV polarization. The proposed approach shows the potential for monitoring damages caused by floods, providing basic information that can help local com munities manage water-related risk, planning land and water manage ment as well as other flood control programs.
Table 8 Inundated area statistics of RF and SVM over cloud-free optical data for Assam region.
Inundated area (Km2) Rest (Km2)
VV Polarization
VH Polarization
RF
SVM
RF
SVM
NDWI –
363.92 300.41
389.21 275.12
387.62 276.70
416.99 247.34
491.47 171.86
thresholding technique for Assam was found 3368.90 km2 (23.46% more than classified VH polarised data using SVM algorithm). For Assam region the variation is very large, a single threshold did not hold well as large swath of a SAR image suffers from environment heterogeneity caused by wind-roughening and satellite framework parameters.
Ethical approval The presented research has not involved human subjects, human material, or human data and therefore ethical approval has not been required.
5. Conclusion Effective and quick response is required during disasters like flood ing. Rapid mapping of such events will be beneficial to urban and infrastructure planners, risk managers and disaster responses during extreme and intense rainfall events. The study shows a simple and efficient method for mapping inundation extent with only the C-band S1A, with coarser geometric resolution and fixed polarizations (VV-VH) by considering the case of Kerala and Assam. Backscattering coefficient values become high as the water roughness causes high signal return, decreasing the contrast and making the separation of the land-water covers difficult. Despeckle filters with good noise removal capabilities often tend to degrade the spatial and radiometric resolution of an original image and cause the loss of image details. This may be acceptable for applications involving large scale image interpretation or mapping. However, the retention of the subtle structures of an image is important and, therefore, the performance of speckle noise suppression technique must be balanced with the filter’s effectiveness to preserve the fine details. The performance evaluation of de-noising methods in this study showed that Lee filter with 3 � 3 kernel size provided a good
Compliance with ethical standards This study is in full compliance with all applicable ethical standards. Informed consent This study does not involve any subject participation and therefore informed consent is not applicable. Declaration of competing interest On behalf of all authors, the corresponding author states that there is no conflict of interest.
16
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 20. (a) Random forest tree classification on filtered VV data (Kerala) (b) Classified filtered VV data with threshold value of machine classification on filtered VH data (Assam) (d) Classified filtered VH data with threshold value of 19.58 (Assam).
References
10.96 (Kerala) (c) Support vector
Fu, B., Wang, Y., Campbell, A., Li, Y., Zhang, B., Yin, S., et al., 2017. Comparison of object-based and pixel-based Random Forest algorithm for wetland vegetation mapping using high spatial resolution GF-1 and SAR data. Ecol. Indicat. 73, 105–117. https://doi.org/10.1016/j.ecolind.2016.09.029. Gitas, I.Z., Polychronaki, A., Katagis, T., Mallinis, G., 2008. Contribution of remote sensing to disaster management activities: a case study of the large fires in the Peloponnese, Greece. Int. J. Remote Sens. 29 (6), 1847–1853. https://doi.org/ 10.1080/01431160701874553. Goodman, J.W., 1976. Some fundamental properties of speckle. JOSA 66 (11), 1145–1150. Janssen, L.L., Vanderwel, F.J., 1994. Accuracy assessment of satellite derived land-cover data: a review. Photogramm. Eng. Remote Sens. 60 (4). Khan, S.D., 2005. Urban development and flooding in Houston Texas, inferences from remote sensing data using neural network technique. Environ. Geol. 47 (8), 1120–1127. https://doi.org/10.1007/s00254-005-1246-x. Lee, J.S., Pottier, E., 2009. Polarimetric Radar Imaging: from Basics to Applications. CRC press. Lee, J.S., Jurkevich, I., Dewaele, P., Wambacq, P., Oosterlinck, A., 1994. Speckle filtering of synthetic aperture radar images: a review. Remote Sens. Rev. 8 (4), 313–340. https://doi.org/10.1080/02757259409532206. Lee, J. Sen, 1981. Speckle analysis and smoothing of synthetic aperture radar images. Comput. Graph. Image Process. 17 (1), 24–32. https://doi.org/10.1016/S0146-664X (81)80005-6. Lopes, A., Nezry, E., Touzi, R., Laur, H., 1990. May). Maximum a posteriori speckle filtering and first order texture models in SAR images. In: 10th Annual International Symposium on Geoscience and Remote Sensing, pp. 2409–2412 (Ieee). Mahdianpari, M., Salehi, B., Mohammadimanesh, F., Brisco, B., 2017. An assessment of simulated compact polarimetric SAR data for wetland classification using random forest algorithm. Can. J. Remote Sens. 43 (5), 468–484. https://doi.org/10.1080/ 07038992.2017.1381550. Mcfeeters, S.K., 1996. The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. Int. J. Remote Sens. https://doi.org/10.1080/ 01431169608948714.
India Meteorological Department, 2018. Rainfall over Kerala during Monsoon Season2018 and forecast for next 5 days. Retrieved from. https://www.imdtvm.gov. in/images/rainfall. over kerala during monsoon season-2018 and forecast for next 5 days.pdf. Amani, M., Salehi, B., Mahdavi, S., Granger, J., Brisco, B., 2017. Wetland classification in Newfoundland and Labrador using multi-source SAR and optical data integration. GIScience Remote Sens. 54 (6), 779–796. https://doi.org/10.1080/ 15481603.2017.1331510. Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5–32. https://doi.org/ 10.14923/transinfj.2015IUP0008. Bruniquel, J., Lopes, A., 1997. Multi-variate optimal speckle reduction in SAR imagery. Int. J. Remote Sens. 18 (3), 603–627. https://doi.org/10.1080/014311697218962. Central Water Commission, 2018. Kerala Floods August 2018. Retrieved from. https:// reliefweb.int/sites/reliefweb.int/files/resources/Rev-0.pdf. Chunming, H., Huadong, G., Yun, S., Jingjuan, L., 2005, July. Detection of the flood boundary in SAR image using texture. In: Proceedings. 2005 IEEE International Geoscience and Remote Sensing Symposium, vol. 5. IGARSS’05, pp. 3697–3699, 2005, (Ieee). Congalton, R.G., 2001. Accuracy assessment of wildland fire and other spatial information. Int. J. Wildland Fire 10, 321–328. de Leeuw, M.R., de Carvalho, L.M.T., 2009. Performance evaluation of several adaptive speckle filters for SAR imaging. Anais XIV Simp� osio Brasileiro de Sensoriamento Remoto 7299–7305. Dumitru, C.O., Cui, S., Faur, D., Datcu, M., 2015. Data analytics for rapid mapping: case study of a flooding event in Germany and the tsunami in Japan using very high resolution SAR images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 8 (1), 114–129. https://doi.org/10.1109/JSTARS.2014.2320777. Durand, J.M., Gimonet, B.J., Perbos, J.R., 1987. SAR data filtering for classification. IEEE Trans. Geosci. Remote Sens. (5), 629–637. https://doi.org/10.1109/ TGRS.1987.289842. GE-25. Frost, V.S., Stiles, J.A., Shanmugam, K.S., Holtzman, J.C., 1982. A model for radar images and its application to adaptive digital filtering of multiplicative noise. IEEE Trans. Pattern Anal. Mach. Intell. 4 (2), 157–166.
17
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Nirupama, N., Simonovic, S.P., 2007. Increase of flood risk due to urbanisation: a Canadian example. Nat. Hazards 40 (1), 25–41. https://doi.org/10.1007/s11069006-0003-0. Qiu, F., Berglund, J., Jensen, J.R., Thakkar, P., Ren, D., 2004. Speckle noise reduction in SAR imagery using a local adaptive median filter. GIScience Remote Sens. 41 (3), 244–266. https://doi.org/10.2747/1548-1603.41.3.244. Raouf, A., Lichtenegger, J., 1997. Integrated use of SAR and optical data for coastal zone management. In: Third ERS Symposium on Space at the Service of Our Environment, vol. 414, p. 1089. Senthilnath, J., Shenoy, H.V., Rajendra, R., Omkar, S.N., Mani, V., Diwakar, P.G., 2013. Integration of speckle de-noising and image segmentation using Synthetic Aperture Radar image for flood extent extraction. J. Earth Syst. Sci. 122 (3), 559. https://doi. org/10.1007/s12040-013-0305-z.
Serpico, S.B., Dellepiane, S., Boni, G., Moser, G., Angiati, E., Rudari, R., 2012. Information extraction from remote sensing images for flood monitoring and damage evaluation. Proc. IEEE 100 (10), 2946–2970. Shamsoddini, A., Trinder, J.C., Wagner, W., Sz� ekely, B., 2010. Image Texture Preservation in Speckle Noise Suppression (na). Sheng, Y., Xia, Z.G., 1996. May). A comprehensive evaluation of filters for radar speckle suppression. In: IGARSS’96. 1996 International Geoscience and Remote Sensing Symposium, vol. 3. IEEE, pp. 1559–1561. Story, M., Congalton, R.G., 1986. Accuracy assessment: a user’s perspective. Photogramm. Eng. Remote Sens. 52 (3), 397–399. Wang, X., Ge, L., 2012. Evaluation of Filters for ENVISAT ASAR Speckle Suppression in Pasture Area. Zhang, F., Zhu, X., Liu, D., 2014. Blending MODIS and Landsat images for urban flood mapping. Int. J. Remote Sens. 35 (9), 3237–3253. https://doi.org/10.1080/ 01431161.2014.903351.
18
Contents lists available at ScienceDirect
Remote Sensing Applications: Society and Environment journal homepage: http://www.elsevier.com/locate/rsase
Evaluation of SAR speckle filter technique for inundation mapping Vikas Kumar Rana *, T.M.V Suryanarayana Water Resources Engineering and Management Institute, Faculty of Technology & Engineering, The Maharaja Sayajirao University of Baroda, Gujarat, India
A R T I C L E I N F O
A B S T R A C T
Keywords: Inundation SAR Machine learning Sentinel-1
The presence of speckle in visual images makes the automated digital image classification a challenging task. Therefore, reduction of speckles is an important pre-processing step. The choice of speckle filter depends on the requirements of an application and the characteristics of the dataset. In this study, some most preferred speckle filters are assessed for the data from Sentinel-1 to map flood extent. The Sentinel-1 (VV-vertical transmit, vertical receive and VH- vertical transmit, horizontal receive) polarizing filter data were used to evaluate machine learning algorithms, namely, random forest (RF) and support vector machine (SVM), to classify an inundated area. The accuracies of the classifications were assessed by kappa coefficient, overall accuracies, and producer’s and user’s accuracies. The present study suggests an approach to monitor damage and provide basic information to help local communities manage water-related risk, land planning, water management, and flood control programs.
1. Introduction In recent years, severe rainfall events have afflicted the state of Kerala in southern India causing damage to houses and infrastructures. Remotely sensed data can provide significant mapping capabilities during such severe rainfall events. However, obtaining remotely sensed data with an ideal combination of fine spatial and temporal resolution with the ability to see through clouds and discriminate flooding under forest cover is a difficult task. The extent of inundation, caused by river flooding and/or coastal storm surges, is required quickly to expedite relief and repair services. The precipitation over Kerala amid June, July and August (1–19 August 2018) were 15%, 18% and 164% above normal, respectively. Due to intense rainfall, all the major reservoirs were full by the end of July 2018 and had no buffer storage to accom modate the inflows from 10th of August 2018 (Central Water Commis sion, 2018). Serious spell of precipitation from the 14 August 2018 to 19 August 2018 brought appalling flood in 13 out of 14 districts. The perpetuated exceptional rainfall in August (170% above normal) in the catchment areas compelled the authorities to resort to hefty downstream release into the rivers (India Meteorological Department, 2018). Remote sensing promises exceptional capacity in catastrophe control owing to its regular acquisition function over a large spatial extent (Serpico et al., 2012; Nirupama and Simonovic, 2007; Gitas et al., 2008; Khan, 2005). Flooding is a complex phenomenon due to its heteroge neity and spectral diversity. The analysis of flood mapping require high
spatial and temporal resolution images to track the rapidly retreating flood process (Zhang et al., 2014). Drastic variability in climate has accelerated the incidence of cata strophic flood events in the last decade (Chunming et al., 2005). In any flood-related study, identification of the flood extent and susceptible areas is a prerequisite to assess the disaster impact. Flood mapping can best be achieved with the help of remote sensing due to the inaccessi bility to the flood-affected regions. However, cloudy conditions reduce flood mapping accuracy below the acceptable levels in optical remote sensing. Synthetic Aperture Radar (SAR) imaging is an efficient remote sensing technique offering well-developed, consistent, efficient, and reliable means of collecting information to extract earth’s surface dielectric properties (Lee and Pottier, 2009). The ability of SAR to penetrate clouds is extremely useful in flood-related studies. Synthetic aperture radar uses microwave radiation to illuminate the earth’s sur face for recording the amplitude and phase of the back-scattered radi ation, which makes the imaging process coherent. The active sensor of Sentinel-1 forms a SAR image by coherently processing the returning signals from successive radar pulses. Stronger or weaker final signals (output) are generated by the out-of-the-phase waves by constructively or destructively interfering with each other. These interferences produce a seemingly random pattern of brighter and darker pixels giving the radar images a distinctly grainy appearance known as ‘Speckle’ (Goodman, 1976; Lee et al., 1994). Speckle noise changes the spatial
* Corresponding author. WREMI, The Maharaja Sayajirao University of Baroda, India. E-mail addresses: [email protected] (V.K. Rana), [email protected] (T.M.V Suryanarayana). https://doi.org/10.1016/j.rsase.2019.100271 Received 16 June 2019; Received in revised form 3 October 2019; Accepted 22 October 2019 Available online 22 October 2019 2352-9385/© 2019 Elsevier B.V. All rights reserved.
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
retain the polarimetric characteristics of certain pixels. However, it over filters the point targets, creates a combination of heterogeneous pixels and degrades the overall spatial information. The filter is simple and fast, however, it is not isotropic (i.e. circularly symmetric), but smooths further along diagonals than along rows and columns. Also, disconti nuities are found in the smoothed image due to an abrupt cut-off of weights rather than decline gradually to null. 1.1.2. Gamma map filter This filter is based on the Bayesian analysis of image statistics. The scene reflectivity of the underlying image in Gamma-Map algorithm is assumed to be Gamma distributed rather than normally distributed, and speckle is noise within it. Thus this filter works best for geospatial im ages containing homogenous areas such as oceans, forests, fields, etc. (Lopes et al., 1990). It is given by following cubic equation (Frost et al., 1982).
Fig. 1. Selected Homogeneous area, linear feature and Edge.
statistics of the underlying scene backscatter making the classification of imageries a difficult task (Durand et al., 1987). A brief introduction of some well-known despeckling methods is presented below.
bI 3
1.1. DE-NOISING methods
2 IbI þ σðbI
DNÞ ¼ 0
bI ¼ required value
The presence of speckle is the major challenge in the SAR image processing. A speckle reduces the resolution of an image and the detectability of the ground targets. It also distorts the spatial patterns of surface characteristics and reduces the accuracy of image classification (Wang and Ge, 2012). Speckles are signal-dependent and, therefore, act like multiplicative noise (LeeSen, 1981).
I ¼ local mean DN ¼ input value σ ¼ original image variance Gamma-Map approach has several advantages compared to the other filters, as it can simultaneously take into account realistic first and second order statistical models for both speckle and underlying scene reflectivity, and combine them through Bayesian inference. Thus this filter works best for geospatial images containing homogenous areas such as oceans, forests and fields.
1.1.1. Boxcar filter A simple averaging filter that replaces the center pixel in a 3 � 3 or a larger sized moving kernel with the mean value of kernel pixels. It has good performance in reducing speckles in a homogeneous area; how ever, it degrades spatial resolution due to indiscriminately averaging pixels from the inhomogeneous area and destroys the polarimetric properties (Lee and Pottier, 2009). This easy operation can very well
Fig. 2. Study area Kerala. 2
– 7624 – Descending 22-Aug18 INS-NOBS
Fig. 3. Study area Asssam
1.1.3. Frost filter This filter uses local image statistics to remove high-frequency noise (speckles) while preserving features (edges) by averaging less in the edge areas. It replaces the pixel of interest with a weighted sum of the values within an n � n moving kernel (Qiu et al., 2004). The despeckled pixel value is estimated using a sub-window of the processing window. The size of the sub-window varies as a function of the target local het erogeneity measured with a coefficient of variation. X Digital number ðDNÞ ¼ kαe αjtj
C-SAR
MultiSpectral Instrument 10 days
nXn
�
S1A_IW_GRDH_1SDV_20180821T004109_20180821T004134_023337_0289D5_B2B2
S2B_MSIL1C_20180822T050649_N0206_R019_T43PFL_20180822T085140 SENTINEL2B
α¼
SENTINEL1A
Product name
Table 1 Specifications of Sentinel-1 and Sentinel-2 products.
Repeat cycle 12 days
Instrument
Product type Ground Range Detected S2MSI1C
Interferometric Wide swath (IW)
Sensing date 21-Aug18
Descending
165
23337
DV (dual VV þ VH polarization)
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Acquisition mode
Pass
Track
Orbit
Polarization
V.K. Rana and T.M.V Suryanarayana
4 nσ 2
�� � σ2 I
2
k ¼ normalized constant I ¼ local mean σ ¼ local variance σ ¼ image coefficient of variation value jtj ¼ |X X0| þ | Y Y0|are present. n ¼ moving kernel size When uniform regions are filtered, the Frost filter acts as a mean filter. When high contrast regions are filtered, the filter acts as a highpass filter with rapid decay of elements away from the filter center. Thus, large uniform areas will tend to be smoothed out and speckle removed, whilst high contrast edges and other objects will retain their signal values and not be smoothed. After application of the Frost filter, the denoised images show better sharpness at the edges. 1.1.4. Lee filter Lee filter is based on the assumption that the filtered or output pixel value is a weighted sum of the reference pixel value and the mean of the values within the kernel (LeeSen, 1981). The filter removes the noise by minimizing either the mean square error or the weighted least square estimation (Qiu et al., 2004). The Lee filter utilizes the statistical dis tribution of the digital number values within the moving kernel to es timate the value of the pixel of interest. This filter assumes the normal distribution for the noise in image data. Iout ¼ ½mean� þ K½Uin
3
mean�
V.K. Rana and T.M.V Suryanarayana
– 12314
VarðxÞ mean2 σ 2 þ VarðxÞ
Variance of � is defined as:
Descending 16-Jul19 INS-NOBS
–
DV (dual VV þ VH polarization) 28113 Ascending 14-Jul19
10 days S2B_MSIL2A_20190716T042709_N0213_R133_T46RDQ_20190716T083530 SENTINEL2B
½variance within kernel� þ ½mean within kernel�2 σ2 þ 1
½mean within kernel�2 Lee’s smoothing filter is adaptive to the local statistics in an image, however, it is an isotropic adaptive filter which cannot remove noise in the edge region effectively. Lee filter is reportedly superior in its ability to preserve prominent edges, linear features, point target, and texture information. 1.1.5. Lee sigma filter This filter is based on sigma probability of the Normal distribution. The sigma (Standard Deviation) of the entire scene is first computed and then each central pixel in a moving window is replaced with the average of only those neighborhood pixels that have intensities within a fixed sigma range of the center pixel. It is well known that, in the normal distribution, the two-sigma likelihood is 0.955. The pixels outside the two-sigma range are considered outliers and ignored. pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi variation S tan dard deviation of an image ¼ mean ¼ Coefficient of Variation ¼ Sigma ðσÞ Due to the use of a fixed sigma computed for the entire scene de Leeuw and de Carvalho, 2009, found that the Lee sigma filter blurred some of the low-contrast edges and linear features.
MultiSpectral Instrument
12 days S1A_IW_GRDH_1SDV_20190714T115653_20190714T115718_028113_032CCD_F972 SENTINEL1A
C-SAR
Ground Range Detected S2MSI1C
Interferometric Wide swath (IW)
41
Polarization Orbit Pass Acquisition mode Product type Repeat cycle
Instrument
K¼
VarðxÞ ¼
Sensing date
Track
Iout ¼ filtered output Uin ¼ unfiltered input mean ¼ average of pixels in a moving kernel
Product name
Table 2 Specifications of Sentinel-1 and Sentinel-2 products.
Remote Sensing Applications: Society and Environment 16 (2019) 100271
1.1.6. Median This filter is not an adaptive filter as it does not account for the particular speckle properties of the image. Destructive and constructive interferences in SAR information are represented by extreme values (low-value and high-value pixels), which are efficiently suppressed by the Median filter (Sheng and Xia, 1996; Qiu et al., 2004). The median filter is successful at removing pulse and spike noise while retaining step and ramp functions. Therefore, the median filter is better than the mean filter in terms of preserving the edges between two different features, but it does not preserve single pixel-wide features, which will be altered if speckle noise is present. Median filter preserves the texture informa tion very well for small window size (3 � 3) but does not retain the mean value at an acceptable level. Since the median is less sensitive than the mean to extreme values (outliers), those extreme values are more effectively removed. 1.2. Measuring performance efficiency of SAR speckle filters A speckle suppression filter is expected to filter the homogeneous areas with reasonable speckle reduction. A good SAR despeckling technique should have the following characteristics (i) scene feature preservation (such as texture, linear features, and point features) (ii) radiometric preservation (iii) speckle-noise reduction, smoothing, blur reduction, and edge preservation. The evaluation of the performance of the filters in de-speckling the SAR image is, therefore, necessary. Selected Homogeneous area, linear feature and Edge from Kerala SAR data is shown in Fig. 1. Following are the parameters to evaluate the performances of a despeckling filter:
4
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 4. Methodology.
Fig. 6. Speckle suppression index.
Fig. 5. Mean square error.
1.2.1. Mean square error (MSE) Mean square error (MSE) is the measurement of the difference be tween the output image and the input image. Higher the value of MSE, higher is the dissimilarity between the unfiltered image and the filtered image. A lower MSE value represents better image quality of the filtered image (Senthilnath et al., 2013). MSE based measurements, however, yield little information about the preservation of specific features as it assesses the whole image. MSE ¼
n 1 hX Iu K i¼1
If
Iu ¼ unfiltered image If ¼ filtered image K ¼ total number of pixels 1.2.2. Speckle suppression index (SSI) The ability of a filter to suppress speckles is measured in terms of the standard deviation of the image to its mean intensity. For homogeneous � � pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi� areas, the ratio VarðIu Þ meanðI Þ is regarded as the measurement u
�2 i
of speckle strength. The speckle suppression index (SSI) is the coefficient of variance of the filtered image standardized by that of the unfiltered
5
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 7. Speckle mean preservation index.
Fig. 10. Percent change in standard deviation.
1.2.3. Speckle mean preservation index (SMPI) SSI is not accurate when the mean value is overestimated due to the existence of extreme values in a relatively lower region of the image. In addition to SSI, therefore, SMPI (Speckle Suppression and Mean Pres ervation Index) is used to evaluate the filter efficiency (Wang and Ge, 2012). In terms of mean conservation and noise removal, lower SMPI values show better filter efficiency (Shamsoddini et al., 2010). ! qffiffiffiffiffiffiffiffiffiffiffiffiffi�ffiffi. ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi p SMPI ¼ Q � Var If VarðIu Þ where Q is calculated as under: � �� �Q ¼ 1 þ meanðIu Þ mean If � 1.2.4. Equivalent number of looks (ENL) Another commonly used evaluation criterion is the equivalent number of looks (ENL), also known as measure of the signal-to-noise ratio. This index is calculated using the following equation: � � �2 mean If � ENL ¼ standard deviation If
Fig. 8. Equivalent number of looks.
Higher ENL value for a filter represents higher efficiency in smoothing speckle-noise over homogeneous areas (Bruniquel and Lopes, 1997). The performance of a filter method is evaluated by considering changes in mean and standard deviation. Ideally, the implementation of filters should not result in any change in the mean of the target image, while it should reduce the standard deviation. To date, different classification algorithms, including the Support Vector Machine (SVM), Maximum Likelihood (ML), Classification and Regression Tree (CART), K-Nearest Neighbor (KNN), and Random Forest (RF), have been applied in various studies. The RF classifier is one of the most effective approaches for classification (Breiman, 2001). Various studies have been conducted using pixel-based RF algorithm for wetland vegetation mapping using high spatial resolution SAR data (Amani et al., 2017; Fu et al., 2017; Mahdianpari et al., 2017). Dumitru et al. (2015) applied the SVM classifier for the rapid mapping of damage assessment for flood in Germany in 2013 and the tsunami in Japan in 2011 using TerraSAR-X pre- and post-flood data. The key element of a quantitative accuracy assessment is the creation of a confusion matrix (Congalton, 2001; Janssen and Vanderwel, 1994; Story and Congalton, 1986). The confusion matrix is represented by a table that shows correspondence between the classification result and a reference image assigned to a particular category, which is relative to the actual category as indicated by the reference data. Senthilnath et al. (2013) used quantitative accu racy assessment for flood mapping, which was estimated from the error or confusion matrix. The overall accuracy was calculated by summing the number of correctly classified values and dividing it by the total
Fig. 9. Percent change in mean.
image, which is defined as: ! � � qffiffiffiffiffiffiffiffiffiffiffiffiffi�ffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffi. � VarðIu Þ meanðI Þ SSI ¼ Var If mean I � � f u VarðIf Þ ¼ variance of filtered image VarðIu Þ ¼ variance of the unfiltered image SSI has an inverse relationship with the suppression ability of the filter. The filtered image has lower variance because of speckle sup pression SSI smaller than 1.0 indicates efficient speckle suppression (Sheng and Xia, 1996).
6
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Table 3 Quantitative evaluation of the filters over Kerala. Kerala Percent Change – Mean
Percent Change – Standard Deviation
Filters
3� 3
MSE 5� 5
SSI 3� 3
5� 5
SMPI 3� 3
5� 5
ENL 3� 3
5� 5
3� 3
5� 5
3� 3
5� 5
Frostrowhead Gamma map Lee sigma Lee Boxcar Median
0.005 0.007 0.007 0.094 0.095 0.124
0.004 0.006 0.113 0.124 0.129 0.148
0.937 0.503 0.986 0.505 0.507 0.248
0.983 0.361 0.979 0.386 0.350 0.165
0.930 0.503 0.963 0.505 0.507 0.221
0.979 0.359 0.974 0.386 0.351 0.142
0.120 0.417 0.109 0.414 0.411 1.712
0.109 0.812 0.110 0.708 0.860 3.883
0.825 0.100 2.636 0.055 0.004 12.58
0.396 0.609 0.554 0.142 0.009 15.77
7.05 49.71 3.99 49.51 49.31 78.28
2.138 64.14 2.651 61.41 64.95 86.11
Table 4 Quantitative evaluation of the filters over Assam. Assam Percent Change – Mean
Percent Change – Standard Deviation
Filters
3� 3
MSE 5� 5
SSI 3� 3
5� 5
SMPI 3� 3
5� 5
ENL 3� 3
5� 5
3� 3
5� 5
3� 3
5� 5
Frost Gamma map Lee sigma Lee Boxcar Median
0.007 0.011 0.004 0.011 0.011 0.015
0.004 0.030 0.003 0.030 0.031 0.043
0.769 0.757 0.964 0.757 0.757 0.668
0.886 0.593 0.953 0.594 0.593 0.431
0.763 0.756 0.968 0.757 0.757 0.625
0.874 0.590 0.976 0.592 0.593 0.379
0.297 0.306 0.189 0.306 0.307 0.394
0.224 0.499 0.193 0.498 0.499 0.945
0.924 0.122 0.292 0.062 0.000 7.085
1.579 0.539 2.155 0.270 0.001 13.26
23.82 24.38 3.28 24.34 24.30 37.95
12.75 41.00 2.63 40.79 40.70 62.60
number of values. Producer’s accuracy is the probability that value in a given class was correctly classified.
the subtle structures of the image is essential, the efficiency of noise suppression must be balanced with the effectiveness of the filter in order to keep fine detail. The most preferred speckle filters are, therefore, assessed in the current study over the data from Sentinel-1, intended for Flooded area properly identified in a classification method Producer’s accuracy ¼ flood mapping applications. The Sentinel-1 (VV-vertical transmit, ver Flooded area in the reference ground truth tical receive and VH-vertical transmit, horizontal receive) polarised filtered data were later used for performance evaluation of machine User’s accuracy is the probability that a value predicted to be in a learning algorithms, namely, random forest (RF) and support vector certain class is really in that class. machine (SVM), to classify an inundated area. The accuracies of the Flooded area properly identified in a classification method classifications were assessed by the confusion matrix parameters, which User’s accuracy ¼ Total flooded area calculated from the method include kappa coefficient, overall accuracies, producer’s and user’s accuracies. The kappa coefficient measures the agreement between classification and truth-values. A kappa value of 1 represents perfect agreement, while 2. Study area and datasets a value of 0 represents no agreement Kappa coefficient ¼
2.1. a. Kerala
Observed accuracy Expected agreement 1 Expected agreement
Kerala is a small, elongated coastal state in peninsular India’s southwestern tip. It is surrounded by the Western Ghats in the east and the Arabian Sea in the west. A part of the state of Kerala was considered in this study. The state faces severe and varied damages due to floods and heavy rainfall. Monsoon circulation dominates the climate of India and Kerala in particular. The wind blows from the oceans to the south of the Asian land masses during the half of the year, while a seasonal wind blows from the Asian land masses to the oceans in the south during the other half of the year causing a spectacular reversal of pressure and wind patterns between the two six-month periods. South-west monsoon (June–September) and post-monsoon (October–November) are the main rainy seasons in Kerala. The state witnessed heavy floods in the year 1924 and 1961. The IMD recorded rainfalls for 15 to 17 August 2018 were found to be comparable to the rigorous storm that occurred in 1924 (Central Water Commission, 2018). Heavy rainfall resulted in high surface runoff in Kerala’s major river basins, filling all dams and sub sequent opening of these dams, causing widespread flooding in down stream areas, low-lying coastal areas, and Kerala’s backwaters. Fig. 2 shows the area covered under the study. The National Aeronautics and Space Administration Alaska Satellite Facility (NASA/ASF) houses a complete archive of Sentinel-1 SAR data processed by the European Space Agency (ESA). The Sentinel-1 Level-1
where Expected agreement ¼
Ct * ft þ nct * nft ðAÞ2
Ct ¼ actual total flooded area ft ¼ actual total flooded area nct ¼ total non-flooded area from a classified method nft ¼ actual non-flooded total area A ¼ total area under the study The objective of this study was to develop an approach for opera tional flood extent mapping. Despeckling is an important preprocessing step as speckle complicates the interpretation of visual images making automated digital image classification a challenging task. The choice of filter relies on the particular application requirements and the data set characteristics used. Despeckle filters with excellent noise extraction capacities often appear to degrade the spatial and radiometric accuracy of the actual image and trigger image reflection deterioration. This may be acceptable for applications involving large-scale interpretation or mapping of images. However, in many cases where the preservation of 7
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 11. Visual comparison of de-noising methods on VV polarization over Kerala test data (a) Boxcar 3 � 3 (b) Frost 3 � 3 (c) Gamma map 3 � 3 (d) Lee 3 � 3 (e) Lee sigma 3 � 3 (f) Median 3 � 3 (g) Boxcar 5 � 5 (h) Frost 5 � 5 (i) Gamma map 5 � 5 (j) Lee 5 � 5 (k) Lee sigma 5 � 5 (l) Median 5 � 5 and (m) Unfiltered test image in VV polarization.
ground range detected (GRD) data acquired in interferometric wide swath (IW) mode, which is the predefined mode over land with VV and VH polarizations, were downloaded via the ASF application program ming interface (API). Sentinel-1 and Sentinel-2 data that are available closest to event date were acquired on 21 August 2018 at 00:40:44 and 22 August 2018 at 05:06:49, respectively. The specific parameters of the Sentinel-1 and Sentinel-2 products are given in Table 1.
owing to the increase in water concentrations of the Brahmaputra river and its related tributaries likely led from high continuous rainfall in the upper catchment regions of the Brahmaputra Basin. Fig. 3 shows the area covered under the study. Sentinel-1 and Sentinel-2 data that are available closest to event date were acquired on 14 July 2019 at 11:57:18 and 16 July 2019 at 04:27:09, respectively. The specific pa rameters of the Sentinel-1 and Sentinel-2 products are given in Table 2. 3. Methodology
2.2. b. Assam
3.1. Pre-processing
Assam is a state in northeast India, situated south of the eastern Himalayas along the Brahmaputra and Barak River valley. The state has recently witnessed heavy flood in July 2019. The Brahmaputra basin falls within the monsoon rainfall regime, getting an average rainfall of about 230 cm. The heavy floods in the Brahmaputra river in Assam
A schematic of the Sentinel-1-based processing chain is outlined in Fig. 4. The downloaded Sentinel-1 Leve-1 GRD data acquired in IW with VV and VH polarizations were loaded onto Sentinel Application 8
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 12. Visual comparison of de-noising methods on VV polarization over Assam test data (a) Boxcar 3 � 3 (b) Frost 3 � 3 (c) Gamma map 3 � 3 (d) Lee 3 � 3 (e) Lee sigma 3 � 3 (f) Median 3 � 3 (g) Boxcar 5 � 5 (h) Frost 5 � 5 (i) Gamma map 5 � 5 (j) Lee 5 � 5 (k) Lee sigma 5 � 5 (l) Median 5 � 5 and (m) Unfiltered test image in VV polarization.
Fig. 13. Point target scattering on (a) Boxcar 3 � 3 (b) Lee 3 � 3 (c) Boxcar 5 � 5 and (d) Lee 5 � 5.
Platform (SNAP). SNAP offers a wide range of tools and features for Sentinel-1 imagery processing and analysis. Due to the large swath width of the Sentinel-1 data, the image was first divided into a subset for
the study sites to reduce the processing time. Multi-looking was then performed to reduce the standard deviation of the noise. The number of Azimuth looks and the number range of looks (2 � 2) with mean GR 9
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 14. Kerala (a) Raw VH amplitude data; (b) Raw VV amplitude data; (c) Multi-looked, calibrated, Filtered (Lee) VH data in dB; (d) Multi-looked, calibrated, Filtered (Lee) VV data in dB; (e) The Range-Doppler terrain corrected VH data; (f) The Range-Doppler terrain corrected VV data; (g) Histogram for Sigma0 VH in dB, and (h) Histogram for Sigma0 VV in dB.
mean pixel of 20 m were applied to a 1 m � 5 m (single look). The multilooked data were then calibrated to transform the pixel values from the digital values recorded by the sensor into backscatter coefficient values or Sigma0 (σ0). This was achieved using the following equation:
σ0 ¼
jDNi j Ai 2
DNi ¼ pixel’s digital number Ai ¼ absolute calibration constant 3.2. Application of filters
2
Speckles inherently corrupt all radar images, degrading the image quality, and making it more difficult to interpret features. Thus, it is 10
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 15. Assam (a) Raw VH amplitude data; (b) Raw VV amplitude data; (c) Multi-looked, calibrated, Filtered (Lee) VH data in dB; (d) Multi-looked, calibrated, Filtered (Lee) VV data in dB; (e) The Range-Doppler terrain corrected VH data; (f) The Range-Doppler terrain corrected VV data; (g) Histogram for Sigma0 VH in dB, and (h) Histogram for Sigma0 VV in dB. Table 5 Comparison of user’s accuracy (UA), producer’s accuracy (PA), overall accuracy (%), and kappa coefficient using random forest tree and support vector machine algorithms for VV and VH polarization over Kerala region. VV Polarization RF Inundation Rest Kappa coefficient Overall Accuracy (%)
VH Polarization SVM
RF
SVM
PA
UA
PA
UA
PA
UA
PA
UA
0.89 0.88 0.72
0.96 0.72
0.79 0.98 0.64
0.99 0.60
0.78 0.98 0.61
0.99 0.58
0.79 0.98 0.63
0.99 0.60
88.80%
Table 6 Inundated area statistics of RF and SVM over cloud-free optical data for Kerala region.
83.80%
82.60%
Inundated area (Km2) Rest (Km2)
VV Polarization
VH Polarization
RF
SVM
RF
SVM
NDWI –
40.25 16.29
34.82 21.72
34.17 22.38
35.19 21.36
41.78 14.77
kernel size, used in the study were available in SNAP and applied using default system parameters.
83.60%
3.3. Evaluating the performance efficiency of filters
Note: UA - User’s accuracy, PA - Producer’s accuracy VV - Vertical-Vertical, VH Vertical-Horizontal, RF - Random Forest, SVM -Support Vector Machine.
Several techniques are available to quantitatively assess the effi ciency of a speckle filter in distinct ways, for example, edge conservation and conservation of features. The findings of the various measurements may be contradictory. Therefore, distinct techniques of evaluation should be used to discover the optimum tradeoff between the various
often necessary to enhance the image by filtering speckles before data can be used in different applications. All of the filters, namely, Boxcar, Median, Frost, Gamma map, Lee and Lee sigma with 3 � 3 and 5 � 5 11
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
producer’s accuracy, user’s accuracy, and the overall accuracy of the classifiers. The overall accuracy gives the correctly classified regions for the image and is calculated by the proportion of the correctly classified pixels to the total number of pixels in the confusion matrix. To calculate inundation for entire scene, thresholding was done in SNAP by using the conditional function given below after carefully analysing the histogram:
Table 7 Comparison of user’s accuracy (UA), producer’s accuracy (PA), overall accuracy (%), and kappa coefficient using random forest tree and support vector machine algorithms for VV and VH polarization over Assam region. VV Polarization RF Inundation Rest Kappa coefficient Overall Accuracy (%)
VH Polarization SVM
RF
SVM
PA
UA
PA
UA
PA
UA
PA
UA
0.77 1.0 0.65
1.0 0.62
0.82 1 0.72
1.0 0.69
0.81 0.99 0.70
0.99 0.67
0.89 0.99 0.81
0.99 0.77
83.60%
87.60%
86.60%
If Sigma0 VV/VH db < X then 1 else 0 X ¼ threshold value 1 ¼ inundated pixels 0 ¼ Rest
92.00%
Note: UA - User’s accuracy, PA - Producer’s accuracy VV - Vertical-Vertical, VH Vertical-Horizontal, RF - Random Forest, SVM -Support Vector Machine.
The resulted output was later used to calculate the difference in inundated areas calculated by thresholding technique and images clas sified using the random forest and support vector machine algorithms having highest accuracy for both the study areas.
elements of the image. The sigma0 values of VV and VH polarised data with the applied filter was terrain corrected using SNAP’s ‘Range Doppler Terrain Correction’ algorithm with an SRTM 1 ArcSecond digital elevation model. The bilinear interpolation method was used for DEM and Image resampling with a pixel spacing of 20 � 20 m. Terrain correction helps in improving the geometric representation of the real-world surface. This is needed because, during image capture, topographical variations and off-nadir distortion unsettle the image.
4. Results and discussion 4.1. Speckle filtering The performance efficiency of filters in speckle suppression, feature preservation, and preventing the loss of meaningful data was evaluated using MSE, ENL, SSI, SMPI, examination of mean, standard deviation, and also by close visual assessment. The MSE value indicates the amount of error present in a filtered image. It allows the comparison of the pixel values of a filtered image to the degraded image before filtering. It is observed that for study area Kerala (window size 3 � 3) Frost, Gamma map, and Lee sigma filters showed very low MSE, indicating their effectiveness in feature preser vation. Median filter showed the highest MSE values indicating poor performance in terms of feature preservation. Boxcar and Lee filters performed moderately. For Kerala (window size 5 � 5) and Assam (window size 3 � 3 and 5 � 5), Frost and Lee sigma showed low MSE, Median filter showed the highest MSE and moderate performance was observed for Boxcar, Gamma map and Lee (Fig. 5). The high value of MSE depicts a greater difference between the original test image and despeckled image which concludes the significant speckle reduction but at the cost of feature loss. Lower SSI and SMPI values were obtained for Median filter (3 � 3 and 5 � 5) in both the study areas indicating high efficiency in speckle suppression. Moderate values were obtained for Gamma map, Lee and Boxcar indicating medium proficiency in the reduction of speckle. None of the filters produced value 0 and 1. A value of 0, indicates complete mean preservation and 100% noise reduction. Whereas, a value of 1 indicates 0% speckle reduction. It was observed that as the window size increased from 3 � 3 to 5 � 5, the SSI values reduced for all the filters (Figs. 6 and 7), indicating superior speckle suppression with an increase in the size of the moving window. The lowest SSI values, as well as SMPI values, corresponds to Median filter (3 � 3 and 5 � 5) indicating high efficiency in speckle suppression. For both the study areas Median filter (3 � 3 and 5 � 5) had high ENL values indicating a higher efficiency in smoothing speckle noise over homogeneous areas, which shows enhanced capacity to distinguish the distinct gray levels within the image. For constant flat areas where the sample variance is null, ENL become ∞, this will repute highly blurred data as excellent. According to ENL, SSI and SMPI values, Gamma map (3 � 3 and 5 � 5), Lee (3 � 3 and 5 � 5) and Boxcar (3 � 3 and 5 � 5) filters showed moderate performance for both the study areas (Fig. 8). The application of filters should ideally not bring about any change in the mean of target image while it should reduce the standard devia tion. The filter, Median (3 � 3 and 5 � 5) for both the study areas was most effective in increasing the standard deviation but also changed the mean value considerably thereby implying that the filters reduced
3.4. Machine learning algorithms for classification The terrain corrected images were classified using the random forest and support vector machine algorithms as a next step. For both the classifiers, the same number of training samples was used. The training inundated pixels covered 5.2 Km2 and the rest of the training pixels covered 3.1 Km2 of the study area Kerala. Similarly, for the study area Assam, the training inundated pixels covered 11 Km2 and the rest of the training pixels covered 54 Km2. During the southwest monsoon season, it is nearly impossible to obtain 100% cloud-free data, however, a small extent of the cloud-free data can be used for validation. The normalized difference water index (NDWI) is defined for Sentinel–2 data as ((B03) (B08)/(B03) þ (B08)), where B03 is a green band and B08 is the near-infrared band (Mcfeeters, 1996). When NDWI is applied over a multispectral image, the water feature has positive values, while soil and terrestrial vegeta tion features have zero or negative values. This is because NIR is absorbed strongly by water but reflected strongly by terrestrial vegeta tion and dry soil, while in a green light, water has high reflectance than terrestrial vegetation and soil. Therefore, the NDWI was applied to extract water from the optical data. A cloud-free part of satellite optical image was collected by Sentinel-2 at 05:06:49 on 22 August 2018, 28 h after the Sentinel-1 pass over the study area Kerala. Similarly, a cloud-free part of satellite optical image was collected by Sentinel-2 at 04:27:09 on 15 July 2019, 40 h after the Sentinel-1 pass over the study area Assam. Sentinel-2 data were converted to reflectance and dark object subtraction atmospheric correction (DOS1) was applied. The corrected Sentinel-2 image was used to validate the extent of the flood. The normalized difference water index (NDWI), established earlier (Mcfeeters, 1996) to extract the water from the optical data, was calculated as: NDWI ¼
ρGreen ρNIR ρGreen þ ρNIR
Moreover, stratified random sampling was used with 500 points for the accuracy assessment. The classification maps were evaluated in terms of their overall accuracy (OA), producer’s accuracy (PA), user’s accuracy (UA), and the kappa index of agreement (k) or kappa coeffi cient. Confusion matrix was created to compare the kappa coefficient, 12
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 16. Kerala (a) Random forest tree classification on filtered VH; (b) Support vector machine classification on VH; (c) Random forest tree classification on filtered VV; (d) Support vector machine classification on VV.
13
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 17. Assam (a) Random forest tree classification on filtered VH; (b) Support vector machine classification on VH; (c) Random forest tree classification on filtered VV; (d) Support vector machine classification on VV.
speckle but also caused considerable loss of meaningful data. The me dian filter is better than the Boxcar filter in terms of preserving the edges between two different features, but it does not preserve single pixel-wide features, which will be altered if speckle noise is present. Median filter preserves the texture information very well for small window size (3 � 3) but does not retain the mean value at an acceptable level. Since the median is less sensitive than the mean to extreme values (outliers), those extreme values are more effectively removed. On the contrary, Boxcar (3 � 3 and 5 � 5) for both the study areas, made the least change in the mean, while reducing the standard deviation moderately. How ever, Lee filter (3 � 3 and 5 � 5) provided a fair balance by reducing the standard deviation without drastically affecting the mean (Figs. 9 and 10, and Tables 3 and 4). The Median filter performed well in terms of ENL, SSI, and SMPI values; however, its performance in terms of speckle reduction and feature preservation was far inferior compared to the Boxcar and Lee filter. However, since the objective of the speckle filtering was to use these SAR images for inundation mapping, the performance of the filters on a water body in terms of reduction of standard deviation while pre serving the mean of the original image was the most important. The
performance of Boxcar and Lee filters was far better in feature preser vation in the filtered images followed by Gamma map. Although quan titative measures are often employed to compare different speckle suppression filters, it has been noted by Raouf and Lichtenegger (1997) and others that visual inspection probably provides the best assessment of the performance of a speckle filter. Visual assessment is an easy and efficient way to investigate both the capability of a filter to suppress speckles and its effectiveness in preserving image details. Lee et al. (1994) stated that, in general, filters using small windows (such as 3 � 3) preserve texture information better. Visual examination was, therefore, carried out and it was observed that the filters which reduced speckles effectively also resulted in considerable loss of meaningful data (Figs. 11 (a)-(m) and 12 (a)-(m)). Lee sigma, and Gamma map clearly resulted in the loss of edges and details. It was difficult to grade Boxcar, Frost and Lee filter visually, since the variation was not perceivable. However, it was observed that point scatters were over filtered, transformed to spread targets and sharp edges were generally blurred in Boxcar filter (Fig. 13). We chose Lee filter for further analysis as it had low MSE, SSI and SMPI values, and a higher percentage change in standard deviation compared to the Boxcar filter for most of the cases. 14
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 18. (a) Natural colour composite R–B04 G-B03 B–B02, (b) Green band (c) Near-infrared (d) Calculated NDWI over cloud-free extent for Kerala region.
The raw SAR data in VH and VV polarization acquired during the crisis events on 21 August 2018 and 14 Jul 2019 for study areas Kerala and Assam are shown in Fig. 14 (a) and (b), and Fig. 15 (a) and (b), respectively. The multi-looked, calibrated, filtered (Lee) data of the VH and VV are shown in Fig. 14 (c) and (d), and Fig. 15 (c) and (d), respectively. Multi-looked, calibrated, filtered (Lee) SAR data were not projected on the map coordinates of each pixel. The pixel was in the original coordinate position of data (rows/columns) in the field of ground range. In the orthorectified imageries, each of the pixels that were corrected and projected using the Range-Doppler terrain correc tion appeared at the actual position. The Range-Doppler terrain cor rected pixels and their respective histograms for Kerala and Assam regions are shown in Fig. 14 (e)–(h), and Fig. 15 (e)- (h), respectively. Tables 5 and 7, and Figs. 16 and 17 show the comparison of classi fication results of random forest classifier and support vector machine classifier for VV and VH polarization. The training data were kept the same for both the classifiers to avoid optimistic bias in the classification. For the study area Kerala, the random forest classifier exhibited maximum overall accuracy of 88.80% with the kappa coefficient value of 0.72. Both the classifiers obtained better accuracy results in VV po larization compared to the VH polarization. The least overall accuracy of 82.60% and a kappa coefficient value of 0.63 were observed with random forest in VH polarization, which was followed by the support vector machine in VV polarization. RF achieved higher classification accuracy than SVM by about 5% in VV polarization. However, both the
classifiers produced comparable overall accuracies in VH polarization (SVM achieved higher classification accuracy than RF by about 1%). The NDWI calculated for the cloud-free extent is shown in Fig. 18 (d). The inundated area in the calculated NDWI over the cloud-free extent is 73.88%, which is 41.78 km2. However, it has also been observed (Table 6) that the inundated area using random forest classification on filtered VV data over the cloud-free extent is 71.18%, which is 40.25 Km2. For the study area Assam, the SVM classifier exhibited maximum overall accuracy of 92% with the kappa coefficient value of 0.81. Both the classifiers obtained better accuracy results in VH polari zation compared to the VV polarization. The least overall accuracy of 83.60% and a kappa coefficient value of 0.65 were observed with random forest in VV polarization, which was followed by the RF in VH polarization. SVM achieved higher classification accuracy than RF by about 5.38% in VH polarization. The NDWI calculated for the cloud-free extent is shown in Fig. 19 (d). The inundated area in the calculated NDWI over the cloud-free extent is 74.09%, which is 491.47 km2. However, it has also been observed (Table 8) that the inundated area using SVM classification on filtered VH data over the cloud-free extent is 62.76%, which is 416.99 Km2. To calculate inundation for the entire scene, threshold value of 10.96 and 19.58 was selected for Kerala and Assam region respectively after analysing the histograms. The calculated inundated area (Fig. 20) with thresholding technique for Kerala was found 204 km2 (2% more than classified VV polarised data using RF algorithm). Similarly, the calculated inundated area with 15
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 19. (a) Natural colour composite R–B04 G-B03 B–B02, (b) Green band (c) Near-infrared (d) Calculated NDWI over cloud-free extent for Assam region.
balance in feature preservation as well in despeckling compared to the other filters used in the study. The accuracy assessment of machine learning algorithms for flood classification over Kerala shows that random forest classifier has higher classification accuracy than the support vector machine (about 5% higher in VV polarization). However, both the classifiers performed better in VV polarization than VH polar ization. For study area Assam, SVM in VH polarization achieved higher accuracy and least performance was observed by RF in VV polarization. The proposed approach shows the potential for monitoring damages caused by floods, providing basic information that can help local com munities manage water-related risk, planning land and water manage ment as well as other flood control programs.
Table 8 Inundated area statistics of RF and SVM over cloud-free optical data for Assam region.
Inundated area (Km2) Rest (Km2)
VV Polarization
VH Polarization
RF
SVM
RF
SVM
NDWI –
363.92 300.41
389.21 275.12
387.62 276.70
416.99 247.34
491.47 171.86
thresholding technique for Assam was found 3368.90 km2 (23.46% more than classified VH polarised data using SVM algorithm). For Assam region the variation is very large, a single threshold did not hold well as large swath of a SAR image suffers from environment heterogeneity caused by wind-roughening and satellite framework parameters.
Ethical approval The presented research has not involved human subjects, human material, or human data and therefore ethical approval has not been required.
5. Conclusion Effective and quick response is required during disasters like flood ing. Rapid mapping of such events will be beneficial to urban and infrastructure planners, risk managers and disaster responses during extreme and intense rainfall events. The study shows a simple and efficient method for mapping inundation extent with only the C-band S1A, with coarser geometric resolution and fixed polarizations (VV-VH) by considering the case of Kerala and Assam. Backscattering coefficient values become high as the water roughness causes high signal return, decreasing the contrast and making the separation of the land-water covers difficult. Despeckle filters with good noise removal capabilities often tend to degrade the spatial and radiometric resolution of an original image and cause the loss of image details. This may be acceptable for applications involving large scale image interpretation or mapping. However, the retention of the subtle structures of an image is important and, therefore, the performance of speckle noise suppression technique must be balanced with the filter’s effectiveness to preserve the fine details. The performance evaluation of de-noising methods in this study showed that Lee filter with 3 � 3 kernel size provided a good
Compliance with ethical standards This study is in full compliance with all applicable ethical standards. Informed consent This study does not involve any subject participation and therefore informed consent is not applicable. Declaration of competing interest On behalf of all authors, the corresponding author states that there is no conflict of interest.
16
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Fig. 20. (a) Random forest tree classification on filtered VV data (Kerala) (b) Classified filtered VV data with threshold value of machine classification on filtered VH data (Assam) (d) Classified filtered VH data with threshold value of 19.58 (Assam).
References
10.96 (Kerala) (c) Support vector
Fu, B., Wang, Y., Campbell, A., Li, Y., Zhang, B., Yin, S., et al., 2017. Comparison of object-based and pixel-based Random Forest algorithm for wetland vegetation mapping using high spatial resolution GF-1 and SAR data. Ecol. Indicat. 73, 105–117. https://doi.org/10.1016/j.ecolind.2016.09.029. Gitas, I.Z., Polychronaki, A., Katagis, T., Mallinis, G., 2008. Contribution of remote sensing to disaster management activities: a case study of the large fires in the Peloponnese, Greece. Int. J. Remote Sens. 29 (6), 1847–1853. https://doi.org/ 10.1080/01431160701874553. Goodman, J.W., 1976. Some fundamental properties of speckle. JOSA 66 (11), 1145–1150. Janssen, L.L., Vanderwel, F.J., 1994. Accuracy assessment of satellite derived land-cover data: a review. Photogramm. Eng. Remote Sens. 60 (4). Khan, S.D., 2005. Urban development and flooding in Houston Texas, inferences from remote sensing data using neural network technique. Environ. Geol. 47 (8), 1120–1127. https://doi.org/10.1007/s00254-005-1246-x. Lee, J.S., Pottier, E., 2009. Polarimetric Radar Imaging: from Basics to Applications. CRC press. Lee, J.S., Jurkevich, I., Dewaele, P., Wambacq, P., Oosterlinck, A., 1994. Speckle filtering of synthetic aperture radar images: a review. Remote Sens. Rev. 8 (4), 313–340. https://doi.org/10.1080/02757259409532206. Lee, J. Sen, 1981. Speckle analysis and smoothing of synthetic aperture radar images. Comput. Graph. Image Process. 17 (1), 24–32. https://doi.org/10.1016/S0146-664X (81)80005-6. Lopes, A., Nezry, E., Touzi, R., Laur, H., 1990. May). Maximum a posteriori speckle filtering and first order texture models in SAR images. In: 10th Annual International Symposium on Geoscience and Remote Sensing, pp. 2409–2412 (Ieee). Mahdianpari, M., Salehi, B., Mohammadimanesh, F., Brisco, B., 2017. An assessment of simulated compact polarimetric SAR data for wetland classification using random forest algorithm. Can. J. Remote Sens. 43 (5), 468–484. https://doi.org/10.1080/ 07038992.2017.1381550. Mcfeeters, S.K., 1996. The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. Int. J. Remote Sens. https://doi.org/10.1080/ 01431169608948714.
India Meteorological Department, 2018. Rainfall over Kerala during Monsoon Season2018 and forecast for next 5 days. Retrieved from. https://www.imdtvm.gov. in/images/rainfall. over kerala during monsoon season-2018 and forecast for next 5 days.pdf. Amani, M., Salehi, B., Mahdavi, S., Granger, J., Brisco, B., 2017. Wetland classification in Newfoundland and Labrador using multi-source SAR and optical data integration. GIScience Remote Sens. 54 (6), 779–796. https://doi.org/10.1080/ 15481603.2017.1331510. Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5–32. https://doi.org/ 10.14923/transinfj.2015IUP0008. Bruniquel, J., Lopes, A., 1997. Multi-variate optimal speckle reduction in SAR imagery. Int. J. Remote Sens. 18 (3), 603–627. https://doi.org/10.1080/014311697218962. Central Water Commission, 2018. Kerala Floods August 2018. Retrieved from. https:// reliefweb.int/sites/reliefweb.int/files/resources/Rev-0.pdf. Chunming, H., Huadong, G., Yun, S., Jingjuan, L., 2005, July. Detection of the flood boundary in SAR image using texture. In: Proceedings. 2005 IEEE International Geoscience and Remote Sensing Symposium, vol. 5. IGARSS’05, pp. 3697–3699, 2005, (Ieee). Congalton, R.G., 2001. Accuracy assessment of wildland fire and other spatial information. Int. J. Wildland Fire 10, 321–328. de Leeuw, M.R., de Carvalho, L.M.T., 2009. Performance evaluation of several adaptive speckle filters for SAR imaging. Anais XIV Simp� osio Brasileiro de Sensoriamento Remoto 7299–7305. Dumitru, C.O., Cui, S., Faur, D., Datcu, M., 2015. Data analytics for rapid mapping: case study of a flooding event in Germany and the tsunami in Japan using very high resolution SAR images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 8 (1), 114–129. https://doi.org/10.1109/JSTARS.2014.2320777. Durand, J.M., Gimonet, B.J., Perbos, J.R., 1987. SAR data filtering for classification. IEEE Trans. Geosci. Remote Sens. (5), 629–637. https://doi.org/10.1109/ TGRS.1987.289842. GE-25. Frost, V.S., Stiles, J.A., Shanmugam, K.S., Holtzman, J.C., 1982. A model for radar images and its application to adaptive digital filtering of multiplicative noise. IEEE Trans. Pattern Anal. Mach. Intell. 4 (2), 157–166.
17
V.K. Rana and T.M.V Suryanarayana
Remote Sensing Applications: Society and Environment 16 (2019) 100271
Nirupama, N., Simonovic, S.P., 2007. Increase of flood risk due to urbanisation: a Canadian example. Nat. Hazards 40 (1), 25–41. https://doi.org/10.1007/s11069006-0003-0. Qiu, F., Berglund, J., Jensen, J.R., Thakkar, P., Ren, D., 2004. Speckle noise reduction in SAR imagery using a local adaptive median filter. GIScience Remote Sens. 41 (3), 244–266. https://doi.org/10.2747/1548-1603.41.3.244. Raouf, A., Lichtenegger, J., 1997. Integrated use of SAR and optical data for coastal zone management. In: Third ERS Symposium on Space at the Service of Our Environment, vol. 414, p. 1089. Senthilnath, J., Shenoy, H.V., Rajendra, R., Omkar, S.N., Mani, V., Diwakar, P.G., 2013. Integration of speckle de-noising and image segmentation using Synthetic Aperture Radar image for flood extent extraction. J. Earth Syst. Sci. 122 (3), 559. https://doi. org/10.1007/s12040-013-0305-z.
Serpico, S.B., Dellepiane, S., Boni, G., Moser, G., Angiati, E., Rudari, R., 2012. Information extraction from remote sensing images for flood monitoring and damage evaluation. Proc. IEEE 100 (10), 2946–2970. Shamsoddini, A., Trinder, J.C., Wagner, W., Sz� ekely, B., 2010. Image Texture Preservation in Speckle Noise Suppression (na). Sheng, Y., Xia, Z.G., 1996. May). A comprehensive evaluation of filters for radar speckle suppression. In: IGARSS’96. 1996 International Geoscience and Remote Sensing Symposium, vol. 3. IEEE, pp. 1559–1561. Story, M., Congalton, R.G., 1986. Accuracy assessment: a user’s perspective. Photogramm. Eng. Remote Sens. 52 (3), 397–399. Wang, X., Ge, L., 2012. Evaluation of Filters for ENVISAT ASAR Speckle Suppression in Pasture Area. Zhang, F., Zhu, X., Liu, D., 2014. Blending MODIS and Landsat images for urban flood mapping. Int. J. Remote Sens. 35 (9), 3237–3253. https://doi.org/10.1080/ 01431161.2014.903351.
18