COMPARISON BETWEEN SPECTRAL , SPATIAL AND POLARIMETRIC CLASSIFICATION OF URBAN AND PERIURBAN LANDCOVER USING TEMPORAL SENTINEL – 1 IMAGES

Landcover is the easiest detectable indicator of human interventions on land. Urban and peri-urban areas present a complex combination of landcover, which makes classification challenging. This paper assesses the different methods of classifying landcover using dual polarimetric Sentinel-1 data collected during monsoon (July) and winter (December) months of 2015. Four broad landcover classes such as built up areas, water bodies and wetlands, vegetation and open spaces of Kolkata and its surrounding regions were identified. Polarimetric analyses were conducted on Single Look Complex (SLC) data of the region while ground range detected (GRD) data were used for spectral and spatial classification. Unsupervised classification by means of K-Means clustering used backscatter values and was able to identify homogenous landcovers over the study area. The results produced an overall accuracy of less than 50% for both the seasons. Higher classification accuracy (around 70%) was achieved by adding texture variables as inputs along with the backscatter values. However, the accuracy of classification increased significantly with polarimetric analyses. The overall accuracy was around 80% in Wishart H-A-Alpha unsupervised classification. The method was useful in identifying urban areas due to their double-bounce scattering and vegetated areas, which have more random scattering. Normalized Difference Built-up index (NDBI) and Normalized Difference Vegetation Index (NDVI) obtained from Landsat 8 data over the study area were used to verify vegetation and urban classes. The study compares the accuracies of different methods of classifying landcover using medium resolution SAR data in a complex urban area and suggests that polarimetric analyses present the most accurate results for urban and suburban areas. * Corresponding author

Early studies that employed radar remote sensing to investigate landuse/landcover information were mainly performed using space shuttle SIR-C/X-SAR (Pierce, Bergen, Dobson, & Ulaby, 1998;Saatchi, Soares, & Alves, 1997).Airborne radar imagery systems also provided data for landuse/landcover mapping (Yeh & Qi, 2015).However, they are only occasionally launched to collect experimental data within a very short period.Regular investigation of timely landuse/landcover information using radar remote sensing has become practicable only after successful operation of orbital radar systems with SAR.These included ERS-1 and ERS-2, JERS-1, RADARSAT-1, TerraSAR-X, RISAT-1 and Sentinel-1.These are made available for regular data collection (Yeh & Qi, 2015).
For the continuation and improvement of SAR applications, the European Space Agency (ESA) launched an European Radar Observatory consisting of a polar orbiting two-satellite constellation (Sentinel-1) (Torres et al., 2012).The satellite has a revisit period of 12 days at the equator.The near polar, circular orbit of the spacecraft at an elevation of 693 Km is sunsynchronous with an inclination of 98.18 o (Sentinel-1 Team, 2013).The centre frequency of the data captured is 5.405 GHz i.e. the data is collected in the C-band of the electromagnetic spectrum.The incidence angle varies from 20 o to 45 o .Data obtained from this satellite can meet explicit research needs relating to land, ocean and cryosphere (Malenovský et al., 2012).Sentinel-1 data has been used in different researches since the time of its availability.The data has been applied in researches on disaster management (Plank, 2014) and areas undergoing rapid changes in a wetland environment (Muro et al., 2016).This data has been widely applied for landcover mapping.It was used for generating level -2 landcover products along with temporal ERS-2 and ENVISAT ASAR AP C-band data (Thiel et al., 2009).INSAR coherence was used for landuse classification and monitoring of temporal changes (Wegmuller, Santoro, Werner, & Cartus, 2015).
The current study assesses different methods of landcover classification in an urban area using Sentinel -1 data.Unsupervised classification by K-means clustering, supervised classification by Maximum Likelihood method and Wishart H-A-Alpha polarimetric decomposition methods were examined.The major objective of this research is to identify the most appropriate method of landcover classification using multitemporal SAR data to discriminate broad landcover classes in urban and peri-urban areas of an expanding metropolis

STUDY AREA
The study is based on the city of Kolkata and its surrounding areas.Kolkata is the third largest metropolis of India with a population of 4.5 million.The city is located in the eastern banks of river Hugli.Due to large-scale migration from adjacent rural areas and immigration from the neighbouring country of Bangladesh, the city is experiencing growth for the last few decades.Due to the presence of river Hooghly on the western side of the city, Kolkata mostly experiences a north south sprawl.The city and its surrounding areas are undergoing rapid changes due to construction of new residential areas, roads and flyovers.As a result large scale agricultural lands are getting converted to non-agricultural uses.Kolkata and its surrounding areas are represented in Figure 1.

DATA USED
Sentinel-1 datasets were downloaded from the data hub of Sentinel.The data were acquired on 24th of July and 15th of December 2015.Level-1 SLC and GRD data were used in this study.The datasets were chosen to match with the wet and the dry seasons of the year.Both the datasets had dual polarization capabilities; vertical transmit and horizontal receive (VH) and vertical transmit and vertical receive (VV).Single Look Complex (SLC) products were geo-referenced using orbit and attitude data from the satellite.They were provided in zero-Doppler slant-range geometry and consisted of complex samples preserving the phase information.Ground Range Detected (GRD) products, on the other hand, were focused SAR data detected, multi-looked and projected to ground range using an earth ellipsoid model.Phase information was absent in these products.(Sentinel-1 Team, 2013).Data captured in the Interferometric Wide Swath Mode (IW) were used in this research.These datasets have a swath of 250 Km and spatial resolutions of 5 X 20 m (Sentinel-1 Team, 2013).
Landsat 8 data collected during the month of March 2015 was used to calculate NDVI and NDBI.These were used to validate the extents of urban built up areas and vegetation as obtained from the classification of SAR images.

METHOD
The method followed in this research was divided into four parts: classification of GRD data, processing of Landsat8 data, polarimetric classification of SLC data and accuracy assessment.

Processing of GRD Data
The GRD datasets of both monsoon and winter seasons were radiometrically calibrated to obtain sigma nought (σ°) or the backscatter values.The process of multilooking was used to convert the images from slant range to ground range.During this process the image pixels became approximately square spaced.Refined Lee filter with an edge threshold of 5000.00 were applied on the images to reduce the noise from random constructive and destructive interferences.Finally both the In the next step, texture elements were calculated using the Grey Level Co-occurrence Matrix (GLCM) algorithm.These included contrast, dissimilarity, homogeneity, angular second moment, mean, variance, correlation and entropy.The backscatter images were merged with the texture elements.These merged images were then classified using K-Means clustering unsupervised method.Four landcover classes (water bodies and wetlands, open land, vegetation and urban built-up areas) as previous steps were identified to facilitate comparison of results.

Processing of Landsat 8 data
A cloud free Landsat 8 imagery from March 2015 over Kolkata was downloaded from the USGS Website.The data was also used to prepare NDVI and NDBI maps of the study area.NDVI is calculated as: . (1) Where, NIR = Near Infra-Red band R = Red band For Landsat 8, this was obtained by the following ratio: (2) NDBI is calculated as (3) . For Landsat 8, this was calculated from: (4)

Processing of Single Look Complex (SLC) data
In the first step, covariance matrices (C2) were calculated for both the images.It is a measure of the power of the received signal.C2 is expressed as: (5) where, and S is the scattering matrix.
Entropy (H), anisotropy (A) and alpha (α) images were obtained for both the time periods.Entropy (H) is a measure of dominance of a certain scattering mechanism in a given resolution cell of the image.It varies from 0 to 1.A value of 0 identifies that all scattering comes from one mechanism (such as flat surface or a tall building by double bounce), while a value of 1 represents a completely random scattering mechanism.Anisotropy (A), on the other hand is a measure of homogeneity of a target relative to the radar look direction.The value of anisotropy also varies from 0 to 1.It becomes 0 if both scattering mechanisms are of an equal proportion.A value greater than 0 represents more anisotropic scattering.Thus anisotropy is complementary to entropy.However, it can be used as a source of decomposition only when H > 0. Landcover classes of the study area were identified by Wishart H-A-α classification.This method used the Entropy (H), Anisotropy (A) and Alpha (α) as inputs.Sixteen classes were initially generated which were merged to four landcover classes (water bodies and wetlands, open land, vegetation and urban built-up areas).

Accuracy Assessment
Hundred ground reference points were randomly collected from the study area for assessing the accuracy of classification.Confusion matrix, producer's accuracy and user's accuracy were calculated for all three types of classification.

RESULTS AND DISCUSSIONS
The outputs from analyses are explained using the results from accuracy assessment.They are divided under the following categories: i. Unsupervised and supervised classification of backscatter image ii.Classification using texture data iii.Polarimetric analyses and classification iv.Comparison with Landsat data

Unsupervised and Supervised Classification of Backscatter images
The overall accuracies of classification of GRD data are less than 50% for both supervised and unsupervised classifications.The user's and producer's accuracies from the classifications are presented in table 1. Producer's accuracy indicates how well the training set pixels of a given class is classified.The user's accuracy, on the other hand, indicates the probability that a pixel classified into a given class actually belongs to the said class on the ground.The urban landcover category exhibited varying accuracies in the classification.The highest user's accuracy for urban built up areas was obtained from supervised classification of the image collected on 15 th December.For the supervised classification on the 24 th July data, the user's accuracy of urban built up class was around 69%.The lowest user's accuracy for the same class was noted in the unsupervised classified image of 24 th July.However relatively low producer's accuracies were obtained for the urban class suggesting significant omission errors for this class.
Comparatively higher user's accuracies were also noted for the water landcover class.The producer's accuracy for the water class was 100% in the supervised image of 24 th July.Both the producer's and user's accuracies are low for open space landcover class indicating high omission and commission errors for this class.Overall, the classifications produced mixed results.

Classification using Texture Data:
Unsupervised classification (K-Means) was conducted on the images with texture variables.The overall accuracies increased to 59% for the winter month and 70% for the monsoon month.The producer's and user's accuracies for the classification are presented in table 2. It was observed that the user's accuracies significantly increased in this method of classification compared to the results in the section 5.1.For the month of July, the user's accuracies for water bodies, open spaces, vegetation and built up areas were respectively 60%, 80%, 65% and 76%.The same from the classification of backscatter image were 70%, 35%, 40.74% and 50% respectively.Only the accuracy for the water class decreased in the classification of the texture image.For all the other classes, there was an increase in accuracy by more than 20%.For the month of December, the user's accuracies from classification of the texture image were 80%, 68.57%, 38.46% and 75% respectively for water, open space, vegetation and urban built up areas respectively.From the backscatter image these values were 70%, 20%, 33.33% and 68.75% respectively.The accuracy for the open space class increased by more than 48.5% in the texture image classification compared to that from backscatter image.Overall, the incorporation of the texture bands increased the classification accuracies of all the landcover categories

Polarimetric Analyses and Classification
In the next step entropy, anisotropy and alpha were calculated for both the images.The inputs from entropy, anisotropy and alpha were used for Wishart H-A-α classification of both the images.The overall classification accuracies significantly increased to 79% for the monsoon month and 81% for the winter month.The user's and producer's accuracies for the classification are presented in table 3. The user's accuracy of urban built-up areas and open spaces increased to 100% for the image captured on 24 th July.This shows that pixel classified as these two classes actually belong to the said classes on the ground.The accuracy was also 100% for built up areas and 90% for open spaces in the image captured on 15 th December.The user's accuracy for vegetation was 74% for winter and 65% for the month of monsoon.The producer's accuracies were also high for all the classes.For the vegetation class, it was 100% for the monsoon month and 95% for winter.The producer's accuracy was also 100% for water bodies during monsoon.Thus the sample pixels have been well classified for these two classes.Overall, there was marked improvement in classification by polarimetric analyses.

Comparison with Landsat Image
The NDVI and NDBI images obtained from Landsat 8 data were then compared with the classified Sentinel-1 products.Vegetation appears brighter in the former one while the built up areas are lighter in the latter.In figures 3 and 4, columns I show the combination of NDBI and classified SAR images while columns II represent overlapping NDVI and SAR classification.

Figure 1 :
Figure 1: Study Area: (A) Landsat 8 image (band 5, band 4, band 3); (B) 24th July, 2015 (VH, VV, VV/VH); (C) 15th December, 2015 (VH, VV, VV/VH) corrected.This process geocoded the images by correcting the SAR processing distortions using a Digital Elevation Model (DEM).In this case, SRTM 3sec DEM was used.The resultant images were projected by default to Geographic Latitude/Longitude projection (Array SystemsComputing and ESA, 2015).Kolkata and its surrounding suburban areas were subset from the orthorectified images for further analyses.The bands with VH and VV polarizations were layer stacked and classification was done on the resultant image.Backscatter values (σ o ) were used for identifying the classes.Two methods of classification were undertaken.In the first step a K-Means clustering method was used for classification(Lillesand, Kiefer, & Chipman, 2004).For the convenience of classification, a minimum of 10 and a maximum of 20 classes were initially set for the K-Means algorithm with a change threshold percentage of 99%.The resultant classes were merged to produce the final four major landcover classes: water and wetlands, open spaces, urban built up areas and vegetation.Next, a maximum likelihood supervised classification was also performed on the backscatter images.Training sample datasets were created.The images were then classified using the training samples.
7. Alpha (α) angle provides the nature of dominant scattering mechanism for a resolution cell if the entropy value is 0 for the said cell.Alpha values identify if the scattering is volume, surface or double bounce.The angle is normalized between 0 o and 90 o .Alpha angle close to zero indicates surface scattering.As it increases from 0, it approaches to volume or multiple scattering at α = 45 o and then it approaches to double bounce scattering till α = 90 o .
Figure 2: H-A-Alpha Segmented Plane Surface scattering in the study area are represented by the blue portion of the graph with alpha values less than 40 degrees.Volume scattering from features are shown in the green part where alpha varies from 40 to 45 degrees.Areas with alpha values more than 45 degrees have multiple scattering and are shown in red.