International Journal of Advancement in Remote Sensing, GIS and Geography
COMPARATIVE EVALUATION OF THE
SENSITIVITY OF MULTI-POLARISED SAR AND
OPTICAL DATA FOR VARIOUS LAND COVER
CLASSES
Sugandh Chauhan* and Hari Shanker Srivastava Indian Institute of Remote Sensing (ISRO), Dehradun-248 001, Uttarakhand, India (*sugandhchauhan27@gmail.com, hari.isro@gmail.com, harishanker_srivastava@iirs.gov.in)
---ABSTRACT: The availability of reliable and timely land cover information is the basis to have sound economic planning and resource management for a modern nation like India. In this study, the capabilities of the dual polarimetric Envisat-1 ASAR and Landsat ETM+ data have been investigated for the land cover mapping. A comprehensive evaluation of the sensitivity of the cross-polarized (VH)/like-polarized (VV) ENVISAT-1 ASAR and optical data for various land cover classes has been done and a class separability analysis has been performed under different band combinations. In order to ensure maximum information retrieval, the bands MPDI (Microwave Polarization Difference Index) and NDVI (Normalized Difference Vegetation Index) have also been incorporated in the study. The separability among the class pairs have been analyzed using the Transformed Divergence (TD) procedure while the classification has been carried out using the Maximum Likelihood supervised classifier. The results of sensitivity analysis indicated that the vegetation is highly sensitive to the VH band owing to volume scattering while the built-up class could be more accurately distinguished in the VV band due to the corner reflector effect. The separability analysis further revealed that with the fusion of optical-VH polarised SAR data and the introduction of MPDI band to the multi-polarised SAR data, the separability among various class pairs greatly improved. The Landsat ETM+ and VH backscatter data fused image thus provided the highest classification accuracy of 91.25% with the kappa coefficient of 0.90, thus demonstrating its potential in land cover assessment and monitoring.
KEYWORDS: Land cover mapping, ENVISAT-1 ASAR, Multi-polarised SAR data, Multi-sensor satellite data, Microwave Polarization Difference Index (MPDI), Separability analysis.
---1. INTRODUCTION: The increasingly rapid deterioration of the environmental quality is now at the prime focus,
prompting many policy makers and legislators to find ways to tackle, stop or even reverse this process (1). An accurate and timely inventory of the land cover classes is thus the need of ours to tackle many socio-ecological concerns like the loss of fertile agricultural lands due to the sprawling unplanned urban growth. In the last three decades, a drastic evolution has been witnessed pertaining to the methods and technologies of remote sensing, with the remotely sensed satellite imageries and aerial photographs being used effectively to map the land cover features (2). Coupled with the data from ground surveys, reduced data cost and increased resolution from the satellites, this technology appears poised in serving the land management initiatives involved in monitoring the land cover change at different spatial scales (3). A number of landscape attributes like the land cover (4), landscape pattern (5) and its condition (6) can be mapped from this technique. These attributes are a major input to the LULC models [(7), (8)].
Traditionally, multi spectral data from Landsat have been used widely for the land cover studies and classifying crop types [(9)-(11)]; however, it comes with a major drawback. Firstly, the data acquisition is hampered in the cloud cover/rainy conditions, a problem that plagues many areas within the humid tropics. Secondly, the reflectance values in the visible and IR region primarily depends upon the molecular resonance of the surface materials, thus the features with varying physical dimensions may give similar spectral response in the optical imagery (12). On the other hand, the microwave remote sensing systems, specifically the synthetic aperture radar (SAR) have proved to be an alternative data source in many areas like forestry, wetland, agriculture, human settlement, grassland etc. [(13)-(16)]. The shortcomings of the optical data have been overcome due to the insusceptibility of this longer wavelength to the atmospheric scattering and the ability to penetrate clouds, fog, smokes, rain, etc. Secondly, there is an added advantage of providing control over factors like polarization, frequency, incidence angle, etc. (17).
The feasibility of the SAR data has already been tested and established for the vegetation studies (18). Studies reveal that the interaction of the SAR signal with the vegetation is volumetric in nature and is quite sensitive to the canopy structure, its orientation and the moisture content [(19), (20)]. However, in order to discriminate the variations
within the major vegetation types or the wetland habitats, the use of single date or single frequency/polarization may not be sufficient ([(21)-(23)]. The use of multi-parametric, Interferometric (InSAR), Polarimetric (PolSAR) and/or Polarimetric-interferometric (PolInSAR) SAR data often result in better discrimination and enhanced classification. Currently it is the uprising and nascent area of Synthetic aperture radar remote sensing [(24)-(27)]. In addition, studies done by [(28)-(30)] indicate that the integrated use of data from the discrete portions of electromagnetic spectrum can be much more advantageous. Such an integration can produce a clearer image with improved classification accuracies & may reduce the redundancy of optical bands. Rosenthal & Blanchard (31) investigated the synergistic effect of the integration of the data from radar and visible infrared sensors on the crop classification and reported an increase of over 20% in the classification accuracy. Blaes, et. al., (32) reported similar improvements in overall accuracies (about 5%), when the optical and SAR data sets were integrated.
A number of process based models, operating at large spatial or temporal scales, which aim at studying the nature of fluxes between various components of the earth system, requires an input of many parameterized land cover attributes [(33), (34)]. It demands an effective classification of the natural diversity into a set of finite manageable classes, such that the resultant classes have a direct functional value to the process being modeled (35). Historically, many attempts have been made to classify the complex mosaic of the terrestrial features by exploiting the microwave region of the electromagnetic spectrum [(36), (24)]. A number of classification techniques have evolved over the years; ranging from the manual classification of airborne SAR imagery in 1970s to the supervised or unsupervised digital techniques (37). The most common of these is the supervised statistical Maximum Likelihood Classifier (MLC) approach, which clusters the data from the training sets provided. In spite of the assumption of the normal distribution of the data, this parametric classifier has shown to provide acceptable accuracies, although with some applied modifications (38). Previous classification efforts also show that the use of this technique may provide very high classification accuracy, often exceeding 90%, if applied on the data on which they were trained [(35), (39), (40)].
In this study, an attempt has been made to perform a comparative evaluation of the sensitivity of the multi-polarized (VV/VH) SAR and optical remote sensing data to different land cover classes. In order to successfully extract the required information and enhance several features, the analysis has also been done on the optical and VH polarised fused image. Moreover, two bands viz., the Normalized Difference Vegetation Index (NDVI) and Microwave Polarization Difference Index (MPDI), have been incorporated for comprehensive information retrieval. Although many other vegetation indices like SAVI, NDBI etc. can also be used in the analysis but only NDVI has been used due to its wider use and well accepted status amongst the remote sensing community. The corresponding land cover maps for different band combinations have been analyzed and the resulting accuracies have been examined.
2. STUDY AREA AND DATASET USED: The study area was chosen over the parts of Haridwar district located in
Uttarakhand. Dominantly, the study area is a flat level terrain having the agricultural land, both in irrigated and un-irrigated form. A diverse variety of soils ranging from fine loamy, coarse loamy to fine silt and sandy soils are present in the area. Apart from different land cover classes, the Gang canal is the characteristic feature of this area. Fine alluvial soil deposited by the mighty Ganga and its tributaries along with the network of canals and tube wells, favor the growth of the wheat crop. The location map of the study area along with the Google Earth sub-image has been shown in Fig I.
Fig I: Location of the study area with the Google earth sub-image
For the present study, co-polarized (VV) & cross-polarized (VH) ENVISAT-1 ASAR IS-4 data acquired on 4 February 2005 has been used. The attributes of the Envisat data used in the study have been presented in Table I. The
Landsat ETM+ data dated 8 February 2005, downloaded from the USGS Earth Explorer
The Landsat ETM+ data with spatial resolution 30m, consists of eight multispectral bands, of which, the bands 2(Green), 3(Red) & 4(NIR) have been used in the study. A reference Landsat ETM+ image (non-stripped) was used for gap filling in the 2005 downloaded dataset. Fig IIa & IIb shows the respective microwave and optical sub-image of the study area.
Parameters Characteristics
Polarization Dual (VV, VH)
Frequency 5.36 GHz
Wavelength 5.6 cm
Spatial Resolution 30 meters
Incidence Angle 360 (central)
Pass Descending
Table I: Characteristics of Envisat-1 ASAR Data
a)
b)
Fig II: Sub image of the study area (a) ENVISAT-1 ASAR (VV-VH) data, (b) Landsat ETM+ standard FCC data
3. METHODOLOGY:
3.1. GROUND TRUTH DATA COLLECTION: An investigation involving the land cover mapping using SAR
data demands a very sound experimental plan, since this is the backbone of any project which utilizes SAR data. The ground truthing was conducted in synchrony to the Envisat-1 ASAR satellite pass. The candidate areas to be surveyed were identified from a reference Landsat ETM+ image and the SOI topo sheet (1:50,000). The minimum field size of
150*150 m2 was taken as the sampling unit using the statistical approach proposed by Patel & Srivastava, 2013 (41).
Nearly homogeneous fields were considered to obtain the ground truth data. GPS based mobile mapping unit was used to record the field geographic locations & trace the field boundaries. The vector layers provided by the GPS unit (in UTM projection) contained the detailed information for each sampling field as an output layer. These layers were then superimposed on the geo-referenced images in order to accurately retrieve the training set for each class. The method proposed by Patel & Srivastava (41) enabled the accurate identification of the field locations irrespective of the need of any location to fall near GCPs like rail/road/canal crossings.
3.2. DATA PRE-PROCESSING: Two scenes of the Landsat ETM+ data corresponding to the SAR image for the
year 2005 were downloaded from the USGS Earth Explorer website. The scenes were mosaicked and the desired area was extracted. However, the optical image contained strips on both sides of each scene which resulted in almost 22%
data loss and thus needed to be rectified. A reference Landsat ETM+ image (without strips), accurately co-registered with the stripped image, was utilized to fill these gaps using landsat_gapfill.sav plug-in of ENVI software. The DN values from the optical image were then converted to the reflectance values while the Envisat-1 ASAR image was
processed to generate backscattering coefficient (σo) image. The SAR image was then co-registered to the
geo-referenced optical image using 113 ground control points (GCPs), in order to match with the base image geometry. The points were selected on clearly delineated junctions and corners of streets or buildings or other linear features. A second order polynomial transformation with the nearest neighborhood sampling approach was applied. The related Root Mean Square Error (RMSE) was approximately 1.02 pixels.
3.3. IMAGE PRE-PROCESSING:
3.3.1. IMAGE FUSION: After the data pre-processing and co-registration of images, the backscatter image of VH
was fused with Landsat ETM+ data (Fig. III). Image fusion is a technique to combine the relevant information from a set of images and is usually done to enhance the interpretability and analysis of the features. More comprehensive information can be retrieved from the resultant fused image which cannot be derived from any one of them individually [(42), (43)]. In the framework of this study, the potential of IHS data integration approach has been assessed to generate the composite image at pixel level for better discrimination of the features.
IHS is a spatial domain fusion/image sharpening technique in which the pixel values are manipulated directly to achieve desired results. The intensity (I) represents the brightness values of the scene and ranges from 0 (black) to 1 (white). Hue tells about the color or the specific wavelength of the pixel and varies from 0 to 360. The saturation refers to the purity of the hue and varies linearly from 0 to 1 (44). It is the standard and the most widely used image fusion procedure for feature enhancement since it separates the color information in ways that is perceptible to the human visual response system [(45), (46)]. In the IHS space, hue and saturation contains mostly the spectral information while the intensity change has very little effect on the spectral variations.
Fig III: Sub-image of the study area - Landsat ETM+ multispectral data fused with the VH backscatter image
3.3.2. DATA PREPARATION: A total of five image sets were prepared for the separability analysis with different
band combinations. The first image comprised of the VV and VH bands (Fig. IIa). For the second image, the band 2 (Green), band 3 (Red) and band 4 (NIR) of the Landsat ETM+ data were stacked together (Fig. IIb). The VH backscatter image fused with the standard FCC formed the third set (Fig. III). Additionally, two bands viz., Normalized Difference Vegetation Index (NDVI) & Microwave Polarization Difference Index (MPDI), were derived from optical and microwave image, respectively. The NDVI image was generated from the Landsat image using the following relation:
NDVI is an indicator of the greenness and health of the vegetation. Its value ranges from -1 to +1. Higher NDVI values (0.5 or more) correspond to the dense vegetation while the lower ones (0.1 or less) denote the areas of barren rock, snow or the sand. The MPDI band was generated from the microwave image using the following relation:
The MPDI is rather sensitive to the variation in plant water content and the soil moisture. Thus, it can be used to characterize the vegetation. The NDVI & MPDI layers were stacked together with the VV and VH bands individually & thus formed the fourth (VV-VH-NDVI) and the fifth set (VV-VH-MPDI), respectively (Fig. IVa, IVb).
a)
b)
Fig IV: Sub-image of the study area in (a) VV-VH-NDVI bands (b) VV-VH-MPDI bands
3.3.3. CLASS SEPARABILITY ANALYSIS AND LAND COVER CLASSIFICATION: The visual analysis of
the Landsat ETM+ color composite along with the ground truth information enabled the identification of a total of nine major land cover categories (Table II).
LULC category Code Description
Wheat WH Wheat crop.
Sugarcane SC Sugarcane crop and Sugarcane ratoons.
Current fallow CF Ploughed fallow lands.
Waterbody WB River Ganga, Ganga canal, Ponds and other small rivers.
Aquatic vegetation AV The hydrophytes submerged in water/at water’s surface.
Forest/Plantation-Dense DFP Orchards & Sal/Sesame tress with dense canopy.
Forest/Plantation-Open OFP Orchards, Sal/Sesame trees, shrubs & grasses with a scarce
canopy cover.
Riverbed RB The smooth riverbed alongside the river Ganga
Built-up BU Commercial and residential areas, other man made features
like-parks, roads, railway lines, bridges, etc.
Nearly 15 signatures were taken, but were subsequently merged into one for each class, such that the distribution was across the entire image. The analysis of the spectral response curves of these selected samples ensured that the response patterns were distinct for different classes. The signatures were then evaluated by computing the signature separability and through an exploratory analysis of the confusion/contingency matrices. The separability was evaluated using the Transformed Divergence (TD) procedure under five band combinations (VV-VH, IR-R-G (Std. FCC), VV-VH-NDVI, Optical data fused with the VH polarised SAR data and VV-VH-MPDI) for the six class pairs viz., i) BU-CF, ii) BU-RB, iii) WB-AV, iv) WB-RB, v) DFP-OFP, and vi) WH-SC. This is among the most widely used distance measure which calculates the statistical distance between two multivariate, Gaussian distributed signatures. The value ranges from 0 to 2.0, with the value >1.9 indicating the optimum statistical separability between the signatures. After the separability test, the land cover classification was performed for the four images (excluding VV-VH image) using the MLC approach. Performing the classification alone is not sufficient, since the users need to know how accurate the maps are so as to use data more efficiently (47). Therefore, a common method i.e. a confusion matrix was utilized to evaluate the accuracy for each image combination. The calculation methods and interpretations for the overall accuracy, kappa coefficient, user’s and producer’s accuracy (UA & PA) were provided by the previous literatures [(47), (48)]. The minimum level of interpretation accuracy to identify a land cover category must be at least 85% (49). The methodological flowchart has been illustrated in the Fig. V.
Fig V: Methodology for the present study
4. RESULTS AND DISCUSSION:
4.1. SENSITIVITY OF THE LIKE-POL (VV) AND CROSS-POL (VH) DATA TO VARIOUS LAND COVER TARGETS: Large variations in the backscatter intensity were observed among the considered classes. Fig. VI
illustrates the mean backscatter response of each land cover class in the VV and VH bands (in terms of σo expressed in
dB). As can be seen from the scatter plot, the WB and the RB classes are almost indistinguishable on the VH axis (dynamic range of only 1.9 dB) due to the smooth appearance of both of these clasess. However, the classes are quite separable on the VV axis with a dynamic range of 7.73 dB. This is mainly due to the smooth water surface which specularly reflects the radar pulses away in opposite direction, thus resulting is a very low backscatter. Relatively, a higher backscatter is observed from the RB due to the large volume of the residual moisture contained in the sand and it was this moisture content that was sensitized by the VV polarization. The separability of the AV class from the WB
on the VH axis (mean σo
backscatter from the smooth underlying water surface and the little amount of the double bounce interaction between the water surface and the canopy elements. This vegetation volume is however, incomparable to the cropped areas or the forested classes, as is evident by its backscatter range.
During the acquisition of the SAR images, wheat and sugarcane were found to be the prime crops that were been grown in the area. Due to the presence of high vegetation volume in case of SC class, it gave higher backscatter in cross-polarised (VH) band as compared to that of WH class, which had relatively a lower amount of vegetation volume. Thus, these classes were clearly distinguished on the VH axis. Moreover, since the SAR backscatter is dependent on the biophysical parameters of the crop like the biomass, plant density, crop height, moisture content etc. and also its orientation (14); sugarcane crop returned higher backscatter in both bands (due to the high vertical structure of the crop and more multiple reflections from the canopy elements). In the context of the forest/plantation
class, the study area encompassed the dry deciduous forests, comprising an extensive vegetation of Sal (
Shorea
robusta
) and Sesame trees. The maximum dynamic range of σo for the forested areas was exhibited in the VH polarization (1.76 dB). The studies done by [(20), (50)] show that the interaction of SAR signal with the plant canopy is volumetric in nature and is mainly governed by the biomass level. Several backscatter mechanisms occurring in the forested terrain contributed to this higher backscatter in VH band which has been found to be sensitive to the plant density variations (20). The direct backscattering from the ground underneath, ground-trunk double bounce scattering, scattering from the upper part of the canopy, including the multiple scattering within the crown of the canopy resulted in such high values. However, one thing that needs to be emphasized here is that due to the excessive leaf fall duringthe time of data acquisition (4 Feb 2005), the σo
values in the forest category were slightly towards the lower side.
The CF class was clearly discernible from the RB class with the mean σo
range of nearly 3.01 dB on the VH axis. This was mainly due to the increased soil roughness that was observed in the fallow lands resulting from the agricultural practices like ploughing before the sowing of seeds. Consequently, this resulted in high backscatter values. Even the BU class was clearly distinctive in terms of its backscattering behavior, comparative to the fallow and riverbed class, despite being a complex assemblage of various land cover types like the sparse vegetation, bare soil and water, among others. An acute peak in the backscatter values, particularly in the VV band (2.4 dB), observed in this class can be ascribed to the concrete and metallic nature of the materials present within this class (BHEL industrial area), and also the well-defined vertical structure of the buildings which served as the dihedral corner reflectors. The VV band being highly sensitive to the corner reflector effect, resulted in a higher backscatter. The SAR image clearly separated out this class from the other classes, proving its advantage in the urban class mapping.
Fig VI: Backscatter response under VV and VH polarization
4.2. CLASS SEPARABILITY UNDER DIFFERENT BAND COMBINATIONS: The class separability analysis
was performed for several class pairs under 5 different band combinations. Table III shows the class separability values (using TD measure) for different class pairs. The comparative evaluation of these values under different band combinations unveiled some noticeable differences. VV-VH, VV-VH-NDVI & VV-VH-MPDI bands clearly distinguished the BU & CF features with the TD values of 1.99, 1.90 & 2.00 respectively. The backscatter from the BU class was comparatively higher on the account of corner reflector effect (both dihedral and trihedral) while the optical (IR-R-G) channels (TD = 1.52) could not distinguish these classes due to the similar spectral variations. Even when the optical data was fused with the VH polarised SAR data, the CF and BU classes could not be distinctly separated. It is due to the fact that in comparison to VV polarised SAR data, VH polarization is relatively less sensitive towards human settlements and more towards the depolarization effect. A similar pattern emerged for the BU & RB class pair. However, a good separability was observed for the WB-AV & WB-RB class pairs in all the band
combinations (TD>1.90). The optical data fused with the VH polarised SAR data served best to discriminate the OFP and DFP classes (TD = 1.9) due to the integrated influence of VH polarised SAR and optical data. The WH and SC class pair showed a unique separation in the VV-VH-MPDI (TD = 1.92) band while the other bands failed to distinctly distinguish these classes (Table III). This was mainly due to the presence of the MPDI band which is known to be sensitive to the variations in the leaf water content and the soil moisture status.
Overall, the optical data fused with the VH polarised SAR data and VV-VH-MPDI band combinations gave superior results in terms of distinctly separating the considered class pairs.
Class Pairs
Separability
VV-VH IR-R-G VV-VH-NDVI OPTICAL FUSED
WITH VH VV-VH-MPDI BU-CF 1.99 1.52 1.9 1.8 2 BU-RB 1.9 1.69 1.97 1.85 2 WB-AV 1.93 1.9 2 1.99 1.99 WB-RB 1.99 1.96 1.99 1.98 2 DFP-OFP 1.89 1.8 1.78 1.91 1.85 WH-SC 1.79 1.76 1.80 1.73 1.92
Table III: Separability of the class pairs under various band combinations using TD distance measure
4.3. LAND COVER CLASSIFICATION FOR DIFFERENT BAND COMBINATIONS: For the classification
purpose, the datasets having minimum three bands were considered (i.e. 4 datasets in this study). The classification maps for these band combinations using the MLC approach have been presented in Fig. VIII. The classification has been implemented in the ERDAS IMAGINE 2014 software using the selected training samples. The accuracy statistics pertaining to these classification results have been shown in Table IV. The overall results demonstrated that, various band combinations of either microwave or optical data, were sensitive to the discrete land cover features in distinct ways. All the four band combinations utilized in this study, to distinguish the nine land cover targets, contributed to improve the classification accuracies in one way or the other.
From Table IVd, it can be inferred that the single-source multispectral data i.e. standard FCC, showed the lowest overall accuracy (81.85%) with the kappa-coefficient of 0.80, in the identification of the considered land cover classes. Although, 81.85% classifications were accurate overall, it did not ensure that each class was also classified successfully at that rate. Therefore, User’s and Producer’s accuracies (UA and PA, respectively) were computed for all the classes. The classification of the spectrally separable classes of WB, AV & CF was done quite efficiently with low commission (i.e. high UA) and omission errors (i.e. high PA). On the other hand, UA and PA was quite low for the BU, RB, WH, SC, OFP & DFP classes. Low UA meant that there was a probability that the pixels classified under these categories did not actually exist on the ground while low PA implied that some of the ground reference points may have been incorrectly classified (Table IVd). A large amount of pixel mixing was observed between the CF and BU classes. The results thus signified the inability of the spectral classifier like MLC to discriminate the statistically overlapping classes due to very similar spectral characteristics. It was thus concluded that the optical sensor data such as that from Landsat ETM+, does not lead to an accurate vegetation classification, since it mainly captures the canopy and its related shadow information. Since the reflectance values from the shadow portions is nearly zero, the overall reflectance values from the vegetation is reduced. Also, the complexity of the forest stands results in saturated values, thus making it troublesome to classify these features even when their biomass densities differ remarkably [(14), (51)].
Incorporation of the NDVI band along with the VV & VH channels further increased the overall accuracy by 3.72% while kappa value by 5.00% as compared to the Std. FCC band (Table IVc). As was expected, this combination improved the classification accuracies for the vegetation class. The pixel mixing among the OFP & DFP classes was slightly reduced resulting in 82.09 & 81.97% of UA while 84.62 & 84.75% of PA, respectively. This increase could have been due to the composite effect of the NDVI band (sensitive to the chlorophyll content of the canopy) and the VH band (sensitive to the depolarization owing to the multiple scattering within the canopy). It is required to emphasize that the NDVI values of DFP (0.56) & OFP (0.37) is lower than that of WH (0.72) & SC (0.75) classes (Fig. VII). This was primarily due to the excessive leaf fall & the dried up forest vegetation that was witnessed during the period of data acquisition (4 Feb 2005). Contrarily, the separability in the case of the WH & SC categories was not quite significant, due to insufficient contrast in their NDVI values at the time of data acquisition (Fig. VII). The WB, AV and RB classes were also classified more accurately (UA increased by 3.1, 2.04 and 8.6%, respectively) as compared to the optical data. It is known that the NDVI is able to provide very precise information about presence/absence of vegetation. Thus, by the addition of the NDVI as a band along with the VV and VH channels, the classification accuracy for these classes should increase and same has been verified by the results obtained during the analysis.
Fig VII: NDVI plot for different land cover categories
However, the Table IVb reveals that with the usage of multi-polarimetric SAR data (VV-VH-MPDI), the overall accuracy considerably improved by 5.6 and 1.87% while the kappa value increased by 7.5 and 2.4% as compared to the Std. FCC & VV-VH-NDVI bands, respectively. The identification of WH and SC classes particularly improved as indicated by the increased UA (92.31 & 88.41%, respectively) & PA (87.27 & 92.42%, respectively). Interesting to note in Table IVb is that this band combination provided the highest accuracy in classifying the WH & SC classes. Even the OFP and DFP categories were classified at a successful rate as compared to the optical sensor data (UA increased by 8.72 & 6.84% while PA by 4.36 & 5.96%, respectively). This can be attributed to the capability of the SAR data to effectively penetrate within the vegetation canopy. Also, the ability of the MPDI band to characterize the vegetation (owing to the sensitivity to the variations in leaf water content) played an important role. The classification results also indicated that the identification of the BU class certainly improved since the UA increased by 3.3, 4.8 & 15.2 % compared to the VV-VH-NDVI, fused & Std. FCC datasets, respectively. The prime reason being the well-defined vertical structure of the buildings which behave as dihedral corner reflectors, thus giving an acute peak in backscatter (mainly in VV band). Similar results obtained by (19), suggest that the SAR data alone is capable of detecting the human settlements.
c)
d)
Fig VIII: Land cover maps for (a) Optical data fused with the VH polarised SAR data (b) VV-VH-MPDI (c) VV-VH-NDVI (d) IR-R-G (Std. FCC) band combinations
LULC classes WH DFP WB BU RB AV CF OFP SC Row total UA (%)
WH 50 0 0 0 0 0 0 0 12 62 80.65 DFP 0 50 0 0 0 0 0 1 0 51 98.04 WB 0 0 52 0 4 1 0 0 0 57 91.23 BU 0 0 0 50 0 0 7 0 0 57 87.72 RB 0 0 2 0 52 0 0 0 0 54 96.30 AV 0 0 2 0 0 54 0 0 0 56 96.43 CF 0 0 0 2 1 0 59 0 0 62 95.16 OFP 0 2 0 0 0 0 0 52 0 54 96.30 SC 11 0 0 0 0 0 0 0 50 61 81.97 Column Total 61 52 56 52 57 55 66 53 62 514 PA (%) 81.97 96.15 92.86 96.15 91.23 98.18 89.39 98.11 80.65 Overall accuracy = 91.25% Kappa coefficient = 0.90
(a)
LULC classes WH DFP WB BU RB AV CF OFP SC Row total UA (%)
WH 48 0 1 0 0 0 0 1 2 52 92.31 DFP 1 33 0 0 0 0 0 4 0 38 86.84 WB 0 0 48 0 0 2 1 3 0 54 88.89 BU 0 2 0 50 0 0 2 0 0 54 92.59 RB 0 0 0 0 30 0 7 0 0 37 81.08 AV 2 0 1 0 0 38 0 0 1 42 90.48 CF 0 0 0 2 6 0 35 0 0 43 81.40 OFP 1 5 1 0 0 0 0 40 2 49 81.63
SC 3 1 0 0 0 0 3 1 61 69 88.41
Column Total 55 41 51 52 36 40 48 49 66 438
PA (%) 87.27 91.67 94.12 96.15 83.33 95.00 72.92 81.63 92.42 Overall accuracy = 87.44% Kappa coefficient = 0.86
(b)
LULC classes WH DFP WB BU RB AV CF OFP SC Row total UA (%)
WH 47 0 0 0 0 0 0 0 9 56 83.93 DFP 0 50 0 0 0 0 0 10 1 61 81.97 WB 0 0 43 0 3 0 0 0 0 46 93.48 BU 0 0 0 50 0 0 6 0 0 56 89.29 RB 0 0 7 0 42 0 0 0 0 49 85.71 AV 0 0 4 0 0 45 0 0 0 49 91.84 CF 0 0 1 1 7 0 51 0 0 60 85.00 OFP 0 9 0 0 0 3 0 55 0 67 82.09 SC 12 0 0 0 0 0 0 0 50 62 80.65 Column Total 59 59 55 51 52 48 57 65 60 506 PA (%) 79.66 84.75 78.18 98.04 80.77 93.75 89.47 84.62 83.33 Overall accuracy = 85.57% Kappa coefficient = 0.84
(c)
LULC classes WH DFP WB BU RB AV CF OFP SC Row total UA (%)
WH 45 0 0 0 0 0 0 0 12 57 78.95 DFP 0 60 0 0 0 0 0 15 0 75 80.00 WB 0 0 47 0 0 5 0 0 0 52 90.38 BU 0 0 0 41 8 0 4 0 0 53 77.36 RB 0 0 0 15 54 0 1 0 0 70 77.14 AV 0 0 5 0 0 44 0 0 0 49 89.80 CF 0 0 0 5 4 0 59 0 0 68 86.76 OFP 0 10 0 2 0 1 0 51 0 64 79.69 SC 13 0 0 0 0 0 0 0 50 63 79.37 Column Total 58 70 52 63 66 50 64 66 62 551 PA (%) 77.59 85.71 90.38 65.08 81.82 88.00 92.19 77.27 80.65 Overall accuracy = 81.85% Kappa coefficient = 0.80
(d)
Table IV: Confusion Matrices for (a) Optical data fused with the VH polarised SAR data (b) VV-VH-MPDI (c) VV-VH-NDVI (d) IR-R-G (Std. FCC) Band Combinations
Furthermore, the classification accuracy improved with the fusion of the Landsat ETM+ multispectral and VH backscatter images. The overall accuracy increased up to 91.25% with 12.5 percent increase in kappa coefficient compared to optical, 7.14 percent increase compared to NDVI and 4.6 percent increase compared to VV-VH-MPDI bands (Table IVa). The forest category (both OFP & DFP) as well as the RB, AV & CF classes were classified most accurately in comparison to other datasets (with UA & PA above 95%) (Table IVa). This was predominantly due to the enhanced identification of these features which resulted from the integrated influence of VH backscatter image & the optical data. The VH channel contributed due to the differences in surface roughness, shape, orientation and moisture content of different features while the optical data facilitated in the understanding of the spectral variations of the targets. In case of the RB and CF classes, this combination enabled a good separation between the vegetated and non-vegetated areas as well as between the smooth and undulating terrain. Even the AV class was classified accurately in comparison to other datasets due to a very high contrast in NDVI and roughness of smooth water surface and rough AV on the optical and SAR remote sensing data, respectively. However, the results from the analysis also indicated that such fusion could not cater to the improved classification accuracy of the BU class. This could be attributed to the fact that the VH band; which is more sensitive to the volumetric scattering, chosen for the fusion had relatively lower significant impact on improving its mapping accuracy. The outcome of this study is in line with the results obtained by other researchers [(26), (50)].
5. CONCLUSIONS: This paper compares the utility of the SAR and optical data for the land cover mapping. Three
basic objectives were set during the onset of this analytical study and were achieved through a well-defined methodology. ENIVISAT-1 ASAR and LANDSAT ETM+ datasets were used for the study. Firstly, a comparative evaluation was done regarding the sensitivity of the cross (VH) and like (VV) polarized ENVISAT-1 ASAR data to the set of nine land cover classes identified in the study area. The results of the study indicate that the BU class, being a complex assemblage of various land cover types, is highly sensitive to the VV band, while the vegetation classes (DFP, OFP, WH, SC), are more separable in the VH band, owing to volumetric scattering. The most remarkable thing observed is the separability of the CF and the BU class in the SAR data. These classes are often misclassified in optical data due to the fact that reflectance of dry fallow lands mixes up with that of built-up class, while on the other hand, the SAR data separated out these classes due to the higher sensitivity towards roughness and orientation of target under consideration. Secondly, a comprehensive separability analysis was performed for six class pairs under five band combinations, using the TD distance measure. Among these, the Landsat ETM+ and VH backscatter fused dataset and VV-VH-MPDI bands exhibited the maximum separability values. Lastly, the land cover maps were produced for different bands and the resulting accuracies were discussed in detail. The results indicated that the fusion of optical and cross polarized (VH) SAR data resulted in highest classification accuracy (91.25%) with the kappa value of 0.90, thus exhibiting its potential in the characterization of the land cover features.
6. ACKNOWLEDGEMENTS: The authors are extremely thankful to Dr. A. Senthil Kumar, Director, IIRS/ISRO,
Dehradun for the encouragement and support. Authors are also thankful to Dr. Sarnam Singh, Dean (Academics), IIRS, Dr. S.P.S. Kushwaha, Former Dean (Academics) & Group Director, ER & SS Group, IIRS, and Dr. Suresh Kumar, Head, ASD/IIRS for useful discussions and support.
REFERENCES:
1. Pardal, M. A., Cardoso, P. G., Sousa, J. P., Marques, J. C. and Raffaelli, D., 2004, Assessing environmental quality: a novel approach, Marine Ecology Progress Series, 267, 1-8.
2. Rawat, J. S., and Kumar, M., 2015, Monitoring land use/cover change using remote sensing and GIS techniques: A case study of Hawalbagh block, district Almora, Uttarakhand, India, The Egyptian Journal of Remote Sensing and Space Science, 18(1), 77-84.
3. Rogan, J., and Chen, D., 2004, Remote sensing technology for mapping and monitoring land-cover and land-use change, Progress in planning, 61(4), 301-325.
4. Srivastava, H. S., Patel, P. and Navalgund, R. R., 2006, Application potentials of synthetic aperture radar interferometry for land-cover mapping and crop-height estimation, Current Science, Vol. 91, No. 6, pp. 783 - 788; http://repository.ias.ac.in/60596/1/7-pub.pdf
5. Shao, G., and Wu, J., 2008, On the accuracy of landscape pattern analysis using remote sensing data, Landscape Ecology, 23(5), 505-511.
6. Aspinall, R., and Pearson, D., 2000, Integrated geographical assessment of environmental condition in water catchments: Linking landscape ecology, environmental modelling and GIS, Journal of Environmental Management, 59(4), 299-319.
7. Ollinger, S. V., Aber, J. D., and Federer, A., 1998, Estimating regional forest productivity and water yield using an ecosystem model linked to a GIS, Landscape Ecology, 13(5), 323-334.
8. Soares-Filho, B. S., Cerqueira, G. C., and Pennachin, C. L., 2002, DINAMICA—a stochastic cellular automata model
designed to simulate the landscape dynamics in an Amazonian colonization frontier. Ecological modelling, 154(3), 217-235.
9. Turker, M., and Arikan, M., 2005, Sequential masking classification of multi-temporal Landsat7 ETM+ images for field-based crop mapping in Karacabey, Turkey, International Journal of Remote Sensing, 26(17), 3813–3830.
10. Yuan, F., Sawaya, K.E., Loeffelholz, B.C. and Bauer, M. E., 2005, Land cover classification and change analysis of the
Twin Cities (Minnesota) Metropolitan Area by multi temporal Landsat remote sensing, Remote sensing of Environment, 98, 317-328.
11. Thenkabail, P. S., Schull, M., Turral, H., 2005, Ganges and Indus river basin land use/ land cover (LULC) and irrigated
area mapping using continuous streams of MODIS data, Remote Sensing of Environment, 95(3), 317–341.
12. Henderson, F. M., and Lewis, A. J., 1998, Principles and applications of imaging radar, Manual of remote sensing, New
York: John Wiley and sons, 3rd ed., vol. 2.
13. Patel, P., Srivastava, H. S. and Navalgund, R. R., 2006, Estimating wheat yield: An approach for estimating number of
grains using cross-polarized ENVISAT-1 ASAR data, Proc. of SPIE, Vol. 6410, 641009, 1-12; http://repository.ias.ac.in/89368/1/47P.pdf
14. Srivastava, H.S., Patel, P., Sharma, Y. and Navalgund, R. R., 2007, Detection and density mapping of forested areas
using SAR interferometry technique, International Journal of Geoinformatics, 3 (2), 1-10, http://repository.ias.ac.in/89337/
15. Srivastava, H. S., Patel, P., Prasad, S. N., Sharma, Y., Khan, B. A., Praveen, B., Prasad, K. C. A., Sharma, S., Vijayan, L.
and Vijayan, V. S., 2007, Potential applications of multi-parameteric synthetic aperture radar (SAR) data in wetland inventory: A case study of Keoladeo National Park (A World Heritage and Ramsar site), Bharatpur, India, in Proc. TAAL 2008: 12th World Lake Conference, Jaipur, INDIA, 1862–1879; http://moef.nic.in/sites/default/files/nlcp/Indian%20Case%20Studies/Q-13.pdf
16. Srivastava, H. S., Patel, P. Sharma, K. P., Krishnamurthy, Y. V. N. and Dadhwal, V. K., 2011, A semi-empirical modelling
approach to calculate two-way attenuation in radar backscatter from soil due to crop cover, Current Science, 100 (12), 1871-1874, http://www.currentscience.ac.in/Volumes/100/12/1871.pdf
17. Sinha, S., Jeganathan, C., Sharma, L.K. and Nathawat, M.S., 2015, A review of radar remote sensing for biomass
18. Lu, D., 2006, The potential and challenge of remote sensing-based biomass estimation, International Journal of Remote
Sensing, 27 (7), 1297–1328
19. Srivastava, H. S., Patel, P., Sharma, Y. and Navalgund, R. R., 2009, Multi-frequency and multi-polarized SAR response to
thin vegetation and scattered trees, Current Science, 97 (3), 425-429 ; http://repository.ias.ac.in/60600/1/11-pub.pdf
20. Patel, P., Srivastava, H. S., Panigrahy, S., and Parihar, J. S., 2006, Comparative evaluation of the sensitivity of
multi‐polarized multi‐frequency SAR backscatter to plant density, International Journal of Remote Sensing 27 (2), 293-305.
21. Patel, P., Srivastava, H. S., and Navalgund, R. R., 2009, Use of synthetic aperture radar polarimetry to characterize
wetland targets of Keoladeo National Park, Bharatpur, India, Current Science, 97 (4), 529-537. http://repository.ias.ac.in/60590/
22. Srivastava, H. S., 2007, Interaction of multi-frequency multi-polarized DLR ESAR data with various targets: A case study
with C, L and P Bands acquired at all the four linear (VV, HH, VH & HV) polarizations, In JEP-MW Conference, 15-16 May 2007, SAC, Ahmedabad, pp. 2:12-23, 2:12-23
23. Patel, P., Srivastava, H. S. and Navalgund, R. R., 2007, Use of Polarimetric SAR to Study Wetland Targets of Keoladeo
National Park (A World Heritage and Ramsar Site), Bharatpur, India, in Proc. TAAL 2008: 12th World Lake Conference, Jaipur, INDIA, 1983-1987, http://moef.nic.in/sites/default/files/nlcp/Indian%20Case%20Studies/Q-32.pdf
24. Alberga, V., 2007, A study of land cover classification using polarimetric SAR parameters, International Journal of Remote
Sensing, 28(17), 3851-3870.
25. Srivastava, H. S., Patel, P., Sharma, K.P. and Krishnamurthy, Y. V. N., 2009, Explored and demonstrated potential
applications of multi-parametric SAR in context of Keoladeo National Park, BHARATPUR, INDIA, in Proc. 2nd Annual Research Seminar on Keoladeo National Park (2nd ARS-KNP), Jointly organized by KNP, Forest Department of State of Rajasthan (India) & WWF, 15th March 2009, Keoladeo National Park, Bharatpur, India, pp. 01 - 30 ; http://ramsar.rgis.ch/pdf/features/features_remote_keoladeo.pdf
26. Patel, P., and H. S. Srivastava, 2013, RADARSAT-2 Announcement of Opportunity Project on Soil Moisture, Surface
Roughness and Vegetation Parameters Retrieval Using SAR Polarimetry, SAC/EPSA/MPSG/CVD/TDP R&D/01/13, SOAR International Closing and Reporting – 2013. Final Report Submitted to Canadian Space Agency (CSA) through MDA, Canada, Indian Space Research Organisation (ISRO), India, January 2013, pp. 01-81 ;
file:///C:/Users/harishanker/Desktop/SOAR%20ID%20Number%202195%20-%20PI%20Parul%20Patel-%20SI%20HariShankerSrivastava-ISRO-India-FINAL%20CONSOLIDATED%20DETAILED%20REPORT%20(5).pdf
27. Sivasankar, T., Srivastava, H. S., Sharma, P.K., Kumar, D. and Patel, P., 2015, Study of hybrid polarimetric parameters
generated from RISAT-1 SAR data for various land cover targets, International Journal of Advancement in Remote Sensing, GIS and Geography (IJARSGG), Vol. 3, Issue 1, 32-42 ; http://www.irosss.org/ojs/index.php/IJARSGG/article/view/496/139
28. Srivastava, H. S., Patel, P., Sharma, Y., and Navalgund, R. R., 2006, Comparative evaluation of potential of optical and
SAR data for the detection of human settlements using digital classification, International Journal of Geoinformatics, 2 (3), 21-28 ; http://repository.ias.ac.in/89336/
29. Srivastava, H.S., Patel, P. and Sharma, Y., 2008, Feasibility of use of optical and SAR data for land cover mapping, In
Proceedings of ISRS-2008 National Symposium, Indian Society of Remote Sensing, Ahmedabad.
30. Amarsaikhan, D., Ganzorig, M., Ache, P., and Blotevogel, H., 2007, The integrated use of optical and InSAR data for
urban land‐cover mapping, International Journal of Remote Sensing, 28(6), 1161-1171.
31. Rosenthal, W.D., and Blanchard, B.J., 1984, Active microwave responses: An aid in improved crop classification,
Photogrammetric Engineering and Remote Sensing, 50(4), 461–468.
32. Blaes, X., Vanhalle, L, and Defourny, P., 2005, Efficiency of crop identification based on optical and SAR image time
series, Remote Sensing of Environment, 96(3), 352–365.
33. Golubiewski, N. E., and Hussein, G. H., 2007, Remote sensing’s functional role in studies of land-use/land-cover change,
Cleveland encyclopedia of earth, Washington, DC: Environmental Information Coalition, National Council for Science and the Environment, Published in the Encyclopedia of Earth.
34. De Roo, A. P. J., Wesseling, C. G., and Van Deursen, W. P. A., 2000, Physically based river basin modelling within a GIS:
the LISFLOOD model, Hydrological Processes, 14, 1981-1992.
35. Dobson, M. C., Ulaby, F.T., and Pierce, L. E., 1995 Land-cover classification and estimation of terrain attributes using
synthetic aperture radar, Remote Sensing of Environment, 51(1),199-214.
36. Qi, Z., Yeh, A. G. O., Li, X., and Lin, Z., 2012, A novel algorithm for land use and land cover classification using
RADARSAT-2 polarimetric SAR data, Remote Sensing of Environment, 118, 21-39.
37. Pant, T., Singh, D., and Srivastava, T., 2010 Advanced fractal approach for unsupervised classification of SAR
images, Advances in Space Research, 45(11), 1338-1349.
38. Conese, C., and Maselli, F., 1992, Use of error matrices to improve area estimates with maximum likelihood classification
procedures, Remote Sensing of Environment, 40, 113-124.
39. De Grandi, F., Lemoine, C. C., de Groof, H., Lavalle, C., and Sieber, A. J., 1994, Fully polarimetric classification of the
Black Forest MAESTRO 1 AIRSAR data, Int. J. Remote Sens., 15, 2755-2775.
40. Lemoine, G. G., De Grandi, G.F., and Sieber, A. J., 1994, Polarimetric contrast classification of agricultural fields using
MAESTRO 1 AIRSAR data, Int. J. Remote Sens., 15, 2851-2869.
41. Patel, P. and Srivastava, H.S., 2013, Ground truth planning for synthetic aperture radar (SAR): addressing various
challenges using statistical approach, International Journal of Advancement in Remote Sensing, GIS and Geography, 1 (2),1-17 ; http://www.irosss.org/ojs/index.php/IJARSGG/article/view/162/64
42. Amarsaikhan, D., Blotevogel, H. H., Ganzorig, M., and Moon, T. H., 2009, Applications of remote sensing and geographic
information systems for urban land-cover changes studies, Geocarto International: A Multi-disciplinary Journal of Remote Sensing and GIS, 24, 257–271.
43. McNairn, H., Champagne, C., Shang, J., Holmstrom, D. and Reichert, G., 2009, Integration of optical and Synthetic
Aperture Radar (SAR) imagery for delivering operational annual crop inventories, ISPRS Journal of Photogrammetry and Remote Sensing, 64, 434-449.
44. E. F. Guide, 1999. Erdas Inc., Atlanta, Georgia, 5th Ed., 170.
45. Foody, G. M., 2002, Status of land cover classification accuracy assessment, Remote Sensing of Environment, 80,185–
46. Srivastava, H. S., Prasad., S. N., Manchanda, M. L. & Adiga, S., 2001, Radar remote sensing applications in wetland
habitats - a case study with multi-incidence angle RADARSAT SAR data, pp. 87-92. In: I.V. Muralikrishan (eds) Spatial Information Technology Remote Sensing and GIS. Vol 2. B.S. Publications, Hyderabad.
47. Smits, P. C., Dellepiane, S. G., and Schowengerdt, R. A., 1999, Quality assessment of image classification algorithms for
land-cover mapping: A review and a proposal for a cost-based approach, International Journal of Remote Sensing, 20, 1461–1486.
48. Anderson, R., Hardy, E. E., Roach, J. T., and Witmer, R. E., 1976, A land use and land cover classification system for use
with remote sensor data, U.S. Geological Survey Professional Paper, 964, 28.
49. Dobson, M. C., Ulaby, F. T., LeToan, T., Beaudoin, A., Kasischke, E. S., and Christensen, N., 1992, Dependence of radar
backscatter on coniferous forest biomass, Geoscience and Remote Sensing, IEEE Transactions on, 30, 412-415.
50. Lu, D., 2005, Aboveground biomass estimation using Landsat TM data in the Brazilian Amazon, International Journal of
Remote Sensing, 26, 2509–2525.
51. Zhang, H., Zhang, Y., & Lin, H., 2012, Urban land cover mapping using random forest combined with optical and SAR
data, In Geoscience and Remote Sensing Symposium (IGARSS), IEEE International, 6809-6812.
View publication stats View publication stats
