کارایی شاخص‌های راداری در استخراج سطوح نفوذناپذیر شهری با استفاده از تصویر رادار تمام پلاریمتریک

Introduction
Impervious surfaces are the surfaces on which water cannot infiltrate. Detection of urban impervious surfaces is of great importance because the extension of these surfaces is an indicator of built-up area expansion and population growth. In recent years, remote sensing images have been widely used for land cover /land use studies. The efficiency of optical images has been widely explored in impervious surface delineation in urban areas. However, detection of impervious surface is not a simple task. Impervious surfaces vary in size, shape and material. Similar spectral responses among impervious surfaces and other types of land cover make the separation of impervious surfaces and other classes challenging.
Synthetic Aperture Radar (SAR) images are getting more and more attention in urban areas mapping. However, most of the studies concentrated on fusion of optical and SAR images or single polarized data. Full polarimetric SAR images offer more capabilities in separation of different land cover classes because, in full polarimetric mode, all characteristics of object’s backscattering will be perceived. SAR indices are computed based on data of two or more polarimetric bands. Therefore, they contain more information of land cover classes. Although the calculation and interpretation of SAR indices is simple, they are not fully understood in impervious surface detection in an urban environment. For impervious surface extraction, different classifiers have been used such as maximum likelihood, support vector machine and neural network. Among them, non-parametric classifiers often reach higher classification accuracies. Therefore, in this study support vector machine (SVM) algorithm has been applied.
Since the efficiency of full polarimetric SAR has not been evaluated for urban impervious surfaces, this study focused on the extraction of these surfaces in the complex urban area by the L-band full polarimetric SAR image. Most of the previous studies focused on optical images as well as the fusion of optical and SAR images. In cloudy and rainy weather, optical images are not available. In such a situation, the use of optical image and fusion of optical and SAR images are not possible. Therefore, we have studied the independent use of SAR images and extracted SAR indices.
Methodology
Tehran has been chosen as the study area since it has a complex structure. Tehran is the capital of Iran and is accounted as the economic and commercial center of Iran. Different impervious surfaces are found in this city. Impervious surfaces include residential areas, commercial areas, highways and parking lots. These surfaces are very diverse in terms of size and materials. As an example, narrow streets, irregular streets as well as wide and regular wide highways all exist in this city. Impervious surfaces may be perceived as dark and bright impervious surfaces in SAR images. Residential area and the area covered by new cement have high backscattering values and appear bright in SAR image. Streets and old cement have lower backscattering values and appear darker in this image. In addition to the impervious surface, vegetation, water body and bare land are also found in Tehran.
One Advanced Land Observing Satellite / Phased Array type L-band Synthetic Aperture Radar (ALOS/PALSAR) scene acquired on 23 April 2009 which encompassed Tehran has been selected. This scene has four polarimetric bands; HH, HV, VH and VV. The image has been processed and DN values have been converted to sigma nought in decibel. In order to reduce the topographic effect, radiometric terrain correction has been applied. Enhance Lee filter has been applied to minimize the speckle effect.
Ratio index, average index, difference index, normalized difference index and NLI index have been computed by the different combination of two polarimetric bands. It is proved that SAR indices are effective in separating different land cover classes. For classification purpose, support vector machine algorithm has been applied. SVM is a non-parametric classifier that has been applied extensively in SAR applications. It has no previous assumption of the statistical distribution of data. Training samples have been chosen on high spatial and historical google earth image. Approximately 500 pixels have been selected for each land cover class. The classification was followed by five scenarios; in the first scenario, only four polarimetric bands were used. In the second scenario, bands HH-HV and driven indices were classified. In the third scenario, bands HV-VV and their SAR indices were considered. Co-polarized bands (HH and VV) and SAR indices calculated based on these two bands were interred in the fourth scenario. Four polarimetric bands and all extracted indices were inserted in the fifth scenario. For accuracy assessment, circa 300 pixels were selected independently for each class as validation samples on high spatial and historical google earth image. By comparing classification results with validation samples confusion matrices were constructed. Based on the confusion matrix, overall accuracy, Kappa coefficient, producer and user accuracies were computed.
Results and discussion
The overall classification accuracy of the first scenario was 92.67% and the kappa coefficient was 0.9. This shows full polarimetric SAR images are capable of delineating impervious surfaces in the complex urban area. Band combination of HH-HV and driven indices yield overall accuracy and kappa coefficient, 84.23% and 0.78, respectively. The third scenario reached to 90.30% for classification accuracy and 0.86 for the kappa coefficient. The highest classification accuracy from two polarimetric bands is achieved by this scenario. These results could be justified by the presence of vertical polarization in both bands. Diverse vertical structures in the urban texture could be better distinguished by vertical polarization. The combination of co-polarized bands and their indices has 79.16% classification accuracy and 0.7 kappa coefficient. The lowest accuracy belongs to this scenario. The absence of cross-polarized bands may cause such relatively poor results. Depolarization is dominant in a well-developed urban area and co-polarized bands are not capable to capture depolarization. The last scenario reached the highest classification accuracy; 95.59% for overall accuracy and 0.96 for the kappa coefficient. The comparison between the first and last scenario shows the importance of SAR indices.
Conclusion
Three main conclusions can be driven from the findings of this study. First, full polarimetric bands are capable of urban impervious surface extraction. This is of great importance, especially in the absence of optical images. Second, dual polarimetric SAR images and their driven indices can extract impervious surface efficiently. Since most of SAR sensors work in dual mode, dual polarimetric SAR images have high availability. This study shows by the help of SAR indices, dual polarimetric can be used alternatively. And the last conclusion implies the importance of vertical polarization. In case, vertical polarization exists in both polarimetric bands, high classification accuracy will be achieved.