Urban classification using preserved information of high dimensional textural features of Sentinel-1 images in Tabriz, Iran

Abstract

Monitoring urban land through satellite images has rapidly developed with the advent of modern technologies, and the increasing number of satellites plays a contributing role. While optical images have a high capability in urban monitoring, they still have some limitations, including their dependence on climatic conditions and spectral information, which lead to difficulty in making a distinction between bare land, buildings and other features. The impossibility of optical imagery at night is another issue that can make the land cover classification difficult. Synthetic aperture radar (SAR) allows imaging in all climatic conditions and at nighttime, with an ability to detect phenomena based on their geometry, roughness, and location, making the land cover classification much easier. In the present study, radar Sentinel-1 images with polarization VV and VH were used for the land classification in Tabriz. Sentinel-2 images for the same time were applied as a reference for the calibration and accuracy assessment. Maximum likelihood (ML) and support vector machine (SVM) algorithms were also employed for supervised classification. In both algorithms, the classification was performed in windows with different sizes once by the SAR backscattering coefficient (σ0) and then by combining the backscattering coefficients with the statistical data obtained from the texture. The results showed that the use of radar images only with backscattering intensity resulted in poor performance while using the gray-level co-occurrence matrix (GLCM) of texture features increased the accuracy. The transmitted frequencies of radar images have different redistributions to different phenomena. The numerical results obtained from the radar image classification show that using only the radar image redistribution led to low accuracies at both VV and VH polarization, but the use of the textural analysis significantly increased the accuracy of the classifications. The statistical results obtained from the ML and SVM classifications for radar images at VV and VH polarization indicated that the latter performed better than the former. When texture analysis was not used in the classes, the classification accuracy was low with kappa values of 0.37 and 0.42 for VV and VH polarization, respectively. The use of texture analysis and obtaining the optimum window size is increase the classification accuracy with a better performance for VH polarization. The SVM classification method with a kappa coefficient of 0.72% showed better performance than the ML one with a kappa coefficient of 0.61%. Conclusively, in the absence of Sentinel-2 datasets, Sentinel-1 images are good alternatives if the preserved texture information is available for the land cover classification. Results of this research are of great importance for developing the remote sensing methods and their techniques can be considered as progressive research in the domain of remote sensing sciences.