Improving the impervious surface estimation with combined use of optical and SAR remote sensing images

Accurate mapping of urban impervious surfaces is important but challenging due to the diversity of urban land covers. This study presents an effort to synergistically combine optical and SAR data to improve the mapping of impervious surfaces. Three pairs of optical and SAR images, Landsat ETM + and ENVISAT ASAR, SPOT-5 and ENVISAR ASAR, and SPOT-5 and TerraSAR-X, were selected in three study areas to validate the effectiveness of the methods in this study. The potential of Random Forest (RF) was evaluated with parameter optimization for combining the optical and SAR images. Experiment results demonstrate some interesting findings. Firstly, the built-in out-of-bag (OOB) error is insufficient for accuracy assessment, and an assessment with additional reference data is required for combining optical and SAR images using RF. Secondly, the optimal number of variables (m) for splitting the decision tree nodes in RF should be some different from the principles reported previously, and an empirical relationship was given for determining the parameter m. Thirdly, the optimal number of decision trees (T) in RF is not sensitive to the resolutions and sensor types of optical and SAR images, and the optimal T in this study is 20. Fourthly, the combined use of optical and SAR images by using RF is effective to improve the land cover classification and impervious surface estimation, by reducing the confusions between bright impervious surface and bare soil and dark impervious surface and bare soil, as well as shaded area and water surface. Even though the easily-confused land classes tend to be different in different resolutions of images, the effectiveness of combining optical and SAR images is consistent. This improvement is more significant when combing lower resolution optical and SAR images. The conclusions of this study could serve as an important reference for further applications of optical and SAR images, and as a potential reference for the applications of RF to the fusion of other multi-source remote sensing data. (C) 2013 Elsevier Inc. All tights reserved.