Integration of multi-seasonal Landsat 8 and TerraSAR-X data for urban mapping: An assessment

Abstract

Accurate land cover maps provide critical information to scientists and decision-makers involved in urban monitoring and management. Satellite remote sensing can be used for producing mid-resolution urban maps at regional scale, especially when integrating multispectral optical information with SAR data. Starting from processing of Landsat 8 and TerraSAR-X multi-seasonal data (March-August 2014) covering a study area located in Lombardy region (Italy), we carried out an assessment of urban mapping performance using different non-parametric supervised classification algorithms and input features. The results show that best overall accuracy is generally reached with Random Forest (95.5%) and Support Vector Machines (93.6%), using both optical and SAR information. Adding X-band backscatter as input information produced an average accuracy improvement around 3%. Among various land cover classes, detection errors were concentrated on urban sparse fabric, and vegetated land cover, especially when SAR features are not used as input.