Supervised land-cover classification of TerraSAR-X imagery over urban areas using extremely randomized clustering forests

This study investigates the impact of the use of scattering intensity and texture features derived from TerraSAR-X intensity images on urban land cover classification accuracy, in combination with the Extremely Randomized Clustering Forests as the visual codebook former and classifier. We propose a multi-orientation ratio descriptor to represent the features of each SAR image patch effectively, and introduce a graph cut optimization based Markov Random Field smoothing processing to reduce block boundary effects due to patch-based classification method. We compare our classification results using one or all features together on 1m resolution TerraSAR-X images and show that the reasonableness of the proposed descriptor and the effectiveness of the Extremely Randomized Clustering Forests classifier.