Patch-based textons for feature extraction from PolSAR images

Abstract

In remote sensing of the earth, polarimetric synthetic aperture radar (PolSAR) is an important imaging system because it is free from the weather and the time of night and day. In contrast to optical image represented by intensity or color, the PolSAR image is represented by measuring the complex scattering matrix as polarimetric properties including the phase and the amplitude information from SAR image acquisition step. This properties is often an obstacle to apply texton approach to PolSAR image directly. Texton approach is known to be a well-established technique in texture research such as texture classification and segmentation. In this work, we proposed patch-based textons for feature extraction from PolSAR images by clustering local image neighborhoods including the PolSAR image properties directly. In clustering step, Wishart-distance measure was used to compute distance between two complex matrices including polarimetric properties instead of Euclidean distance in K-means clustering algorithm. Texton dictionary was used as a texture descriptor for texture classification. The performance of patch-based textons in PolSAR images was evaluated with texture classification in the way of machine learning and 5-fold cross validation. The experimental data is divided into five predefined categories such as city, field, forest, grassland, and street. The experimental results depending on the way of choosing sample patches, speckle reduction, feature dimension for textons, the number of textons, patch size, and the parameter K for K-nearest neighbor were evaluated by the balanced accuracy. The balanced accuracy was used to evaluate the classification result instead of overall accuracy to adjust the influence of imbalanced samples. The proposed textons was compared with grayscale textons. Finally, we showed suitability of patch-based textons for PolSAR images.