Unsupervised Objects Classification in ALOS-2 PALSAR-2 Images

The paper suggests a new method for the unsupervised classification of multi-channel radar images on the basis of full polarimetric information. The main stages of the method are data preprocessing (with radiometric and geometric correction), synthesis of the hue-saturation-value image, threshold pixel-based segmentation according to the criterion of maximum a posteriori probability in each channel, local spatial post-processing (classification over k-nearest neighbors in the sliding filter window), and logical convolution of the classification results over each signature. The thresholds essential to make a decision about the class are found via the procedure of automatic splitting of Gaussian distributions mixture. The results of clustering of radar data received by the full polarimetric synthetic aperture radar PALSAR-2 of Japanese satellite ALOS-2 launched in 2014 are presented. Classification accuracy is estimated with the image fragments which represent the generalized surface classes of 'water' and 'not water'. The quality criteria were the general probability of error and the F-measure. It is shown that the developed method is applicable for analysis and interpretation of polarimetric data at solving the tasks of space monitoring and thematic mapping with satellite images.