Semantic Segmentation of Urban Areas in Polarimetric SAR Imaging using Deep Neural Networks and Decision Trees

Data obtained from polarimetric SAR satellite imagery can be fed to machine learning models to automatically detect buildings & monitor the movements of the population in remote areas of interest. In this research, we implement and train (i) a state-of-the-art deep learning architecture, U-Net & (ii) a deep feature extraction network using VGG-16 and feature classification using the C5.0 decision tree, for the pixel-wise classification of orthogonal and non-orthogonal buildings. Based on the manually annotated ground truth labels from PolSAR images, we train and test our model on a custom built and annotated urban dataset. Model evaluation was performed using the Intersection over Union (IoU) metric for this semantic segmentation task, which outperforms other models in this research domain.