Comparison of Dimension Reduction Methods Using Polarimetric SAR Images for Tensor-based Feature Extraction

Inhalt:
Polarimetric Synthetic Aperture Radar (PolSAR) data can be used to classify materials based on their scattering properties. Using the radar returns that are back-projected onto a digital elevation model (DEM), we can combine several polarimetric decomposition techniques to create a feature vector for each return. PolSAR data of the same area collected at three different angles were used to compare two different dimension reduction methods for tensorbased feature extraction using an unsupervised clustering algorithm. In this experiment, we wanted to examine the effects of pre-processing PolSAR data using dimension reduction methods for machine learning algorithms. The two dimension reduction methods we examined are Principle Component Analysis (PCA) and Non-negative Matrix Factorization (NMF) and we tested the two algorithms for a range of components. The pre-processed data from these two methods were clustered using the k-nearest neighbor (KNN) algorithm for a various number of clusters to create classification maps that were analyzed for separability of materials and the differences in the clustering results due to collection geometry were also examined. This method for selecting the pre-processing parameters for analyzing data can be used to optimize information extraction for tasks that use machine learning algorithms.