A multi-temporal binary-tree classification using polarimetric RADARSAT-2 imagery

Abstract

Abstract The polarimetric Synthetic Aperture Radar (PolSAR) signal contains more parameters than single or dual polarized SAR when using a scattering matrix to characterize targets. The increased information content of PolSAR provides more potential inputs for machine learning and classification applications; however, polarimetric parameters tend to be simply used as input variables and the optimum parameters to efficiently separate land classes and their physical meaning has received little attention. This research application proposes a Multi-temporal Binary-Tree Classification framework (MBTC) to identify and integrate optimum scattering parameters and machine learning methods in a meaningful way. First, the optimum scattering mechanism that most effectively distinguishes pairs of land classes is derived by a Lagrange multiplier. Next, for each pair of classes, machine learning classifiers are trained by the optimum scattering power ratio and elevated by Out-Of-Bag (OOB) cross validation. Three machine learning algorithms- Support Vector Machine (SVM), Random Forest (RF) and Neutral Network (NN)- are investigated. Finally, a multi-temporal binary-tree classifier is constructed, in which each pair of land classes are distinguished by the optimized machine learning algorithms. Two independent study sites in Canada are used for evaluating the MBTC framework using RADARSAT-2 observations. The London site with 6 classes is used to analyze the optimum scattering mechanisms and execute a simple classification. The Carman site with 10 classes allows for an indepdent and comprehensive assessment of the MBTC by comparing against an advanced Model-Based Decomposition (MBD). At the London site, the MBTC achieves the maximum power ratio with the optimum scattering mechanism between each pair of classes and high overall accuracy (OA) of 91% and kappa coefficient (Kappa) of 0.9. At the Carman site, comparisons indicate that MBTC significantly outperforms the MBD with NN and SVM classifiers but has a similar accuracy to the MBD for RF classifier with OA of 85% and Kappa of 0.82. In cases with pairs of classes that are difficult to separate, such as barley and wheat, MBTC is shown to be superior in this research application.