Multitemporal polarimetric RADARSAT-2 SAR data for urban land cover mapping through a dictionary-based and a rule-based model selection in a contextual SEM algorithm

This paper presents a dictionary- and rule-based model selection approach in an adaptive contextual semi-supervised algorithm for improving urban land cover classification using high-resolution multitemporal RADARSAT-2 polarimetric SAR (PolSAR) data. Six-date PolSAR data were acquired from June to September, 2008, over the Greater Toronto Area. Contextual information and the capabilities of different PolSAR distribution models were explored by the spatially variant Finite Mixture Model (FMM) with an adaptive Markov Random Field (MRF) in a Stochastic Expectation–Maximization (SEM) algorithm. This algorithm can obtain homogenous results while preserving shape details in the complex urban environment with high accuracy. Commonly used PolSAR distribution models such as Wishart, G0p, Kp, and KummerU were compared through the proposed approaches for urban land cover mapping. According to a Goodness-of-Fit test based on Mellin transformation, an accurate PolSAR distribution model could be selected with the dictionary-based classification. However, the results showed that improvement from the dictionary-based approach was limited. Therefore, further improvements were expected by exploring expert knowledge. The initial results showed that G0p and KummerU performed better for distinguishing between low density built-up areas and forest. G0p, Kp, and KummerU are better for the low scattering classes. The Wishart model has superior capacity in separating high density built-up areas and the adjacent roads. Based on such knowledge, a set of rules was developed to integrate the advantages of alternative models. Significant improvement on the overall classification accuracy could be observed by this rule-based approach. The biggest improvement was achieved using the HD–Road rule on the G0p model with the best overall classification accuracy at 89.99% (kappa: 0.87). This represented 4.1% (kappa: 0.045) improvement over that of G0p without model selection.