Performance analysis of similarity measures between multichannel optical and multipolarization radar images

This paper investigates the problem of measuring similarity between multimodal Remote Sensing (RS) images using both area-based and feature-based structural similarity measures (SMs). For many RS platforms, optical image is multichannel and radar image is multipolarization. Thus, vector-to-vector SMs could be applied to optical-to-radar image pairs in contrast to scalar-to-scalar SMs considered in the literature. Using two real Landsat8 - SIR-C image pairs, we demonstrate that vector variants of state-of-the-art SMs outperform their scalar counterparts. This is especially evident for Normalized Correlation Coefficient (NCC), which in vector case performs as good as or better than advanced structural SMs.