In this paper, we present an approach for producing side scan sonar image mosaicking under a robust SLAM scheme. A Pose Graph based SLAM algorithm is used to perform a correction over the sensor trajectory for enabling image registration, using observation constraints extracted from the images. However, due to the operational context, the available odometry data carries a high degree of uncertainty and could also be discontinued, making the process highly sensitive to false observation data associations. As the automation of the feature extraction and pairing process will eventually produce false positives, we suggest an approach based on the use of switchable constraints and a coarse modeling of the sensor motion when odometry is not recorded for addressing this issue.