Second-order blind deconvolution of single input multiple output (SIMO) finite impulse response (FIR) channels is considered. A major drawback of several blind diversity techniques using antenna arrays/temporal-oversampling is high sensitivity to the choice of the model order. In this contribution, a robust method using only the second-order statistics is described. It provides high and robust estimation accuracy even for a limited sample size and an unknown model order. In contrast to other suggested approaches, that often exploit properties valid only in the large sample case, the proposed method is applicable both in the large sample and high signal-to-noise ratio (SNR) scenario.