Channel state information is vital for exploiting diversity in cooperative networks. The existing literature on cooperative channel estimation assumes block lengths are long and that channel estimation takes place within a fading block. To reduce the forwarding delay, short block lengths are preferred and adaptive estimation through multiple blocks is required. In this paper, we consider estimating the relay-to-destination channel in decode-and-forward relay systems for which the presence of training data is probabilistic since it is unknown whether the relay participates in the forwarding round. A detector is used before the update of the least mean square channel estimate so that an update is made only when the detector decides the presence of training data. We use the generalized likelihood ratio test and focus on the detector threshold for deciding whether the training sequence is present. To improve the convergence speed and reduce the estimation error, a heuristic objective function is proposed which leads to an proper threshold. Extensive numerical results show the superior performance of using this threshold as opposed to fixed thresholds.