We use CFAR detection when noise variance is not known which means estimate depend upon noise variance. As the decision threshold has to satisfy probability of false, we assume that testing is under $H_0$ rather than $H_1$, this makes our estimate biased. This in turn decreases detection probability $P_D$. Most text books use the idea of employing reference noise $\eta_R$ samples to avoid the dependability of ML(maximum likelihood) estimate on noise variance.
Can someone please explain what is the role of reference noise samples and how they are able to enhance the $P_D$?