A critical need in the development of any automatic target recognition system is the ability to accurately predict and quantify the detection and classification performance under various operational and environmental conditions. In this paper, we propose new methods capable of predicting and estimating the performance of a multichannel detection system using multiple synthetic aperture sonar. Performance prediction and estimation is accomplished by analyzing the multichannel coherence statistics and characterizing the background conditions within the image. The ability of the method to provide an assessment of image complexity for various background conditions is studied. The saddlepoint approximation is employed to approximate the empirical null distribution of the test statistics for threshold selection and to achieve a prescribed false alarm rate. Test results on two real and one synthetic sonar imagery data sets with different target and background conditions are provided, which indicate the capability of the proposed methods in describing the distribution of the likelihood ratio and predicting the detector's performance in low to medium clutter environments.