This paper addresses the problem of matched-field source localization in the presence of uncertainties in the ocean environment. Because signal wavefront mismatch can cause anomalous source location estimates, development of robust localization methods is critically important. In this paper, a robust maximum-likelihood estimator is proposed. It is based on a decomposition of the field into predictable and unpredictable subspaces of the acoustic normal mode representation. The estimator uses the predictable subspace for source localization. Identification of the predictable modes is made according to the second-order joint statistics of the horizontal wave numbers. The performance of the method is evaluated and compared to other matched-field methods using simulations and acoustic array data from the Mediterranean Sea. In the presence of mismatches, the algorithm has superior probability of correct localization than the maximum-likelihood, matched-mode-processing, and Bartlett methods.