This paper considers the generalized empirical likelihood (GEL) estimation and tests of spatial autoregressive (SAR) models by exploring an inherent martingale structure. The GEL estimator has the same asymptotic dis-tribution as the generalized method of moments estimator explored with same moment conditions for estimation, but circumvents a first step estimation of the optimal weighting matrix with a preliminary estimator, and thus can be robust to unknown heteroskedasticity and non-normality. While a general GEL removes the asymptotic bias from the preliminary estimator and partially removes the bias due to the correlation between the moment conditions and their Jacobian, the empirical likelihood as a special member of GELs further partially removes the bias from estimating the second moment matrix. We also formulate the GEL overidentfication test, Moran's I test, and GEL ratio tests for parameter restrictions and non-nested hypotheses. While some of the conventional tests might not be robust to non-normality and/or unknown heteroskedasticity, the corresponding GEL tests can.