In distributed information fusion systems, local estimation errors may be highly correlated arising from the common process noise and dependent measurement noises. The problem of estimation fusion based on sensor selection in the presence of the cross-correlation of local estimation errors is addressed. By introducing the objective function according to the fusion accuracy and the cardinality constraint of the selected subset, the sensor selection problem is turned into a combinatorial optimization one. The cross entropy optimization method is employed to solve it, which implements the sampling and updating the sample distribution alternately.