The preprocessing of the Microarray data is a hot topic in the bioinformatics research. The key point of a successful preprocessing method is to remove the noise of nonspecific binding and to keep the information of specific binding as much as possible. One way to solve these problems is to understand the principle of the binding between probes and target sequences, and to distinguish specific binding from nonspecific binding correctly. In this paper, we introduce MM probe intensities into position dependent nearest neighbor (PDNN) model, which contain much information of nonspecific binding.We use two-step model to estimate the parameters,which can simplify the computation. Based on the Wilcoxon rank test, we can determine whether a gene is present, with which we can obtain the training data set for the specific binding and non specific binding parameters. We also apply our model to gene expression data (HGU133plus2.0 and HGU133A) . We find that all these improvements increase the precision and stability, and show better result compared to the other four methods( Mas5.0, dChip, RMA and PDNN ).