Artificial neural network (ANN) is an important part of artificial intelligence, it has been widely used in remote sensing classification research field. Wetlands remote sensing classification based on ANN is difficult, because of the complex feature of wetlands areas. The purity of training samples for remote sensing image supervised classification is difficult to guarantee that will affect the classification results based on ANN. This article proposed a method for sample purification based on statistical analysis theory which could purify training samples for improved wetlands remote sensing classification based on ANN. The BP ANN with a nonlinear mapping function can give good classification results for complex areas. We selected a TM image of Honghe Wetlands National Nature Reserve as study material. First, we used the statistical analysis theory to remove noise in training samples; second, we used the original samples and purified samples to train the BP ANN separately, and produced two classification maps of TM image based on two trained BP ANN; finally, we compared the classification accuracy between the two maps. The results showed that BP ANN trained with purification sample improved the wetlands classification accuracy significantly.