Wetlands which are the planet’s most important ecosystem have high scientific research -value and will bring us both social and economic benefits. However, duing to various natural and man made factors, more and more wetlands have converted to agricultural land and urban land. Now, the changes in wetlands’ area and quantity have caused public’s widespread concern. And wetland’s management and protection will benefit from the improvement of the wetland information abstraction’s precision. Improving the classification precision of the RS image is a difficult problem because of the small scale of remote sensing images. This paper which is about the wetland remote sensing images extraction is based on the LANDSAT ETM remote sensing data, and the result of the Wavelet Packet reconstruction will be used as the sample set of the Active Support Vector Machine .At the end of this paper, a comparative analysis of the experimental results will show between the single classification (SVM, BPNN) method and the solution which is proposed in this article. This method can be proved to obtain very good classification results through many experiments on remote sensing image classification I’ve done. Experimental results show that this algorithm’s classification accuracy is better than the single classification’s. Moreover, in the active learning process, the bad influence of the image’s isolated and intersection points on the classification is avoided, and the number of training samples are reduced greatly.