In the farming-pastoral area, the land types are complex and the mixed pixel phenomenon is serious, so the automatic extraction of wetland information is difficult. In view of the characteristics and difficulties of wetland remote sensing information extraction in these regions, the authors selected the TM remote sensing data covering Dacangxiang area in Duolun County to study a new method of wetland automatic extraction. Firstly, the NDVI threshold method was used to extract water body and then water types were extracted using water body shape features such as area (s), perimeter (p) and shape index (k). Afterwards, the linear spectral mixture model (LSMM)was chosen to extract wetland information after water removal. Finally, the meadow component which was decomposed by LSMM was amended by SPOT-5 satellite image to further extract high coverage, medium coverage and low coverage meadow. The results show that the overall accuracy of the wetland information extraction is 85.3%, and the Kappa coefficient is 0.79. Compared with the results of the traditional maximum likelihood method, the overall accuracy is improved by 19.9% and the Kappa coefficient is improved by 0.22. This method can provide a reference of remote sensing information extraction of wetland for other regions with mixed spectral characteristics.