The landuse classification accuracy based on single remotely sensed data and simple supervised classification is unsatisfactory to the landuse investigation in loess hill and gully area. By taking the Wuding River watershed in Northern Shaanxi Province as a test area, the TM mult-i spectral data and SPOT pan data were merged by the method of Principal Components Analysis. Based on the merged image, the landuse categories were then extracted by applying an integration of supervised classification and unsupervised classification. The combination of two methods remarkably improved sampling method. Compared to the classification based on the single TM mult-ispectral data and supervised classif icat ion, the total accuracy increased from 82.0% to 89.2% , especially the accuracy of city and town area, paddy field, water area increased over 10%, the mixture of sloping land and forest (grassland)decreased remarkably, and the accuracy of the two categories increased over 5%. The result is of critical significance in landuse dynamic monitoring in the area.