In land use/land cover classification, the utilization of large-scale routine approaches in diverse

areas often fails to obtain satisfactory results. In this paper, the Qiantang river watershed was chosen as the

study area. A stratified and regionalized supervised classification (the Maximum Likelihood Classification)

approach was employed. With this approach, water and mountain areas were first stratified and extracted through a

set of equations that were used to compute parameters from Landsat TM bands. Subsequently, by using the mask

method, the authors obtained plains and foothills, which were subdivided into six sub-regions according to the

geomorphic features and land use/cover characteristics. Additionally, the plains and foothills should be

classified separately in the case the images were acquired in different seasons. The supervised classification

could be carried out after respective signatures in every region were identified. The classification accuracy

reached 90. 7% with a Kappa coefficient of 0.881, which was much higher than that obtained from the routine

classification approach that had a classification accuracy of 51.6% and a Kappa coefficient of 0.411. This study

shows that the stratified and regionalized approach is very efficient in land use/cover classification in a fairly

large region, such as the watershed level in southern China.