Landform classification is one of the most important steps tor eveal the mechanisms of surface matter flows and energy conversion, which could inform the scale and layout of human construction activities. However, traditional landform classification methods based on Digital Elevation Model (DEM) often use a small number of topographical derivatives or landform characteristics, resulting in insufficiently precise classification results. However, object-oriented landform classification performs better in that reliable classification can be achieved by maximizing the homogeneity within and between objects. But how to set conditions in object segmentation remains a challenge. In this paper, a geomorphological classification method based on watershed unitwas proposed, by accounting for many characteristics of watershed unit including statistics of basic topographic factors, feature point and feature line, basin and texture characteristics. Firstly, hydrological analysis based on DEM divided the study area into different small basins as the experimental units. Then, 29 features were extracted within each unit to represent watershed morphology using digital terrain analysis; feature selection and parameter calibration were carried out based on Random Forest (RF) algorithm. RF is a supervised integrated learning model aggregating different outputs of a single decision tree to reduce variances that may lead to classification errors in the decision tree. Finally, the watersheds in training set were selected to train the RF classifier, and the trained classifier was used to classify the landform of the whole study area, based on which we studied the landform spatial differentiation pattern. This experiment achieved good results in the landform classification of the Loess Plateau in northern Shaanxi Province. It is one of the areas with the most serious soil erosion and the most fragile eco-environment in the world. Most of them are covered by thick loess, and the topography is fluctuant. Result shows: (1) Compared with manual interpretation, excellent classification results based on small watershed in the study area were obtained, with the classification accuracy reaching 85% and the Kappa coefficient 0.83. (2)All small watersheds were divided into eight types of landforms. The same type of landforms showed obvious spatial aggregation. There were boundaries and transitional zones between different types of landforms. (3) Different geomorphological regions explained different situations of loess deposition and runoff erosion in different regions. Our findings suggest that the combination of RF algorithm and DEM data can achieve better classification results.