Land type is the base of land change analysis or landscape analysis. Land type classification is often based on land resources survey. Updating land type is generally difficult, mainly due to the lack of appropriate information. Hence, it is of importance to develop a method for land type classification using remote sensing images. The study was to propose the geographic image cognition (GEOIC) approach for land type classification. The approach was realized by the segmentation of land units, using Hyperion image, geographic information, vegetation, soil, DEM, and geosciences knowledge. It is the extension of the methodologies of object-based image analysis. Results showed that the GEOIC approach is an integrated approach with objectification cognition on remote sensing images and multi-source information using geo-knowledge. The GEOIC approach included three aspects: spatial feature perception, spatial object cognition and spatial pattern cognition. The use of the GEOIC approach in land type classification was tested in a study area in the agriculture-pasture mixed region of Loess Plateau in China. Results of land type classification at different scale levels showed that the overall accuracy ranged from 72.4% to 88.3%, with an average about 80%. The accuracy of classification at similar pixel level was relatively low, with an overall accuracy of 73.1% and Kappa coefficients of 0.69. The classification at scale level of 100 was effective for mapping land types with an overall accuracy of 88.3% and Kappa coefficients of 0.86. The classification accuracy through the segmentation of land units at an appropriate scale level was higher than that for pixel to pixel methods. This study concluded that the GEOIC approach on land type classification is significant and appears potential for land type classification aiming to land assessment and planning.