Abstract: Abstract: In this paper, SfM-MVS (structure from motion with multi-view stereo) method was introduced to construct a high precision terrain model of the typical gully on the Losses Plateau of northern Shaanxi in China, with oblique aerial photos acquired by a COTS (commercial off-the-shelf) UAV (unmanned aerial vehicle) (DJI INSPIRE-1) and 30 high-precision pre-deployed ground control points (GCPs) measured by FIFO A30 RTK (real-time kinematic). A sequence of 194 oblique photos were captured by UAV camera with 70° pitch angle following a dual-Z shaped flight route, which were in comparison with 74 orthophotos captured by a nadir-point UAV camera in single Z shaped flight route. The photos were imported into PhotoScan software for terrain construction along with POS (position and orientation) information. Firstly, a preliminary alignment of aerial photos was performed as well as a rough estimation of camera parameters. The RMS (root mean square) reprojection error of tie points was 0.808 pixel and the maximum reprojection error was 41.143 5 pixel. Secondly, the corresponding projections of GCPs were marked on each photo and a set of GCP references were established in PhotoScan. Thirdly, camera estimation was iteratively optimized with high precision GCP references until errors of GCPs and reprojection errors of tie points met desired standard. After 4 iterations, the GCP errors were stabilized and its reprojection error was down to 0.538 pixel, and the RMS reprojection error of tie points also decreased to 0.51 pixel and its maximum reprojection error was down to 7.8 pixel. Fourthly, based on the optimized camera parameters and original aerial photos, depth image of each photo was calculated and a dense gully point cloud model consisting of 9 537 948 points was built through PMVS (patch-based multi-view stereo) algorithm in PhotoScan. And fifthly, the 30 GCPs were classified into 2 categories; 10 GCPs that numbered multiples of three were selected as check points to evaluate the overall accuracy of gully model, while the others were used as geo-reference to the WGS-84 system. Finally, the georeferenced gully dense point cloud was rasterized to gully DEM (digital elevation model) and triangulated to gully TIN (triangulated irregular network) from which the gully DOM (digital ortho-photo map) was built. Afterwards, the accuracies of oblique photos derived gully point cloud model as well as DEM and DOM results were analyzed and compared with that of ortho-photos. Box plot of GCP errors verified the consistency between the X/Y/Z errors of check GCPs and reference GCPs in oblique photo result. Thus the constructed gully dense point cloud has a roughly 0.096 m planimetric RMSE and 0.018 m vertical RMSE, which fulfills the requirement of DLG (digital line graphic) and DOM in 1:500 scale (GH/Z 3003-2010). The achieved accuracy well meets the requirement of high precision gully modeling and monitoring. The result of ortho-photos has a 0.105 m planimetric RMSE and 0.036 m vertical RMSE. That confirms a 50% improvement of vertical accuracy with oblique photos. The overall density of the gully point cloud is 12 680 points/m2, which is comparative to terrestrial laser scanning, and it avoids the uneven sampling caused by multi-station assembling of TLS (terrestrial laser scanning). Except from a small patch of inner part in concave gully head, there aren't obvious holes in gully head or walls in the model. In addition, the area covered with vegetation can also be correctly constructed. On the contrary, there are open holes in concave gully head and vegetated area in ortho-photo result. Seen from the DEM and contour lines map of the gully, the constructed terrain model gives a fine description of gully morphology. However, there are burrs and breaks in contour line caused by vegetation outside the gully, which shows the necessity of vegetation removal. In general, the proposed method has advantages in high precision gully terrain modeling and shows great ability to further application.