Abstract: Abstract: The point cloud data of a gully region in loess plateau via Terrestrial Laser Scan (TLS) was characterized by uneven distribution of laser footprints, rapid geomorphologic change, and high density of herbaceous vegetation. In order to improve the precision of gully DEM, this paper proposes a vegetation filtering method of TLS point cloud. We first use laser return intensity to make an applicable classification。It is significant to compensate intensity attenuation which is brought by distance, angle of incidence, and environment, and establish a unified relationship between object and return light intensity. Available data indicates that return intensity is represented by an inverse second-order-dependent function of distance and other parameters can be treated as a constant in one experiment. We built a distance attenuation model of return light intensity. We can calculate the attenuation factor based on it and then compensate for laser return intensity of the whole point cloud. In this study, the return intensities of six sphere targets are used to build an attenuation model, and we obtained the attenuation factor as 0.3173. With the unified return intensity, each point's intensity deviation with intensity of the ground was used as a weight to enlarge the difference of non-ground points and ground points. Then we used segmentation and surface fitting method to calculate each point's distance to the trend surface, and set the threshold to separate the ground points and vegetation points. In this study, we propose an adaptive mesh grid filtering method integrated with return light intensity. In this method, we updated the distance to the trend surface though each point's intensity weight which has a linear relationship with its intensity deviation. Besides, the adaptive segmentation is more fast and effective than the K neighborhood search method. The method's reliability was tested through a point cloud acquired from a typical gully in Qiaozi Valley, Tianshui City of Gansu Province. It was located in 105°43'2''E, 34°36'59''N. The section is a typical V shape, and 90% of the surface is covered with low vegetation like grass and leguminous plant. The gully was scanned with a Leica HDS6100 3D laser scanner with a precision of 1mm. The cloud data containd 1,498,191 points, return light intensity varied from -2048 to -2047, covered an area of 14.0975 m2 and the average point density was 3345.1 points/m2. We practiced the adaptive mesh grid-filtering algorithm in Matlab, and iterated three times to get filtering result. There were 160 517 vegetation points which were removed from data, and they were 34.04% of the whole point cloud. Comparing the filtered DEM with the original one, we proved that this filtering method can overcome negative influences of uneven terrain and high vegetation coverage and filter efficiently. It reduces the elevation root-mean-square error (RMSE) of point cloud from 0.1430 to 0.1324, and most rags of original contour lines decrease and ground characteristics are well preserved. In addition, we compared this method with the intensity-classification method and surface fitting method, and found that this paper's filtering method performs better. Furthermore, we got a typical gully vegetation probability distribution of Qiaozi Valley by calculating the cumulative probability of filtered and original DEMs' deviation. It can explain the gully morphologic change, and the change coincides with the observation data. It proved that vegetation is effective in sand binding and reduction and slope impaction. Consequently, this study not only provides a new approach to filter gully vegetation in point cloud and acquiring high-precision DEM, but also helps to set the stage for future research to monitor gully morphologic change and control soil erosion.