Most deep learning networks for Lidar semantic segmentation have been devoted to small-scale indoor data and only few of them have focused on large-scale outdoor data. To bridge this gap, this research explores the influences of changing features of deep learning networks on the accuracy of large-scale outdoor Lidar semantic segmentation. Surface normal information and random downsampling layers are the two features considered. Eight scenarios are designed to test them. Point clouds acquired from Kapiti Coast, New Zealand in 2021 with five labeled classes are used for training, validation, and testing stages. Mean intersection over union (mIOU) is the main metric in the validation and test. The findings show that the network adding surface normals with four random downsampling layers whose sampling ratios are 4, 4, 4, and 4 of those layers performs best because of its high mIOU. Moreover, IOU results reflect that the segmentation of buildings performs best between all tested classes.