LiDAR data widely replaces 2-dimensional data for geographic data representation because of its information complexity. One of the LiDAR data processing tasks is semantic segmentation which has been developed by deep learning models. These algorithms use Euclidean distance representation to express the distance between the points, whereas LiDAR data with random properties are not suitable to use this distance representation. Therefore, this study proposes the non-Euclidean distance representation which is adaptively updated using their covariance values. The proposed method results the accuracy of 75.55%, better than the baseline PointNet of 65.08% and Dynamic Graph CNN of 72.56% with the dataset from the author. This performance improvement is because of multiplication with the inverse covariance value of point cloud data increasing the points similarity to the class.