Abstract: Aiming at the inaccurate and inefficient results of traditional individual tree segmentation approaches for large-scale remote sensing point cloud data due to the challenges of data noise, data anomalies, dense canopy, tree occlusion, etc., a latent feature similarity driven two-stage individual tree segmentation network is proposed by combining PointNet++ and convolutional neural network in an end-to-end manner. In the pre-segmentation stage, the local and global features of the input point cloud data are hierarchically extracted, and the latent feature matrix of the point cloud data is constructed; thus their similarity matrix and the confidence map can be created by two network branches respectively. Here, the similarity matrix measures their similarity between each pair of sampling points and determines the sampling points with high similarity to be the same instance thus forming a prediction group, whilst the confidence map reflects the degree of confidence of the prediction of each instance and filters the prediction group with low confidence to ensure its segmentation reliability and accuracy. In order to further improve the segmentation results, an adaptive K-nearest neighbor (A-KNN) propagation module for point cloud data is introduced in the fine segmentation stage to smooth the pre-segmentation results and also to reduce the misjudgment of the individual tree segmentation of sampling points. Experimental results demonstrate that our proposed method can effectively employ the spatial information and local features of point cloud data, and also greatly improve the accuracy and robustness of individual tree segmentation results owing to A-KNN label propagation, such as the average accuracy of individual tree segmentation on our labelled dataset is 96.33%, and its mIoU is 0.920 6.