In forest inventories, the species information is crucial for economical, ecological and technical reasons. Species recognition is currently a bottleneck in practical remote sensing applications. Here, we examined species discrimination using tree-level LiDAR features in discrete-return data. The aim was to examine the robustness and explanatory power of the intensity and height distribution features. A dataset consisting of 13890 trees from 117 stands in southern Finland (61°50'N, 24°20'E) was used. The data of two LiDAR sensors was fused using intensity normalization in natural targets. Age dependency of first-return intensity was observed in spruce and birch trees, which needs to be considered in using LiDAR intensity metrics. Classification of Scots pine, Norway spruce and birch was tested and accuracy was 81−85%. Separation of pine and spruce was more accurate, 91−93%. We also present results for 15 rare conifer and broadleaved species. To enhance the classification accuracy of birch, we propose co-use of image features.