Tree species classification based on stem-related feature parameters derived from static terrestrial laser scanning data

Abstract

ABSTRACTTree species information is crucial for forest ecology and management, and development of techniques efficient for tree species classification has long been highlighted. In order to fulfil this task, a large variety of remote-sensing technologies have been attempted. Static terrestrial laser scanning (TLS) is such a representative case, which has proved to be capable of deriving explicit tree structure feature parameters (ETSPs) and has been primarily validated for tree species classification. However, in practice for each forest plot mapped by TLS, this kind of ETSP-based solutions can only work for the first circle layer of individual trees surrounding the TLS systems, because the trees at the outer circle layers tend to show incomplete crown representations due to the effect of laser obscuration. This adverse circumstance even may occur to the scenario of TLS-based inventory in the multi-scan mode. To break through this restriction, this study focused on tree stems that tend to be more readily mapped by TLS in the complicated forest environment, and then, their comparatively complete forms were used to comprehensively derive primarily stem-related feature parameters (SRPs) for distinguishing different tree species. Specifically, in this study 14 SRPs were proposed, mainly based on stem structure and surface texture characteristics. Based on a Support Vector Machine (SVM) classifier, the classification was operated in the leave-one-out cross-validation (LOOCV) mode. In the case of four typical boreal tree species, that is, Picea abies, Pinus sylvestris, Populus tremula, and Quercus robur, tests showed that the optimal total classification accuracy reached 71.93%. Given that tree stems generally display less features than crowns, the result is acceptable. Overall, the positive results have validated the strategy of fulfilling TLS-based tree species classification by deriving predominantly stem-related feature parameters, and this, in a broad sense, can expand the effective range of TLS on forest ecological studies.