Abstract: Remote sensing technology has become increasingly effective for forest mapping but its operational use in forest management and planning is still in its infancy. One of the most critical concerns is that remotely sensed forest attributes are not compatible with those traditionally defined in forestry practice. Tree species composition as a fundamental forest attribute is referred to by per-species tree volume or basal area proportion in conventional forestry but is quantified as tree counts or canopy cover percentage in remote sensing. These differences in the definition of tree species composition imply a barrier for effectively applying remote sensing in forestry decision-making. This study developed a remote sensing framework to derive tree species composition in a mixed-species, complex forest landscape based on tree attributes obtained by integrating UAV LiDAR and hyperspectral data. We classified 11 tree species with machine learning and obtained F-score values of 0.43–0.95. By incorporating tree species into tree diameter at breast height (DBH) prediction models, DBH was estimated with accuracy much higher than a general model of all tree species. The magnitude of increase in DBH-estimation accuracy was proportional to tree species-classification accuracy. Consequently, species composition coefficient estimation error was largely below 20% in the plots where forest type classification accuracy exceeded 90%. The error propagation from tree crown detection to DBH modeling cannot be overlooked for the integrated use of UAV LiDAR and hyperspectral data toward automatic, model-imbedded forestry-oriented surveys.

Key words: Tree species composition, UAV remote sensing, DBH estimation, Individual tree-based approach (ITA)