Mapping mangrove species using combined UAV-LiDAR and Sentinel-2 data: Feature selection and point density effects

Mapping mangrove species is important for mangrove conservation and rehabilitation. However, due to the similar spectral signatures of most mangrove species, accurately mapping mangrove species remains an ongoing challenge. Unmanned aerial vehicles paired with light detection and ranging sensors (UAV-LiDAR) have the potential to increase separability between mangrove species by providing vertical structure information with high flexibility and low cost and could be complementary to optical imagery. In this study, we combined UAV-LiDAR and Sentinel-2 data to classify mangrove species in two representative mangrove forests in China. Moreover, an architecture diagram of important features for discriminating mangrove species was first constructed using a modified recursive feature elimination method and a proposed step-by-step method, including hierarchical clustering, scatterplots, and boxplots. This architecture diagram could indicate the specific role of each feature (or group) in demarcating specific mangrove species and help other mangrove studies prepare candidate features. The combined UAV-LiDAR and Sentinel-2 data produced the highest overall accuracies of 85.60% and 91.61% in the Dongzhaigang National Nature Reserve and Qinglangang Provincial Nature Reserve areas, respectively, compared to those of the Sentinel-2 (80.00% and 80.42%, respectively) and UAV-LiDAR (77.60% and 75.52%, respectively) data alone by using the novel canonical correlation forest algorithm. Important LiDAR features included metrics describing the top, bottom, and overall (e.g., Hstd and GM2nd) morphological characteristics of the mangrove canopy, which characterized the vertical stratification and shape of the canopy. The added value of the UAV-LiDAR data mainly lies in improving the separability between shrub-like mangroves and arbor mangroves that had similar spectral signatures, such as A. marina and R. stylosa. The point density might have little effect on feature selection and classification accuracy, while 10% of the initial point density (∼10 pts/m2) could produce similar accuracy (differences less than 0.5%). This study suggests that the combined UAV-LiDAR and Sentinel-2 data provide a cost-efficient and accurate approach for mangrove species classification.