OOS 5-6 - Vegetation structure from spaceborne lidar in species distribution modeling

Background/Question/Methods

Current rapid loss of species diversity linked to habitat loss and fragmentation requires effective actions to identify and quantify areas supporting high biodiversity as well as habitats of high importance for key wildlife species. Insofar, range-wide species distribution models (SDM) have been defined by only two-dimensional characteristics of the vegetation ignoring vertical vegetation structure which is notoriously hard to measure with remote sensing. However, there is increasing evidence that information on canopy structural heterogeneity is crucial to understand the factors driving species distributions, particularly in dense humid tropical forest ecosystems. Our objective was to integrate The Global Ecosystem Dynamics Investigation (GEDI) Lidar datasets characterizing various aspects of canopy structure with other sources of geospatial data, combine with extensive in situ data on species observations to develop robust multi-scale SDMs of mammalian community in Borneo. We used an ensemble of multi-scale machine learning algorithms to optimize prediction of species occurrences and identify the factors most strongly related to their patterns of distribution. We tested more than 40 variables (e.g. vegetation structure, composition and phenology, productivity, climate, geomorphology, disturbance, and human influence) at several special scales likely to affect species distribution.


Results/Conclusions

We generated predictive habitat suitability layers for a set of mammalian species covering entire Borneo, identifying key biodiversity hotspots. The SDM models showed the importance of GEDI-derived canopy structure metrics in modelling distribution of forest depending species, especially those which occur in primary, tall and dense forests. We produced two sets of SDM for each species, one including GEDI structure variables and the other without it. In nearly all cases the models including GEDI vertical structure data outperformed those without it based on independent cross-validation. Also, the spatial pattern of the prediction was substantially different, as well as the variables and scales included. This shows that including vertical vegetation structure in SDM models improves their predictions and substantially changes the interpretation and spatial prediction of species habitat relationships. This has large implications for the use of SDM for both research and conservation applications, particularly in areas with high and complex structure in vegetation, such as the humid tropics.