The utility of dynamic forest structure from GEDI lidar fusion in tropical mammal species distribution models

Patrick Burns^1*Zaneta Kaszta²Samuel A. Cushman^2,3Jedediah F. Brodie^4,5Chris R. Hakkenberg¹Patrick Jantz¹Mairin Deith⁶Matthew Scott Luskin⁷James G.C. Ball⁸Jayasilan Mohd-Azlan⁵David F.R.P. Burslem⁹Susan Cheyne^10,3Iding Haidir^11,3Andrew James Hearn³Eleanor Slade¹²Peter J. Williams¹³David W. Macdonald³Scott J. Goetz¹

• ¹School of Informatics, Computing and Cyber Systems, Northern Arizona University, Flagstaff, Arizona, United States
• ²Department of Biological Sciences, College of the Environment, Forestry, and Natural Sciences, Northern Arizona University, Flagstaff, Arizona, United States
• ³Wildlife Conservation Research Unit, Department of Zoology, Mathematical, Physical and Life Sciences Division, University of Oxford, Oxford, England, United Kingdom
• ⁴Division of Biological Sciences, College of Humanities and Sciences, University of Montana, Missoula, Montana, United States
• ⁵nstitute of Biodiversity and Environmental Conservation, Universiti Malaysia Sarawak, Kota Samarahan, Sarawak, Malaysia
• ⁶Institute for the Oceans and Fisheries, Faculty of Science, University of British Columbia, Vancouver, British Columbia, Canada
• ⁷School of the Environment, The University of Queensland, Brisbane, Queensland, Australia
• ⁸Department of Plant Sciences, Faculty of Biology, School of Biological Sciences, University of Cambridge, Cambridge, England, United Kingdom
• ⁹School of Biological Sciences, University of Aberdeen, Aberdeen, Scotland, United Kingdom
• ¹⁰Borneo Nature Foundation International, Penryn, England, United Kingdom
• ¹¹Kaltim Lestari Utama, Samarinda, Indonesia
• ¹²Asian School of the Environment, College of Science, Nanyang Technological University, Singapore, Singapore
• ¹³Ecology, Evolution, and Behavior Program, College of Natural Science, Michigan State University, East Lansing, Michigan, United States

Species distribution models (SDM) often rely on remotely-sensed geospatial datasets to predict probability of occurrence and infer habitat preferences. Lidar measurements from the Global Ecosystem Dynamics Investigation (GEDI) are shedding light on three dimensional forest structure in regions of the world. Here we combine a large camera trap dataset of mammal species in Borneo and Sumatra with a diverse set of geospatial data to predict the probability of occurrence of 47 species. Multi-temporal GEDI predictors were created through fusion with Landsat time series. The availability of these GEDI-based forest structure predictors and other temporally-resolved predictor variables enabled temporal matching of species occurrences and hindcast predictions of species probability of occurrence at years 2001 and 2021. Our GEDI-Landsat fusion approach worked well for forest structure metrics related to canopy height (relative height of the 95th percentile of returned energy R2 = 0.62 and relative RMSE=41%) but, not surprisingly, was less accurate for metrics related to interior canopy vegetation structure (e.g. plant area volume density from 0 to 5 m above the ground R2 = 0.05 and relative RMSE=85%). For the SDM analyses, we tested several combinations of predictor sets and found that when considering a large pool of multiscale predictors, the exact composition, and whether GEDI Fusion predictors were included, didn't have a large impact on generalized linear modeling and Random Forest model performance. Adding GEDI Fusion predictors to a baseline set only meaningfully improved performance for some species (n=4 for RF and n=3 for GLM). However, when GEDI Fusion predictors were used in a smaller predictor set that is more suitable for hindcasting, more SDMs showed meaningful performance improvements relative to the baseline model (n=9 for RF and n=4 for GLM) and the relative importance of GEDI-based canopy structure predictors increased relative to when they were combined with the baseline predictor set. Moreover, as we examined predictor importance and partial dependence, the utility of GEDI Fusion predictors in hindcast models was evident in regards to ecological interpretability. We produced a catalog of probability of occurrence maps for all 47 mammals species at 90m spatial resolution for years 2001 and 2021.