Modeling forest canopy height across heterogeneous ecoregions through the integration of Global Ecosystem Dynamics Investigation (GEDI) lidar with multispectral and SAR imagery.

The Global Ecosystem Dynamics Investigation (GEDI) lidar instrument has been collecting data on vegetation structure since 2019. Efforts are underway to understand the effects of tuning different GEDI parameters (e.g., ground algorithms, sensitivity, power beam, slope, time of collection) in the error reduction of canopy height models using multispectral and SAR predictors. Error variation in canopy height models as a function of environment and forest type has been less evaluated. Colombia is comprised of five natural regions, each presenting specific environmental conditions and forest characteristics. We created canopy height maps for each of the five natural regions of Colombia at 25m spatial resolution. We used the relative height at the 95 percentile of returned energy (RH95) of GEDI footprints as ground truth measures for canopy height and temporal and textural metrics of Sentinel-1, Sentinel-2 and PALSAR as predictors. Five-fold cross-validation presented significant differences for RMSE and MAE among regions (RMSE: F=1589; p= 0.001 and MAE: F=882; p= 0.001). The regions with the highest annual rain and/or topographic slope, Choco and Andes, showed the highest errors (RMSE = 8.23m 8.94m and MAE = 5.55m 5.8m) while the region with the lowest topographic and slope variation presented the lowest errors, the Orinoquia (RMSE = 4.8m and MAE = 3.07m) (Figure 1). These differences suggest that integrating regional maps (Figure2), as opposed to fitting one area-wide map, may reduce errors in canopy height models that encompass large heterogeneous regions. We believe that our methods can be apply to map other GEDI forest structural variables. Figure 1. RMSE (Root Mean Squared Error) and MAE (Mean Absolute Error) estimations for the regional maps of canopy height (in meters). Figure 2. Canopy height (in meters) map of Colombia.