Vegetation Products from ICESat: A database of GLAS (Geosciences Laser Altimeter System) waveforms and a global map of forest canopy height (Invited)

Data from the GLAS sensor can be used to accurately estimate forest canopy heights. However, deriving these heights from GLAS data is complicated because (1) no single data product combines all sensor information relevant for vegetation, (2) height estimates are biased on sloped terrain and require calibration, (3) there are no established methods for ensuring spatial and temporal consistency of the height estimates, and (4) the sensor samples portions of the earth’s surface but does not create the wall-to-wall raster coverages expected by end-users. The first of these complications is specific to ICESat data, but any spaceborne lidar mission designed to map ecosystem structure must address the final three issues. The ICESat Vegetation Product (IVP) combines all of the vegetation relevant information from GLAS and addresses the final three data issues. We identified an index of forest height, Loreydc, which is the crown-area-weighted height of dominant and co-dominant crowns. This index can be estimated from field inventory, airborne small-footprint lidar, and spaceborne large-footprint waveform recording sensors. This allows us to relate it to regional and national field inventories (which can also be used to estimate aboveground biomass) and to calibrate estimates of the index from easily obtained airborne lidar data collections. Estimates of Loreydc require that background noise and return signal be separated using a threshold. While airborne waveform recording instruments can adopt a mission-specific threshold, a long lifetime satellite mission must consider the effect of changing characteristics of the laser over time along with changes in sensor sensitivity, atmospheric conditions and solar background strength. Coincident pairs of GLAS waveform observations collected within and between observation periods were identified, and the difference between the two sets of waveform indices was minimized using optimization. This allowed comparisons between observations from all possible pairs of observation periods and led to an approach for estimating threshold values from signal-to-noise ratios. For the foreseeable future, spaceborne lidar sensors will remain capable of sampling no more than a fraction of the earth’s vegetated regions. Complete coverage of ecosystem structure would accelerate the incorporation of these data into mapping and modeling. Developing this wall-to-wall coverage will require combining the lidar with a second data source like polarmetric or interferrometric SAR, but these data sources are not currently available globally. To develop the IVP raster product we used image segmentation of multi-temporal MODIS data. Source data and texture indices for each patch were used to develop equations to estimate patch height using those patches in which GLAS observations fell as training and testing datasets. Equations to estimate patch height had correlation coefficients of about 0.7 and residuals with RMSEs of about 4m, with prediction strength varying by ecoregion. Independent training and testing datasets had correlation coefficients and RMSEs within 10% of each other. Draft versions of a global map of Loreydc and the individual GLAS observations are available through the website ceal.cnr.colostate.edu.