Silvilaser 2019 - Poster Presentations »
Mapping forest biomass across the US using national forest inventory data, spaceborne lidar and a two-stage hierarchical model
Satellite lidar missions, such as GEDI and IceSat-2, are designed to collect laser altimetry data from space for narrow bands along orbital tracts. As a result lidar metric sets derived from these sources will not be of complete spatial coverage. This lack of complete coverage, or sparsity, means traditional regression approaches that consider lidar metrics as explanatory variables and do not account for error cannot be used to generate wall-to-wall maps of forest inventory variables. We implement a two-stage hierarchical spatial modeling framework to produce wall-to-wall (30m) predictions of forest aboveground biomass (AGB) across the contiguous US. The first stage models the sparsely sampled lidar to make wall-to-wall predictions using Landsat variables (e.g., NDVI). The second stage predicts forest AGB wall-to-wall using the stage-one lidar predictions as explanatory variables. Because the model is fitted using Bayesian inference, we can propagate stage-one lidar prediction uncertainty through to the stage-two predictions of forest AGB. This helps ensure the validity of any model-based confidence/credible intervals generated for predictions. We inform the model with USFS Forest Inventory and Analysis (FIA) forest measurements and GLAS spaceborne lidar to spatially predict AGB across the contiguous US. To circumvent computational difficulties that arise when fitting complex geostatistical models to massive datasets, we use a Nearest Neighbor Gaussian process (NNGP) prior. Results suggest that a two-stage hierarchical modeling approach to leveraging sampled lidar data to improve AGB estimation is effective. Further, fitting the hierarchical spatial model within a Bayesian mode of inference allows for AGB quantification across scales ranging from individual pixel estimates of AGB density to total AGB for the continental US with uncertainty. The two-stage hierarchcial spatial modeling framework examined here is directly applicable to spaceborne lidar acquisitions from GEDI and IceSat-2. Pairing these lidar sources with the extensive FIA forest monitoring plot network using a two-stage prediction framework offers the potential to improve forest AGB accounting certainty and provides maps for analysis of the spatial distribution of AGB.