Confronting remote sensing product with ground base measurements across time and scale

Abstract:

Ecosystem models are essential tools in forecasting ecosystem responses to global climate change. One of the most challenging issues in ecosystem modeling is scaling while preserving landscape characteristics and minimizing loss of information, when moving from point observation to regional scale. There is a keen interest in providing accurate inputs for ecosystem models which represent ecosystem initial state conditions. Remote sensing land cover products, such as Landsat NLCD and MODIS MCD12Q1, provide extensive spatio-temporal coverage but do not capture forest composition and structure. Lidar and hyperspectral have the potential to meet this need but lack sufficient spatial and historical coverage. Forest inventory measurements provide detailed information on the landscape but in a very small footprint. Combining inventory and land cover could improve estimates of ecosystem state and characteristic across time and space. This study focuses on the challenges associated with fusing and scaling the US Forest Service FIA database and NLCD across regional scales to quantify ecosystem characteristics and reduce associated uncertainties.

Across Southeast of U.S. 400 stratified random samples of 10x10 km² landscapes were selected. Data on plant density, species, age, and DBH of trees in FIA plots within each site were extracted. Using allometry equations, the canopy cover of different plant functional types (PFTs) was estimated using a PPA-style canopy model and used to assign each inventory plot to a land cover class. Inventory and land cover were fused in a Bayesian model that adjusts the fractional coverage of inventory plots while accounting for multiple sources of uncertainty. Results were compared to estimates derived from inventory alone, land cover alone, and model spin-up alone. Our findings create a framework of data assimilation to better interpret remote sensing data using ground-based measurements.