Training Data

We used two different training data sources to model forest structure. We predicted canopy height and % canopy cover using the Global Ecosystem Dynamics Investigation (GEDI) space-based LIDAR data. Specifically, we used the Level 2A Geolocated Elevation and Height Metrics product available in the Google Earth Engine data catalog. This dataset contains monthly summaries of canopy height and % canopy cover for paths flown by the mission globally from 2019 to present. We selected the 95th percentile of both metrics in the month of July (corresponding to the month of our Landsat covariates) as our response variables. From across our modeling boundary, we sampled 5000 observations of each variable evenly across the years 2019-2022.

In addition, we also derived a training dataset for tree diameter following a 'transfer learning' approach, with the goal to 'learn' to predict tree diameter from the LEMMA GNN QMD_DOM model, which predicts the quadratic mean diameter of all dominant and codominant trees. We extracted 5000 observations of the QMD_DOM variable distributed across our modeling region, sampled evening across the years 1990, 2000, 2005, 2010, and 2015.