Local Prediction of Temperate Forest Structure in Eastern North America Using LiDAR, Radar, and Optical Data

Chenyang Wei, Colin Sweeney, Trevor Roberts, Hikaru Keebler, Daniel Fink, Benjamin Zuckerberg, Marta A. Jarzyna, Kaiguang Zhao

Forest structure underpins the emergence of ecological patterns and processes yet remains costly and labor-intensive to measure at broad scales. NASA’s Global Ecosystem Dynamics Investigation (GEDI) mission provides three-dimensional Light Detection and Ranging (LiDAR) measurements at discrete footprints, leaving spatial gaps that complicate wall-to-wall mapping. Few studies have produced high-resolution, broad-extent predictions of multiple GEDI-derived metrics while explicitly accounting for spatial nonstationarity in predictor–response relationships. We addressed this gap with a local modeling framework that predicted 11 GEDI-based structural metrics at 30-m resolution across temperate broadleaf and mixed forests of eastern North America (1.17 million km2) for 2019–2022. Using Google Earth Engine, we integrated Landsat and Sentinel-2 multispectral imagery, Sentinel-1 synthetic aperture radar, and auxiliary variables (topography, land cover, leaf traits, and soil properties) to derive 93 environmental covariates. We partitioned the study area into 1,693 overlapping tiles, trained tile-specific random forest (RF) models with 80% of GEDI observations, and aggregated overlaps using weights based on model performance and pixel location. Across all metrics, local model predictions correlated strongly with GEDI measurements (Pearson’s r > 0.65). On the 20% held-out test set, median R2 of local models exceeded 0.4 for seven metrics, with canopy height and canopy cover both reaching 0.63. Sentinel-2, topography, and Landsat ranked among the most important predictor groups in at least 69.6% of local models for each metric. Across 30 randomly sampled tiles, local models outperformed a single global RF model in 56.7% of cases, with the largest gains where the global model performed worst. Our results show that integrating spaceborne LiDAR with multisource environmental covariates in a local modeling framework delivers robust predictions of forest structure and offers a transferable approach across broad geographic regions.