Assessing hurricane-driven forest changes using satellite-based lidar and multispectral imagery

Abstract

Coastal forests in Alabama and Florida are increasingly vulnerable to storms and hurricanes, which can cause widespread canopy loss, decline in forest structure, and long-term disruption of ecosystem services. Accurate and scalable methods for assessing hurricane-driven forest damage are essential for understanding ecological impacts, informing restoration efforts, and guiding long-term forest resilience planning. This thesis addresses the pressing need for high-resolution, spatially comprehensive tools to monitor hurricane-driven forest changes by leveraging spaceborne lidar, multispectral imagery, and machine learning techniques. In the first study, we evaluated the Ice, Cloud and land Elevation Satellite-2 (ICESat-2) land and vegetation height product, or ATL08, at two resolutions (100 m segment and 20 m sub-segment), to assess pre- and post-hurricane changes in canopy structure using repeat-track observations within the damage extent of Hurricane Sally (September 2020). Strong agreement was observed between ATL08-derived 98th percentile canopy height (RH98) and airborne lidar data RH98, particularly at the 20-meter sub-segment scale (RMSE = 3.44 m, r = 0.80). Structural analyses revealed significant height reductions in mature trees (−1.51 m) and increases in understory vegetation (+0.77 m), reflecting both canopy damage and regeneration. Building upon these findings, the second study entailed the production of pre- and post-hurricane canopy height maps, changed canopy height maps, and associated thematic transitions. Machine learning algorithms, primarily Random Forest (RF) and Extreme Gradient Boosted (XGB) regression models, were trained using ICESat-2 data, Sentinel-2 imagery, Landscape Fire and Resource Management Planning Tools, National Land Cover Database derived predictors, and topographic variables. The integration of ATL08 lidar with high-resolution imagery, such as data from Sentinel-2, presents an opportunity to develop spatially explicit hurricane assessment tools. The RF showed superior performance (R² = 0.44, RMSE = 4.30 m) compared to XGBoost (R² = 0.41, RMSE = 4.76 m), and predicted pre-hurricane canopy height maps showed strong agreement with an existing ICESat-2-derived 2020 canopy height product (r = 0.57, RMSE = 3.49 m). Landcover change analysis revealed shifts from evergreen forests to herbaceous, scrub, and barren classes, and woody wetlands to emergent herbaceous wetlands with mean canopy height losses of 2.3 to 5.2 m. Canopy cover analysis showed dense (>60%) and sparse (<30%) cover experienced severe canopy height loss (up to 8.3 m), while moderate covers (30–60%) were resilient. Together, these studies demonstrate the potential of repeat-track ICESat-2 observations, synergistic use of multi-temporal ICESat-2 and Sentinel-2, and machine learning–driven canopy modeling to assess changes in forest structure and hurricane-driven canopy changes. This integrated approach provides a robust, scalable methodology for post-hurricane forest damage assessments, as well as informing adaptive strategies in hurricane-prone coastal forest ecosystems.