Land Cover mapping by combining GEDI-derived vertical metrics with SAR and multispectral data

Abstract

Accurate land cover mapping is essential for understanding ecological processes and informing sustainable land management. This study presents a two-stage remote sensing framework that first generates wall-to-wall maps of four key vegetation structure variables [canopy height (RH95), foliage height diversity (FHD), plant area index, and aboveground biomass (AGB)] using random forest models trained with Global Ecosystem Dynamics Investigation (GEDI) footprints and satellite-derived predictors. The second stage integrates these maps with multispectral (Sentinel-2) and synthetic aperture radar (SAR) (Sentinel-1C-band and ALOS-2/PALSAR-2 L-band) data to improve land cover classification across two contrasting landscapes in mainland Portugal. From the first stage, the resulting canopy height maps were independently validated using airborne light detection and ranging (LiDAR), showing strong agreement (R2 = 0.64–0.63; RMSE = 2.6–5.3 m). In the second stage, these structural layers were combined with seasonal and spectral-temporal metrics from Sentinel-2, as well as with texture and backscatter data from Sentinel-1 and PALSAR-2, totalling 230 predictor variables. Land cover classification achieved high accuracy (overall accuracy (OA): 96.0% and 91.7% in the two study areas), with GEDI-derived structural metrics significantly improving class separability, particularly among forest and shrubland types. Variable importance analyses highlighted the synergistic contribution of LiDAR, SAR, and spectral data, with FHD, vegetation canopy height (VCH), and AGB ranking among the most influential predictors. This study demonstrates the value of integrating spaceborne LiDAR with complementary remote sensing data for both vegetation structural modelling and high-resolution land cover mapping. The results reinforce the potential of GEDI in operational land monitoring frameworks, particularly in complex or structurally diverse ecosystems.