Mapping 30-m annual vegetation height in Guangdong Province based on GEDI and SDC with a hybrid deep learning model

In this study, we developed a hybrid deep learning model (U-Swin-Net) that integrates GEDI and multi-source remote sensing data, mainly from Seamless Data Cube (SDC), to produce annual vegetation height maps at 30 m resolution for Guangdong Province from 2015 to 2022. The resulting maps show strong spatial consistency and stable temporal patterns across different years. With the proposed GEDI preprocessing strategy, the maps effectively distinguish vegetation from non-vegetation areas without relying on external land-cover products during post-processing, enabling reliable estimation and monitoring of spatiotemporal changes in vegetation height. Compared with the GLC_FCS30D land-cover product, the derived vegetation classification reaches annual precision above 97%. When evaluated with the reserved GEDI test set, the estimated vegetation height achieves R² = 0.632, RMSE = 3.785 m, and MAE = 2.981 m. When compared with field measurements, the results show R² = 0.448, RMSE = 3.386 m, and MAE = 2.709 m. The proposed SWH Loss effectively mitigated the underestimation of tall trees and kept high accuracy for common vegetation heights. Specifically, for vegetation heights exceeding 30 m, RMSE and MAE were reduced by 1.364 m and 1.665 m, respectively. Furthermore, comparisons with independent GEDI observations from non-training years demonstrated that our maps achieved higher consistency than existing vegetation height maps, providing evidence for the temporal transferability of the framework. The annual vegetation height maps are distributed as GeoTIFF files in float32 format with a spatial reference system of WGS84 (EPSG:4326) and a spatial resolution of 30 m.