Multi-temporal Canopy Height Modeling using Deep Learning and Sentinel-2: The Influence of Training Data and Input Feature Representations

Precise information on canopy height is crucial to assess biomass and climate-induced structural changes within forests. Canopy height maps derived from optical satellite data can provide valuable wall-to-wall information, but often remain subject to substantial uncertainties. In this study, we systematically investigate the influence of training data and Sentinel-2 feature representations on the statistical performance of multi-temporal, deep learning-based canopy height modeling within Rhineland-Palatinate, Germany. We trained separate models using GEDI canopy height metrics or photogrammetric digital surface models (pDSM) as training data, and compared Sentinel-2 composites with spline coefficients derived from annual satellite time series as input representations. Hence, four different canopy height models were generated for the period 2018-2025, which we validated with independent airborne LiDAR data and compared with established global canopy height products. The best predictions were achieved using a combination of pDSM training data and Sentinel-2 spline coefficients, reaching a pooled RMSE of 3.14 m, compared to 4.06 m - 5.98 m for the other model configurations. This improvement was mainly rooted in the higher inherent accuracy of pDSM data, as well as the phenological information provided by the spline coefficients. The best-performing model also outperformed three global canopy height products and exhibited high temporal stability, supporting its suitability for a range of forest monitoring applications such as disturbance and recovery monitoring.