Mapping canopy heights from space using deep learning with Sentinel-2 time series and LiDAR data

Canopy height is one of the most important forest structural variable, generated by remote sensing for applications such as forest inventory, sustainable management, carbon assessments and disturbance monitoring. But generating accurate and frequent canopy height maps over large areas remains a challenge. Airborne laser scanning (ALS) provides highly reliable and detailed canopy height models, repeated acquisitions are often limited by cost and availability. With the spaceborne LiDAR from NASA’s GEDI (Global Ecosystem Dynamics Investigation) there is globally distributed relative canopy height observations but these are affected by noise and terrain-related uncertainties. This has created a gap for generating consistent, wall-to-wall canopy height products at annual timescales. With the growing availability of high temporal multispectral imagery from satellite missions such as Sentinel 2 raises the question to what extent dense optical time series can be used to support operational canopy height mapping when combined with LiDAR observations. In this work, we investigate the potential and limitations of using dense Sentinel-2 time series in fusion with LiDAR data for generating CHMs at 10m resolution across Bavaria. We downloaded and processed all available Sentinel-2 imagery for Bavaria from 2019 to 2024 (~9 TB) by correcting it radiometrically and geometrically and regridding it into a non-overlapping datacube structure. From this datacube, we generated multi-seasonal composites and interpolated time series to capture forest phenology at the pixel level. Using the Sentinel-2 time series products created, we trained a CNN based model (UNet) with (i) high-resolution ALS derived CHMs and (ii) GEDI waveform relative height metrics as reference data. Preliminary results demonstrate that integrating multi-seasonal Sentinel 2 information substantially improves model performance at generating annual CHMs at 10m reoslution. At the same time, we also highlight limitations related to the choice of training supervision data and that models trained with higher quality ALS based CHMs yield the most reliable canopy estimates whereas GEDI based supervision can introduce increased uncertainty in heterogeneous terrain and areas with limited footprint samples. We thus provide a technically workable, scalable and semi-automatic forest canopy monitoring approach which - once trained for a region - uses only open-scource data, making it highly reproducible.