Abstract

The mapping and monitoring of forest canopy height over large areas at a fine, deca-metric spatial resolution is important for the quantification of carbon budgets in forests, the assessment of tree growth rates and for the detection of deforestation events. The height information can in principle be derived from photogrammetric analysis of stereo imagery (SfM) as well as from airborne laser scanning (ALS) data. Both approaches, however, do not scale well and entail relatively large costs for data acquisition. We assess the suitability of open and free Sentinel-2 time series for mapping and monitoring of tree height using sparsely sampled, space-borne GEDI laser data as reference. Instead of relying on (monthly) compositie images as predictor variables, we encode the noisy and irregularly sampled Sentinel-2 times series into a few orthogonal, gap-free and information-rich representations. The representation learning is fully self supervised and based on an approach known as Barlow Twins. A simple neural net is used to map the derived representations to the target heights from GEDI . The trained network is afterwards applied to the entire region of interest to create gap-free forest height maps at 10m spatial resolution.