A country-level deep-learning approach for canopy height estimation from TanDEM-X InSAR data

Abstract:
The estimation of forest parameters, such as canopy height and above-ground biomass (AGB), at large scales is of paramount importance for forest disturbance analysis, carbon-cycle modelling, wild fire propagation simulations and resource inventorying. In this work, we propose a fully convolutional deep learning architecture, trained with a dedicated data set of TanDEM-X features, to generate wall-to-wall forest parameters products. We investigate the challenges imposed by the proposed deep learning approach for large-scale applications, concentrating in particular on the design of an effective training dataset on the basis of theoretical requirements on single-pass InSAR theory. We test the regression performance of our approach over the five tropical regions mapped by the ESA/NASA AfriSAR campaign in Gabon, Africa. The obtained CHM estimation accuracies are extremely competitive with those of the state-of-the-art methods, with the advantage to be achieved with only a single input TanDEM-X bistatic pair.