Speaker
Description
Forests form the largest terrestrial ecosystem by area, and hold primary relevance in the carbon cycle.
Continuous and precise monitoring of its dynamics, is of paramount importance to study and quantify the impact of anthropogenic caused degradation from events, such as logging and farming activities, but also wildfires, floods or storms of increasing intensities and linked to the man-made climate change. In this context, the generation of readily available, precise and large-scale estimates of forests’ properties, and of their temporal evolution is essential.
Spaceborne Synthetic Aperture Radar (SAR) systems are invaluable to this goal, as they operate on the required geographical and temporal scales, independent of weather conditions or time of day.
In this work, we present some of our findings related to the synergistic use of satellite-acquired SAR imagery and deep learning algorithms for the estimation of bio-physical parameters.
We focus our investigation on the added value of interferometric SAR (InSAR) products in relation to the more commonly used backscattering information, by presenting the benefits and the challenges associated to available feature pools. Furthermore, we quantify the benefits of single-pass over repeat-pass InSAR acquisitions, emphasizing the relevance of current and future bistatic missions, such as TanDEM-X and Harmony, in comparison to their repeat-pass counterparts. Finally, we provide a basic comparison of our proposed deep learning approach with the more commonly used physical-based models, highlighting both the key benefits and the challenges associated with AI-based approaches. All of the presented results have been obtained using DLR’s TanDEM-X and ESA’s Sentinel-1 products, and have been validated over the tropical sites of NASA/ESA 2016 AfriSAR campaign in Gabon, Africa. The obtained results are extremely promising in view of large-scale mapping, displaying the distinctive advantage of requiring only a single input acquisition for inference.
