Talks and Poster Presentations (without Proceedings-Entry):
"Advancing the Understanding of Spaceborne Radar Observations using Machine Learning and Physical Models";
Keynote Lecture: 1st workshop on Complex Data Challenges in Earth Observation (CDCEO) 2021,
Earth observation satellites equipped with radar sensors have become indispensable for monitoring the land surface from local to global scales. In particular, the Sentinel-1 satellites series represents a breakthrough as it acquires new radar imagery with a spatial resolution of 20 m every few days. Since the launch of the first Sentinel-1 satellite in 2014, several Petabyte of data have already been collected. Given the large data volume and the difficulties in the interpretation of the radar data, machine learning (ML) techniques are increasingly used alongside physical models to derive geophysical data sets such as soil moisture, water bodies, or forest type. In this contribution, I review a number of studies that have used ML techniques such as Support Vector Regression, Gradient Boosted Regression Trees, or Long-Short Term Memory for the classification and interpretation of Sentinel-1 radar data. The success of all these studies was routed in the availability of high quality reference data for ML training and validation. The ML models often worked surprisingly well, however, sometimes for the wrong reasons. In addition, the limits of the applicability of the trained models are often not clear. Therefore, the usefulness of ML for turning radar measurements into higher-value data sets depends critically on the context and intended application of the data. Further improvements in our understanding of spaceborne radar data can be expected from the combination of physical models and ML approaches.
Created from the Publication Database of the Vienna University of Technology.