Automatic ML-based water body detection as part of the hydrological conditioning of the TanDEM-X DEM

Abstract:
The TanDEM-X mission provides a global digital elevation model (DEM) with high spatial resolution and therefore is able to capture the local geomorphic appearance of the world’s surface. In particular, the high quality and homogeneity enable new possibilities for hydrologic products, where the earth’s relief is a key source of information. As the TanDEM-X DEM is provided in an unedited version some distortions are still present and require pre-processing. In particular, decorrelation effects at open water areas cause a noisy and rough appearance of naturally flat surfaces, which impair hydrological assessments. In this paper an automated global water body classification is presented, utilizing TanDEM-X synthetic aperture radar, optical and terrain data to generate a water mask tailored towards these areas in the DEM. The generated water mask is a vital product for the hydrologic conditioning. The state-of-the-art Gradient Boosted Decision Trees is chosen as the underlying classifier. It is combined with a Bayesian hyperparameter optimization to avoid manual tuning. For the generation of training data, different global water masks are evaluated and appropriate training pixels are identified. First results show the potential of this approach and particularly the advantage in detecting water north of 60deg degrees, when using the lowest Amplitude measured in the feature space.