Automatic machine-learning based water body detection algorithm as part of the hydrological conditioning of the TanDEM-X DEM

Abstract

The TanDEM-X mission provides a high-resolution global digital elevation model (DEM) which works as a base for data analyses where the Earths relief is a key source of information. The unedited version of the TanDEM-X DEM contains artefacts such as rough surfaces of open water due to decorrelation effects, which impair the derivation of streams and hydrological assessments. To counteract distortions inherent to the interferometric synthetic aperture radar acquisition and processing and to support the conditioning of the elevation data in order to improve hydrological analyses, a global water body mask (WBM) is generated. The presented technique is a major part of the pre-conditioning of the global TanDEM-X DEM for producing a second and refined version of the HydroSHEDS database. Its core is an automatic machine-learning (ML) based water body detection using a Gradient Boosted Decision Tree (GBDT) algorithm as underlying classifier with Bayesian hyperparameter optimization. The features considered by the algorithm are mainly based on the TanDEM-X datasets assisted by a global Normalized Difference Vegetation Index (NDVI) derived from Sentinel-2 data. The ML model is sustained by hyperparameter tuning in order to learn the spectral signatures and their varieties for water bodies and land surfaces, enabling accurate detection and classification. As sufficient training data is required and to avoid manual hyperparameter tuning, the automation generates training data by using a maximum WBM created from reference data which allows for automatic random sampling throughout the feature space to train the classifier. For each TanDEM-X tile, a classified WBM is generated by the automatic ML algorithm as well as an alternative WBM based on the WBM quality layer of the Copernicus (COP) DEM modified using the TanDEM-X dataset. Then, either the superior water body mask or a combination of both is selected. First results show that the classified WBM is able to improve the alternative COP WBM in approximately 45% of tiles processed so far, underlining the quality of the used ML algorithm. However, regional analysis uncovers large differences. While the COP WBM achieves better results in arid areas, the classified WBM is superior in northern regions. Frozen water imposes great challenges to water detection algorithms. The ML based approach shows large improvements in northern regions compared to the COP WBM as it has the flexibility of including the minimum observed amplitude preferably capturing open water. Arid regions often show too little water resulting in an insufficient number of training pixels to perform a classification.