Hydrologically-Informed DTM Super-Resolution for Rapid Flood Depth Estimation

Kurzfassung

Reliable estimation of flood depths from satellite-derived inundation extent information critically depends on the spatial resolution and hydrological consistency of the underlying digital terrain model (DTM). Accurate, very high-resolution DTMs are typically not publicly available, difficult to access within the time constraints of rapid mapping, and lack consistent coverage. Although open-access DTMs such as the Forest and Buildings removed Copernicus DEM (FABDEM) provide global coverage, their coarse spatial resolution often fails to represent important small-scale terrain features that control flow paths, slopes, and local water accumulation. To address these limitations, this study proposes a deep learning framework for DTM super-resolution that combines low-resolution DTMs with optical satellite imagery by integrating hydrological knowledge into the training process to force the reconstruction of relevant topographic features for improved flood inundation depth estimation.

The proposed approach employs a residual channel attention network (RCAN) enhanced with optical satellite imagery as auxiliary input to upscale low-resolution terrain data. Central to the methodology is a collaborative hydrologic loss function that guides network optimization beyond elevation-based accuracy. In addition to the mean absolute elevation error (MAE), the loss integrates slope deviation and flow direction disagreement to focus the learning on the reconstruction of terrain features that are directly relevant for hydrologic applications.

Unlike other super-resolution approaches, which are often using downscaled versions of the low-resolution inputs to learn super-resolved DTMs, the proposed framework was trained on a growing set of aligned patches of real-world globally available low-resolution elevation data, optical satellite imagery, and high-resolution reference DTMs derived from airborne LiDAR. Model performance is evaluated against conventional interpolation and standard super-resolution baseline architectures, including convolutional neural networks (CNN) as well as geospatial foundation models (GFM). To assess the practical impact on flood mapping, the super-resolved DTMs are tested on a set of real-world flood events in Germany by using the well-known Flood Extent Enhancement and Water Depth Estimation Tool (FLEXTH) to derive inundation depth metrics.

Results show that integrating DTMs derived using hydrologically guided super-resolution into flood depth tools can lead to more accurate flood depth estimates compared to low-resolution or other super-resolved inputs. The added hydrologic loss significantly improves the preservation of slopes and flow directions while maintaining elevation accuracy.

Overall, the presented framework offers a method to generate hydrologically meaningful high-resolution DTMs from globally available low-resolution inputs to benefit flood depth estimation in areas, where no high-resolution terrain information is available.