Data-driven Super-Resolution of Flood Inundation Maps using Synthetic Simulations

TL;DR

This work tackles the need for frequent, high-resolution flood inundation maps by downscaling daily low-resolution water fraction maps (WFM) to 30 m FIM using data-driven super-resolution models trained with synthetic, physics-based simulations. It evaluates three architectures—RCAN, RDN, and ESRT—within a penalized loss framework that aligns WFM fractions with high-resolution outputs, and demonstrates that synthetic data can effectively substitute scarce real-world data for training. The results show that DL-based downscaling generally outperforms traditional baselines and can exhibit zero-shot transfer to hydroclimatologically similar regions, though transfer effectiveness declines for climatically dissimilar areas. Overall, the method enables high-resolution flood monitoring with daily cadence, offering practical benefits for hydrological modeling, disaster response, and risk assessment, while highlighting the need for domain adaptation and integration of topographic features for global applicability. In formal terms, the problem is to learn a map from a low-resolution map $oldsymbol{X} \\in [0,1]^{L \times L}$ to a high-resolution binary map $oldsymbol{Y} \in \{0,1\}^{H \times H}$ with $H = fL$ and $f>1$, here $f=10$, leveraging synthetic data to train models that generalize to real-world observations.

Abstract

The frequency of extreme flood events is increasing throughout the world. Daily, high-resolution (30m) Flood Inundation Maps (FIM) observed from space play a key role in informing mitigation and preparedness efforts to counter these extreme events. However, the temporal frequency of publicly available high-resolution FIMs, e.g., from Landsat, is at the order of two weeks thus limiting the effective monitoring of flood inundation dynamics. Conversely, global, low-resolution (~300m) Water Fraction Maps (WFM) are publicly available from NOAA VIIRS daily. Motivated by the recent successes of deep learning methods for single image super-resolution, we explore the effectiveness and limitations of similar data-driven approaches to downscaling low-resolution WFMs to high-resolution FIMs. To overcome the scarcity of high-resolution FIMs, we train our models with high-quality synthetic data obtained through physics-based simulations. We evaluate our models on real-world data from flood events in the state of Iowa. The study indicates that data-driven approaches exhibit superior reconstruction accuracy over non-data-driven alternatives and that the use of synthetic data is a viable proxy for training purposes. Additionally, we show that our trained models can exhibit superior zero-shot performance when transferred to regions with hydroclimatological similarity to the U.S. Midwest.

Figures (5)

• Figure 1: Chosen regions for this study. Iowa is the region where the model was trained over, whereas all the other regions comprise of climatically similar (Europe) and dissimilar (Ghana, Red River) regions.
• Figure 2: Sample outputs for the RW Iowa Des Moines region; i.e., the region over which the model was trained.
• Figure 3: Sample outputs for the RW EU region; i.e., external to the regions over which the model was trained.
• Figure 4: Sample outputs for the RW Ghana region; i.e., external to the regions over which the model was trained.
• Figure 5: ROC curves for (a) RW-IA (CR) and (b) RW-IA (DM) dataset. Naïve model here represents a model whose output is solely a "no Flood" for all pixels. Star here represents the pixel-wise classifier with a threshold of 0.5.