Fast and uncertainty-aware super-resolution of compound flooding using latent diffusion models

Upscaling coarse-grid flood maps to achieve fine-grid accuracy using machine learning has emerged as a promising hybrid pathway for operational flood mapping, as fine-grid hydrodynamic models remain computationally prohibitive for real-time and ensemble-based applications. In this context, latent diffusion models (LDMs), a class of generative AI models, have recently demonstrated superior accuracy and generalisability in flood map super-resolution. However, despite their inherently stochastic nature, it remains unclear to what extent LDM-generated ensembles provide meaningful representations of predictive uncertainty in flood depth estimates.

In this study, we develop a conditional latent diffusion framework to generate fine-grid high-resolution flood depth maps using coarse-grid flood simulations and digital elevation models (DEMs) as conditioning inputs. The approach is demonstrated for a coastal floodplain near Tacloban, Philippines, which is subject to complex compound flooding driven by inland rainfall and storm surge. Hydrodynamic simulations are performed using a subgrid-based shallow water solver. The coarse-grid model contains approximately 95 times fewer computational cells than the fine-grid model and executes around 188 times faster, albeit with reduced accuracy (pixel-wise RMSE of 81.2 cm for maximum flood depth map).

Fine-grid model outputs are treated as deterministic ground truth, allowing uncertainty arising solely from the stochastic behaviour of the LDM to be isolated. By repeatedly sampling the trained model for identical inputs (up to 100 stochastic runs), we systematically evaluate accuracy–uncertainty–cost trade-offs using RMSE and pixel-wise 90% confidence interval (CI) coverage of flood depths.

Results show that individual stochastic predictions substantially improve upon the coarse-grid baseline but exhibit notable variability, with RMSE ranging between approximately 19–24 cm (Figure 1). Ensemble averaging rapidly enhances accuracy, with ensemble-mean RMSE converging within 20–40 runs, yielding 26–14 times speed-up compared to fine-grid hydrodynamic simulations. However, despite increasing ensemble size, empirical 90% CI coverage stabilises at around 70%, indicating systematic under-capture of uncertainty. Increasing the number of reverse diffusion steps from 500 to 1000 does not significantly alter this behaviour (Figure 2) suggesting that uncertainty limitations are not driven by insufficient sampling resolution.

Further analysis indicates that uncertainty under-representation arises from overly strong conditional signals learned during training rather than ensemble size. Introducing controlled stochastic perturbations (Figure 2), into the latent representation of coarse-grid flood maps at inference time increases ensemble spread and substantially improves CI coverage, reaching approximately 86% for a noise factor of 0.2, while only marginally increasing RMSE (~0.3 cm).

The study highlights three key insights: (i) stochastic LDM ensembles provide a practical balance between accuracy and computational efficiency for operational flood mapping; (ii) increasing ensemble size alone yields diminishing returns for uncertainty representation under strong conditioning; and (iii) future research should focus on incorporating uncertainty-aware conditioning during training and leveraging advanced diffusion solvers to further reduce inference cost. Together, these findings establish a principled pathway toward fast, uncertainty-aware flood inundation modelling using generative AI.

Figure 1: Accuracy–uncertainty–cost trade-off.

Figure 2: Uncertainty coverage–accuracy trade-off.