Speaker
Description
Forecasting flash floods, landslides and debris flows in Alpine terrain is hard for a structural reason: the conventional warning chain runs from quantitative precipitation forecast (QPF) through hydrological model to alert, so its skill is capped by its least predictable link. Convective, orographically forced rainfall is routinely misplaced by numerical weather prediction (NWP). During the catastrophic August 2023 Slovenian floods, operational models captured the large-scale setup --- moisture transport, instability, orographic ascent --- yet failed to localise the extreme rainfall, limiting the spatial precision of catchment-scale warnings.
FLOODCAST, an early-stage project being developed with the Slovenian Environment Agency (ARSO), reframes the task as direct probabilistic classification of hazard occurrence from the atmospheric state. For catchment $c$, hazard type $h$ and lead time $\tau$, the model estimates
$$p\big(\,y^{h}_{c,\,t+\tau}=1 \;\big|\; \mathbf{s}_{c},\ \mathbf{a}_{c,t},\ \mathbf{X}_{\mathcal{G},\,t-L:t}\,\big),$$
where $\mathbf{s}{c}$ encodes static susceptibility, $\mathbf{a}{c,t}$ the antecedent land state (multi-layer soil moisture, snowmelt, antecedent precipitation), and $\mathbf{X}{\mathcal{G},\,t-L:t}$ the evolving NWP atmospheric state --- instability indices, integrated vapour transport, mid-tropospheric circulation, terrain-relative flow --- over a lookback window $L$ on a catchment graph $\mathcal{G}$. Forecast precipitation is downweighted rather than depended upon.
The system couples static and dynamic layers. The static layer provides hazard-specific spatial priors $\mathbf{s}{c}$: the first, a machine-learned Slovenia-wide flood susceptibility map built from terrain, hydrographic, land-cover and soil predictors, conditions the dynamic hazard probabilities. For the dynamic layer, $\mathcal{G}$ comprises 227 catchment nodes with contiguity, directed-routing and synoptic edges; we are developing a spatio-temporal graph neural network with a state-space (Mamba) sequence backbone, whose linear-time scaling suits long antecedent windows, alongside gradient-boosted-tree and convolutional U-Net baselines, pretrained on ERA5 reanalysis and fine-tuned on operational forecasts.
Evaluation follows a leave-one-event-out protocol on historical Slovenian events, with AUPRC, critical success index and reliability diagrams as headline metrics, and probability calibration under severe class imbalance as a central design constraint. We will present preliminary susceptibility-mapping results and progress on an end-to-end pilot targeting the August 2023 event, discussing the transferability of environment-based hazard learning across domains.