Probabilistic event attribution of precipitation and streamflow extremes in the spring 2022 Manitoba flood using LSTM-based hydrological modeling
Year: 2026
Type: article
Abstract: The Spring 2022 flood was among the most extensive and damaging on record in southern Manitoba, yet the contribution of anthropogenic climate change to such cold-region events remains poorly quantified. This thesis quantifies the influence of anthropogenic climate change on the Spring 2022 precipitation and flood event using a probabilistic event attribution framework applied across both meteorological and hydrological domains. The observational records from 19 meteorological stations and 28 HYDAT streamflow gauges spanning the period 1950–2023 were considered, with the Whitemouth River sub-basin (gauge 05PH003) serving as the focal catchment. Long Short-Term Memory (LSTM) networks reconstructed six incomplete streamflow records. The meteorological event was the most extreme 29-day April–May precipitation accumulation, and the hydrological event the most extreme 27-day mean streamflow. Attribution used three approaches contrasting factual and counterfactual climates: General Circulation Model (GCM)-based, non-stationary scaling, and time-slice methods. Streamflow was simulated with LSTM networks and the Raven process-based model forced by climate model inputs. To explicitly account for the role of antecedent snow conditions in flood generation, a compound snow–temperature preconditioning analysis was conducted at Indian Bay using bivariate copulas, jointly characterising the influence of winter snowpack accumulation and early spring temperatures as compound precursors to flooding. Median risk ratios (RR) indicated that precipitation magnitude was likely exacerbated by anthropogenic forcing (1.00 to 1.53), with the strongest signal in southeastern Manitoba, underlain by the Canadian Shield and behaves distinctly from the lacustrine Red River Valley to the west. Flood attribution did not reach statistical significance: streamflow risk ratios leaned above unity but were not robust across either approach (LSTM: RR 1.45 [0.65, 2.66]; Raven: RR 3.14 [0.83, 8.48]), and the change in compound snow–temperature preconditioning was non-significant (RR 1.11 [0.13, 5.12]). These results reveal a partial decoupling between precipitation and flood attribution in cold-region Prairie basins, where flood magnitude is governed by snowpack, melt timing, antecedent wetness, and storage rather than precipitation intensity alone, so precipitation-based attribution cannot serve as a flood-risk proxy. The findings advance climate change attribution in snow-dominated basins and can inform flood risk assessment and adaptation planning across the Canadian Prairies.
Source: Mspace (University of Manitoba)
Author: Zain Ul Abideen, Muhammad
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
Funder: University of Manitoba
OpenAlex ID: W7204617038
