Machine-learning enhanced forecast of tropical cyclone rainfall for anticipatory humanitarian action
- 1Politecnico di Milano, Department of Electronics, Information, and Bioengineering, Milano, Italy
- 2Fondazione Centro Euro-Mediterraneo sui Cambiamenti Climatici, Bologna, Italy
- 3European Centre for Medium-Range Weather Forecasts, ECMWF, Reading, UK
- 4Department of Meteorology, University of Reading, Reading RG6 6BB, UK
- 5Red Cross Red Crescent Climate Centre, The Hague, 2521 CV, the Netherlands
Tropical Cyclones (TCs) have the potential to cause extreme rainfall and storm surge, which in turn can lead to riverine and coastal flooding with huge damage to property and loss of lives.
The use of precipitation forecasts in the context of decision-making and anticipatory action is currently hampered by the limited skill of numerical weather prediction models in forecasting the characteristics of such extreme rainfall events (especially their severity and location) with a sufficiently long lead time.
In this study, we present a post-processing scheme for precipitation forecasts based on a popular deep-learning algorithm (U-Net). We design our Machine Learning (ML) model to reduce the local biases of precipitation forecasts from TCs and adjust the spatial distribution of extreme rainfall. For this, we use a composite loss function to train the model, based on the combination of the Mean Absolute Error (MAE) and the Fractions Skill Score (FSS). We first demonstrate the potential of our ML-based approach working on ERA5 reanalysis data and subsequently apply it to the ensemble mean of ECMWF sub-seasonal forecasts with a lead time up to 10-days. As for the ensemble spread, we investigate possible post-processing adjustments based on the improvement of the spread-error relationship and of action-relevant scores of interest for humanitarian agencies, namely False Alarm Ratios (FAR) and Hit Rates (HR). We train and validate the model on a historical dataset of global TC precipitation events, using ECMWF re-forecasts over 20 years and a multi-source observational dataset (MSWEP) as reference. The results are evaluated with a multi-criteria approach including MAE, FSS, FAR, and HR, to assess the capacity of improving the predicted severity and spatial patterns of TC precipitation, as well as their potential for triggering anticipatory actions. Finally, we discuss how the outputs of our model can be used and further improved to support humanitarian actions aimed at saving lives in vulnerable communities in Mozambique.
How to cite: Ficchì, A., Ascenso, G., Giuliani, M., Scoccimarro, E., Magnusson, L., Emerton, R., Stephens, E., and Castelletti, A.: Machine-learning enhanced forecast of tropical cyclone rainfall for anticipatory humanitarian action , EGU General Assembly 2023, Vienna, Austria, 24–28 Apr 2023, EGU23-15188, https://doi.org/10.5194/egusphere-egu23-15188, 2023.
