Rainfall nowcasting with machine-learning for landslide early warning

Real-time data plays a crucial role in predicting short-duration rainfall. These predictions have a significant impact on our daily life, especially in emergencies that can imperil human safety and the environment. A tragic example of this occurred on November 26, 2022, in Casamicciola, island of Ischia, in the Gulf of Naples (Italy), when heavy rainfall triggered landslides, resulting in loss of lives and extensive damage to buildings and infrastructure. Such destructive consequences highlight the urgent need for accurate short-term rainfall forecasts.

In fact, it is very important to have reliable short-term rainfall forecasting for early warning purposes. This can help save lives and property by preventing the effects of flooding, landslides, and other hazards caused by heavy rainfall . It is hard to understand and model how rainfall changes over time, and therefore, predicting rainfall in the short term is a complex challenge. Many of the current models use complicated statistical methods that are often too expensive and time consuming. In contrast, machine-learning (ML) models can find hidden patterns in rainfall data and predict the hourly or sub hourly amount of rain with limited computational burden.

In this study, a novel ML model is developed for nowcasting rainfall, to explore how it can make effective and quick short-term forecasts of precipitation. Specifically, the random forest (RF) algorithm is used, as recent studies have found this ML approach to be suitable (Mdegela et al., 2023). The study area is located on the island of Ischia, where 4 rain gauges (Ischia, Monte Epomeo, Forio, Piano Liguori) with a temporal resolution of 10 minutes are installed. The proposed model uses both the precipitation time series of the rain gauges and surface rainfall intensity provided by radar managed by the Civil Protection agency with temporal resolution of 5 minutes and a spatial resolution of 1 km, to predict the future precipitation. Different grid sizes (20x20, 30x30, 40x40 and 50x50 km) centred on the island of Ischia are considered to select the best radar input data (features) for the RF algorithm. The datasets were randomly selected for RF model training (70% of the data) and validation (30% of the data). The Minimum Inter-arrival Time (MIT) criterion was adapted for the definition of rainfall events within the rain gauge precipitation records (Heneker et al., 2001). A rainfall event is defined as a rainfall period preceded and followed by dry periods longer than MIT.

The results indicate that the RF model provides reliable short-term precipitation forecasts using only observed values as input, making it a fast, simple, and convenient method for nowcasting. The resulting precipitation forecast has the potential to be used in an early warning system to mitigate the impact of landslides.

 

References

Heneker, T.M., Lambert, M.F., Kuczera, G., 2001. A point rainfall model for risk-based design. J. Hydrol. 247, 54–71.

Mdegela, L., Municio, E., De Bock, Y., Luhanga, E., Leo, J. and Mannens, E., 2023. Extreme Rainfall Event Classification Using Machine Learning for Kikuletwa River Floods. Water, 15(6), p.1021.