Deep learning sub-seasonal predictions of flooded area in South Sudan

South Sudan has recently been experiencing severe, persistent flooding, particularly in the country's central wetland region, known as the SUDD. Flooding in the SUDD is driven by rainfall that spans outside the country's borders, including the African Great Lakes. A predictive model of the flooding dynamics can help support preemptive adaptation and release financing before the onset of an event by identifying areas at risk of severe flooding in the coming months.

We use a Long Short-Term Memory Network (LSTM) to make predictions of fractional flooded areas (FFA) at the administrative level 2 in South Sudan. We use diurnal fluctuations of MODIS Land Surface Temperature over a 10 day period to estimate FFA (Boeck et al., 2022). The model takes three types of inputs: 1) Daily precipitation from Global Precipitation Measurement aggregated over four watersheds, which encompass the upper White Nile, Bahr al-Arab and the Baro Akobo Sobat, 2) Monthly gravitational anomaly from GRACE/GRACE-FO from the African Great Lakes and the SUDD wetland and 3) Static characteristics of the administrative unit on which flooding predictions are made. We trained a single LSTM simultaneously on all admin units, creating a single national model. The output of the model is a time series of FFA at each administrative unit. A cross validation of the model performance shows a range of Nash-Sutcliffe Efficiency (NSE) values from -0.5 to 0.8, with a median NSE value of 0.4.

Citations:

Boeck, S., Bonifacio, R., Pozzi, L., and Bockowska, P.: Using MODIS thermal data for mapping and monitoring of a massive multi-year flooding event in South Sudan for humanitarian response and decision making , EGU General Assembly 2022, Vienna, Austria, 23-27 May 2022, EGU22-11964, https://doi.org/10.5194/egusphere-egu22-11964, 2022.