Random Forest approach to forecast onset date and duration of rainy season in Tanzania

In the SADC region of Eastern Africa the onset and duration of the rainy season is of high importance to the agriculture and general water resource management. The planting time, selection of crops and success of different crops is linked to how skillfully this date can be forecasted.  
 
As part of the Horizon 2020 project called FOCUS-Africa, in order to forecast this specific onset-date and duration for a specific location in Tanzania, we have constructed a statistical model utilizing the Random Forest algorithm. This is being trained using a mix of observation of past teleconnection indices such as IOD and ENSO3.4 from recent months that from earlier studies have shown to be connected to the onset date and dynamical seasonal forecast of precipitation with a daily temporal resolution. At this stage three dynamical models are included. Finally, the observed precipitation of the previous months is being used as predictors as well.  
 
The first results have shown an improvement of the statistical model over using climatic information such as mean onset date as the reference forecast. This can be achieved 2-3 months ahead of the onset date. Furthermore, a relatively large importance of the seasonal forecast systems and the teleconnection indices seems to be present several months ahead of the observed onset date. This also indicates the importance of mixing observations and dynamical models in order to optimize the best possible overall skill of the system in predicting the onset date of the rainy season and thereby supporting local agriculture.