EGU General Assembly 2022
© Author(s) 2022. This work is distributed under
the Creative Commons Attribution 4.0 License.

Forecasting a proxy of humanitarian drought impact with machine learning using meteorological predictors; a case study for Zimbabwe

Marijke Panis1, Phuoc Phùng2, Bouke Pieter Ottow2, and Aklilu Teklesadik2
Marijke Panis et al.
  • 1The Netherlands Red Cross, 510, Netherlands (
  • 2The Netherlands Red Cross, 510, Netherlands

The impacts of drought are complex due to the multidimensionality (intensity, duration, and extent) and slow-onset nature of droughts. To be able to forecast the impact of droughts, one needs to prioritize and disentangle the diversity of impacts. In Zimbabwe, our country of interest, the Zimbabwe Red Cross Society prioritized crop loss, livestock loss, child malnutrition, and stunting. However, no high-quality data with national spatial coverage on these impacts is available. Therefore, it is necessary to use a proxy indicator for these impacts (or one of these impacts). As Zimbabwe is strongly dependent on rainfed- agriculture for its livelihood, our assumption is that a crop yield anomaly can be used as a proxy for crop loss impact. A negative crop yield anomaly derived from global historical yield series was used to determine the drought status (yes or no impact) in April and forms the target or predictand. The meteorological indicators to predict the crop yield are the observed 3-month-averaged El Niño–Southern Oscillation (ENSO) and the observed monthly rainfall from CHIRPS for each lead time. Also, a combination of monthly rainfall and ENSO was used as predictor. Our forecasting ML classification model, XGBoost, is run at lead times of one to seven months and at the livelihood zone/agro-climatic zone level. The entire dataset for 1983-2015 is divided into train (80%), test, and validation sets. Statistical performance is measured with the Probability of Detection and False Alarm Ratio of both the test and validation set. Our findings show the potential of ENSO-based data in forecasting our proxy for drought impact over various lead times. The addition of rainfall does not improve forecast skill. Future research will investigate if additional meteorological- and biophysical predictors such as soil moisture and Vegetation Condition Index improve the forecast skill. Our IBF Trigger Model for drought is currently a sequence of automated tasks that feed into an IBF-Portal with comprehensive visualizations for decision-makers. Both the development of the trigger model and the portal result from close collaborations and co-designs with the Zimbabwe Red Cross Society and its in-country partners.

How to cite: Panis, M., Phùng, P., Ottow, B. P., and Teklesadik, A.: Forecasting a proxy of humanitarian drought impact with machine learning using meteorological predictors; a case study for Zimbabwe, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11590,, 2022.