Forecasting and Quantifying Risks of Crop and Water Supply Failures Using Machine Learning and Remote Sensing

The increasing intensity and severity of droughts, coupled with the challenge of generating reliable forecasts, significantly contribute to food insecurity in developing countries. Enhanced agricultural drought forecasts offering early, impact-based warnings can assist communities in preparing for the detrimental effects of drought, thus diminishing the number of individuals impacted. Drought Early Warning Systems (DEWS) convert physical drought observations into actionable insights, guiding decisions about water-intensive agricultural practices and alternative irrigation water sources. These systems require extended lead times, ideally spanning months or seasons, due to the time-sensitive nature of decisions aimed at mitigating drought vulnerability. Prior research on impact-based forecasting primarily concentrated on precipitation indicators, frequently neglecting the integration of vegetation health, climatic, and hydrological factors, which are crucial in determining crop conditions. This study addresses the limitations of current DEWS by investigating the application of machine learning models, specifically Fast and Frugal Trees (FFTs) and Extreme Gradient Boosting (XGBoost), to identify agricultural impact triggers well before the harvest period. These algorithms forecast agricultural drought impacts by constructing decision trees from a comprehensive set of drought indicators. We illustrate the efficacy of machine learning models in predicting maize yield failures and anomalies in the Water Requirement Satisfaction Index (WRSI), a vital measure of crop performance during the growth season. This paper presents the outcomes of a pilot DEWS project, InfoSequia-4CAST, executed by the World Food Programme (WFP) in Mozambique in 2022. The user-end model successfully provided monthly DEWS bulletins for Tete and Gaza provinces with satisfactory reliability. End users had the flexibility to modify various parameters, including input variable selection, lead times of interest (up to six months), optimization statistical metrics, the geographic scale for model development, and the threshold method or value for alert initiation. Furthermore, we discuss prospects for enhancing the reliability of these impact-based forecasts by contrasting the applied and transparent FFT model with the more complex, 'black box' XGBoost model. While simpler approaches like FFTs may be more appealing to decision-makers due to their straightforwardness, their potential for oversimplification might limit their effectiveness, stemming from their restricted use of variables. XGBoost exhibits superior performance over FFTs in identifying drivers of agricultural drought and providing more reliable, data-driven forecasts. Nevertheless, FFTs hold value in contexts where transparency is required, as this can strengthen trust in the forecasts, which is equally crucial for their adoption and the initiation of responsive actions.