Application of Machine Learning to Forecast Agricultural Drought Impacts for Large Scale Sub-Seasonal Drought Monitoring in Brazil

Abstract: Drought events have increased in frequency and severity over recent years and can result in significant economic losses, as well as impacts on both global and regional food security. Drought is a slow onset hazard taking place over months or years. This makes forecasting the propagation of drought from rainfall deficits to impacts upon soil moisture and vegetation health challenging. Drought impacts can depend on societal vulnerability making drought monitoring and forecasting an important task. In Brazil, half of all natural disaster events are drought related. Agricultural impacts can be significant, most severe impacts are historically in the semi-arid northeast. Drought is a significant challenge for farmers across Brazil. The previous La Niña was associated with severe drought in southern Brazil, this had significant impact upon soybean production, affecting food and milk prices as well as harming the country’s agricultural GDP.   

Recent years have seen significant advances in machine learning techniques and the availability of remote sensing data. These advances allow new insights into the propagation of drought and improvements in forecasts and early warning systems. Here we explore methods for forecasting vegetation health and soil moisture using machine learning techniques and the standardized precipitation-evapotranspiration index (SPEI). Models provide estimates of root zone soil moisture and vegetation health for sub-seasonal timescales relevant for agricultural adaptation. Models are trained and evaluated across agricultural regions of major crops, soybean and maize. The study area ranges across contrasting biomes in Brazil, including the semi-arid northeast, south, and major soybean growing region in the central Mato Grosso. This presents a challenge of building a forecasting system that can be accurate across a broad range of environments. The techniques developed as part of this study aim to inform operational drought forecasting at CEMADEN the national centre for monitoring and early warning of natural disasters in Brazil by providing forecasts of future drought in addition to current monitoring information. This will help to improve resilience against agricultural drought in Brazil. 
 

Biography: Joe Gallear is a Postdoctoral researcher at Rothamsted research. Joe's current work consists of using machine learning and satellite data to produce forecasts of drought impacts on vegetation health. This work is in collaboration with the UK Met office and the National Center for Monitoring and Early Warning of Natural Disasters in Brazil (CEMADEN) as part of the wider CSSP Brazil project. Joe completed his PhD at the university of Leeds on using machine learning and process-based crop modelling for regional scale  yield prediction.