Spatiotemporal changes of drought area as input for a machine-learning approach for crop yield prediction

Abstract

More severe and prolonged droughts observed in recent decades require improved methods to predict impacts on agriculture. Crop-growth models estimate yield and plant development variables and are widely used to assess drought impacts; however, they are not explicit forecasting tools, as their accuracy is constrained by physical assumptions, data availability, and multiple sources of uncertainty. To address these limitations, machine learning (ML) models have been increasingly applied for crop yield prediction, typically using drought indices as input, while spatial drought characteristics remain underexplored. This research develops an ML framework that incorporates the spatial extent of drought to predict seasonal crop yield. The framework combines artificial neural network (ANN) and polynomial regression (PR) models, with PR providing baseline estimates and ANN delivering refined predictions. The approach was tested using 50 years of historical crop yield data and drought areas derived from the Standardised Precipitation Evapotranspiration Index at multiple aggregation periods (1–12 months). Results show ANN models consistently outperform PR models, achieving lower prediction errors, with root mean square error values as low as 48.1 kg/ha in best-performing cases. The results demonstrate that spatiotemporal drought area dynamics and their temporal aggregation provide an effective preprocessing strategy for ML-based drought impact prediction.