A data-driven framework for quantifying flash drought impacts on winter wheat production

Abstract

Flash droughts are increasingly recognized as a major threat to agricultural productivity due to their rapid onset and intensification of drought conditions. Recent evidences from winter wheat systems indicate that yield responses depend strongly on the timing of drought development, highlighting the importance of crop phenology in mediating drought impacts on yield. While newer studies consider crop phenology, there is lack of evidence comparing flash droughts with other extreme weather events that have impacts on yield. Therefore, a quantitative framework that integrates flash drought, crop phenology, and other weather drivers to explain yield variability lacks development, limiting our ability to infer on winter wheat productivity. Here, we investigated the effects of flash droughts and other climate factors during different phenological phases of winter wheat yield and abandonment. We first estimated crop phenological dates using generalized additive models (GAMs) with Gaussian process smoothing for nonlinear spatial and temporal variation. Using the NDLAS-3 dataset at 1km spatial and hourly temporal resolution, we derived several weather indices, including crop evapotranspiration, a key indictor of flash drought development. We then quantified the effects of flash droughts and other weather drivers on county-level winter wheat yield and abandonment across the Great Plains from 2002-2020 using both statistical and machine learning methods. Our phenological date estimation framework performed well in prediction, both in interpolation and extrapolation. Mean water-use efficiency during flash droughts were among the strongest predictors of both yield and abandonment, ranking above most other weather variables in variable importance via permutation. Incorporating flash drought metrics improved overall model performance for both linear and random forest models. In particular, mean water-use efficiency during the earlier stages of the winter wheat growing season accounted for almost 80% of the explained variation in crop abandonment with early freezing, spatial, and temporal features, outranking late-season frost and extreme heat. These findings suggest that current winter wheat production systems in the Great Plains remain insufficiently adapted to flash drought risk and show to have non-linear effects on yield and abandonment.