Exploring dynamic response of agrometeorological droughts towards winter wheat yield loss risk using machine learning approach at a regional scale in Pakistan

Context: Crop yield is a major agriculture sector affected by climate change; especially agrometeorological droughts experienced by south Asian countries in past decades.Research objective: The main goals of this research were to explore the spatiotemporal characteristics of seven agrometeorological drought indices at a regional scale in Pakistan. Secondly, to forecast the wheat yield loss risk (YLR) due to droughts under current and future climate scenarios by employing three machine learning (ML) methods; random forest (RF), gradient boosting machine (GBM), and generalized additive model (GAM).Method: The relationship between detrended wheat yield and a combination of five remote sensing indices Normalized Difference Water Index (NDWI), Vegetation Condition Index (VCI), Temperature Condition Index (TCI), Vegetation Health Index (VHI), and Drought Severity Index (DSI)), and two meteorological drought indices (Palmer's Drought Severity Index (PDSI), and Standardized Precipitation Evapotranspiration Index (SPEI)) was analyzed. Mann-Kendall trend (MK), Sens's slope, and Sequential Mann-Kendall (SQMK) tests are applied to explore the trend and trend-changing years for all indices over the historical time of 20 years. The YLR and all indices were projected (2021-2050) from the baseline period (2001-2020) using PROPHET time series forecasting and CMIP6 climatic models. YLR was forecasted on present and future projected time series by employing three non-linear ML regression models. Results: The output of the drought analysis revealed that the study area was hit by high to severe drought events in 2001-2004, 2006, 2008, 2010, 2012, and 2017. Trend analysis revealed intersection years breaking the rising trend of drought indices. All drought indices are significantly correlated with meteorological wheat yield with a sequence of NDWI>DSI>VCI>VHI>PDSI>SPEI>TCI. Future projections under high emission scenarios revealed a rise in YLR associated with frequent projected droughts from VHI, DSI, SPEI, and PDSI. YLR forecasting from agrometeorological indices is best predicted by random forest with the lowest RMSE = 0.005314. NDWI (26%) and VCI (19%) are found to be significant relative predictors associated with 51% high YLR in the baseline period and SPEI (20%) and NDWI (17%) as the most important relative predictors associated with 39% high YLR in future.Conclusion: The region is vulnerable to agrometeorological droughts with more susceptibility to less rain and high temperature affecting crop health and a high risk of yield loss in the future.Implication: The study provides a direction to stakeholders and policymakers to develop and adapt better strategies to mitigate and prevent drought-related yield loss risk in the future.