Projection of future drought and its impact on simulated crop yield over South Asia using ensemble machine learning approach

Understanding the development mechanism of drought events, characterization of future drought metrics, and its impact on crop yield is crucial to ensure food security globally, and more importantly, in South Asia. Therefore, the present study assessed the changes in future projected drought metrics and evaluated the future risk of yield reduction under drought intensity. We characterized the magnitude, intensity, and duration of future drought by means of the SPEI drought index using CMIP6 (Coupled Model Inter-comparison Phase-6) climate models. The impact of future drought on crop yield was quantified from the ISI-MP (Inter-Sectoral Impact Model Inter-comparison Project) crop model by a proposed non-linear ensemble of Random Forest (RF) and Gradient Boosting Machine (GBM). Results suggested that high drought magnitude with a longer drought duration is projected in some regions of South Asia while high drought intensity comes with a shorter duration. It was also found that Afghanistan, Pakistan, and India will experience a longer drought duration in the future. Our proposed ensemble machine learning (EML) approach had high predictive skill with a minimum value of RMSE (0.358–0.390), MAE (0.222–0.299), and a maximum value of R2 (0.705–0.918) compared to the stand-alone methods of RF and GBM for yield loss risk projection. The drought-driven impact on crop yield demonstrates a high risk of yield loss under extreme drought events, which will encounter 54.15%, 29.30%, and 50.66% loss in the future for rice, wheat, and maize crops, respectively. Furthermore, drought and yield loss risk dynamics suggested a one unit decrease in SPEI value would lead to a 14.2%, 7.5%, and 10.9% decrease in yield for rice, wheat, and maize crops, respectively. This study will provide a notable direction for policy agencies to build resistance to crop production against the drought impact in the regions that are critical to climate change. [Display omitted] •Future changes of drought metrics evaluated from CMIP6-GCM models.•A non-linear ensemble machine learning approach is proposed for quantifying drought impact on crop yields.•The southwestern regions of Afghanistan and India, as well as the northeastern region of India, are expected to have the highest drought intensity in the future.•Rice is projected to have the highest risk of yield loss, followed by maize and wheat in South Asia.•Future yield loss risk is projected to non-linearly increase with an increase of drought intensi