Bayesian analysis of variance for quantifying multi-factor effects on drought propagation

Drought poses a significant threat to agricultural sector, and meteorological drought (MD) is the main driver and origin of agricultural drought (AD). The process is susceptible to multiple potential factors, whereas there is a challenge in quantifying the effects of factors. This study develops a novel method named as BANOVA by combining Bayesian Model Averaging (BMA) with analysis of variance (ANOVA), and is tested in Central Asia which is an agricultural dominant area. Standardized Precipitation Index and Standardized Soil Moisture Index are calculated to characterize MD and AD from 1982 to 2014. Precipitation and soil moisture are provided by Climatic Research Unit Gridded Time Series and Global Land Data Assimilation System. The maximum Pearson correlation coefficients (MPCC) and drought propagation time (DPT) are selected to depict the process of drought propagation. The results suggest that BMA can effectively integrate the results of each machine learning model and reduce the uncertainty of model structure. Several findings can be summarized: (1) there is a stable relationship between AD and MD that MPCC are significant (p < 0.01) for all grids, and the average DPT is 2.8 months; (2) eight factors are selected for drought propagation simulation, and temperature plays the dominant role in both MPCC and DPT with the contribution of 86.2 % and 35.0 %; (3) the interactive effects of factors on drought propagation characteristics are significant, such as temperature and precipitation with the contribution of 13.2 % on drought propagation time. This study highlights the importance of temperature in drought propagation. Under the context of global warming, the propagation time and relationship between MD and AD will become shorter and closer, respectively.