From meteorological to agricultural drought: Propagation time and influencing factors over diverse underlying surfaces based on CNN-LSTM model

Abstract

As global warming intensifies and extreme weather events become more frequent, the severity of drought conditions in China's Xinjiang region has escalated. This exacerbates socio-economic pressures in the area and presents increasingly formidable challenges for the future. In response to these challenges, researching drought phenomena in Xinjiang is imperative. This study employs Bayesian methods and copula functions to estimate drought propagation time. It utilizes a hybrid deep learning model (CNN-LSTM) to analyze the process of drought propagation and its influencing factors across four land cover types: crops, forest land, grassland, and unused land. The findings indicate that Cropland experiences the longest average time of drought propagation (5.27 months), while forests have the shortest duration (4.2 months). Unused land and grassland exhibit similar average durations of propagation (4.8 months). On an annual scale, drought propagation time for each land type manifests in two phases: from January to May and from June to December. The former phase shows propagation time ranging from 6 to 9 months, while the latter ranges from 1 to 5 months; both demonstrate an increasing trend over time. Seasonally, all Land Cover Types exhibit a pattern of shorter propagation times in summer and autumn compared with winter and spring. Moreover, a longer time of drought propagation correlates with a greater disparity between meteorological and resultant agricultural drought severity. In analyzing the influence of factors on drought propagation, soil moisture content and El Niño-Southern Oscillation(ENSO) were found to significantly impact all Land Cover Types, progressively strengthening their influence over the years.