Spatiotemporal characteristics and forecasting of short-term meteorological drought in China

Abstract

Intense short-term meteorological drought may lead to rapid declines in soil moisture, triggering flash drought that can cause major agricultural or socioeconomic damage. Machine learning methods have proven effective in forecasting hydrometeorology, but short-term drought forecasting is still inadequate. We developed a pentad-time-scale (5-day) standardized precipitation evapotranspiration index (SPEI-5d) to quantify short-term meteorological drought and analyzed its spatiotemporal distribution characteristics in China during the period 1962–2018. We used historical SPEI-5d as input and employed five machine learning methods for drought hindcasting: an autoregressive integrated moving average (ARIMA), random forest (RF), recurrent neural network (RNN), long short-term memory (LSTM), and convolutional long short-term memory (ConvLSTM). The results show the following: (1) During the period 1962–2018, 61.9% of the study region showed decreasing trends in drought severity, while 69.6% of the region showed increasing trends in drought intensity. (2) Drought duration, severity, and intensity have distinct seasonal characteristics, and different forecasting models perform differently in each season, with generally lower forecasting accuracy in summer and higher forecasting accuracy in winter. (3) The ConvLSTM model can capture spatiotemporal information well compared to traditional time-series forecasting models; it has the best performance (root mean square error = 0.29, and Nash-Sutcliffe efficiency = 0.92) in the test set and has high forecasting accuracy (R2 > 0.8) for lead times of 1–5 days (with accuracy decreasing as lead time increases). Our findings highlight the spatiotemporal variability of short-term meteorological drought and provide valuable scientific insights for short-term meteorological drought forecasting at 1–5 days of lead time.