Prediction of spring agricultural drought using machine learning algorithms in the southern Songnen Plain, China

Abstract

Winter climate conditions have a great effect on spring soil moisture (SM) in cold regions. The frequent occurrence of spring drought events in the Songnen Plain poses a significant threat to food security. In this study, we selected the random forest (RF) algorithm from three currently popular machine learning algorithms to predict the spatial pattern of average SM for the period of April 1–15 and April 16–30 with 1 km resolution in the dry cropland of southern Songnen Plain (SSNP) using 0–10 cm SM data for each April from 81 agro‐meteorological stations and winter climate data (e.g., temperature, snowfall), soil properties and topographic relief. Compared to the existing ERA5_Land and SMAP SM products, we obtain higher precision SM data (April 1–15: r = 0.74, RMSE = 0.050 m3 m−3; April 16–30: r = 0.73, RMSE = 0.051 m3 m−3). Drought grades were classified based on predicted SM data, and the results indicate agricultural drought was mainly influenced by the change of winter snowfall/snow depth, with western SSNP being more susceptible to drought because of soil properties. Compared with current SM data products, the RF model proposed in this study can implement more accurate prediction of spring soil drought based on winter climate, provide important information for agricultural management departments to prepare for spring cropping and irrigation, and avoid further soil salinization caused by drought.