Abstract：Drought is a complicated and costly natural hazard and identification of critical drought factors is critical for modeling and forecasting of droughts and hence development of drought mitigation measures (the Standardized Precipitation-Evapotranspiration Index) in both space and time. Here we quantified relationships between drought and 23 drought factors using remote sensing data during the period of 2002–2016. Based on the Gradient Boosting Algorithm (GBM), we found that precipitation and soil moisture had relatively large contributions to droughts. During the growing season, the relative importance of Normalized Difference Water Index (NDWI-7) for SPEI3, SPEI6, SPEI9, and SPEI12 reached as high as 50%. However, during the non-growing season, the Snow Cover Fraction (SCF) had larger fractional relative importance for short-term droughts in the Inner Mongolia and the Loess Plateau which can reach as high as 10%. We also compared Extremely Randomized Trees (ERT), H2O-based Deep Learning (Model developed by H2O.deep learning in R H2O.DL), and Extreme Learning Machine (ELM) for drought prediction at various time scales, and found that the ERT model had the highest prediction performance with R2 > 0.72. Based on the Meta-Gaussian model, we quantified the probability of maize yield reduction in the North China Plain under different compound dry-hot conditions. Due to extreme drought and hot conditions, Shandong Province in North China had the highest probability of >80% of the maize yield reduction; due to the extreme hot conditions, Jiangsu Province in East China had the largest probability of >86% of the maize yield reduction.