Droughts and floods alternately occur over a large karst plateau (Yun–Gui Plateau) in Southwest China. Here we show that both the frequency and severity of droughts and floods over the plateau are intensified during the recent decade from three-decade total water storage anomalies (TWSA) generated by Gravity Recovery and Climate Experiment (GRACE) satellite data and artificial neural network (ANN) models. The developed ANN models performed well in hindcasting TWSA for the plateau and its three sub-regions (i.e., the upper Mekong River, Pearl River, and Wujiang River basins), showing coefficients of determination (R2) of 0.91, 0.83, 0.76, and 0.57, respectively. The intensified climate extremes are indicative of large changes in the hydrological cycle and brought great challenges in water resource management there. The TWSA of the plateau remained fairly stable during the 1980s, and featured an increasing trend at a rate of 5.9 ± 0.5 mm/a in the 1990s interspersed extreme flooding in 1991 and during the second half of the 1990s. Since 2000, the TWSA fluctuated drastically, featuring severe spring droughts from 2003 to 2006, the most extreme spring drought on record in 2010, and severe flooding in 2008. The TWSA of the upper Mekong has decreased by ~ 100 mm (~ 15 km3) compared with that at the end of the 1990s. In addition to hindcasting TWSA, the developed approach could be effective in generating future TWSA and potentially bridge the gap between the current GRACE satellites and the GRACE Follow-On Mission expected to launch in 2017.