An improved hidden Markov prediction model (IL-HMMs) based on incremental learning was developed, which was based on the prediction of regional groundwater table in the typical county of Dengkou, a typical arid region in Northwest China. In order to test the IL-HMMs model prediction results, the predicted results was compared with the measured data of 2013, and the results of hidden Markov model (HMMs), weighted Markov chain (WMCP) and BP neural network (BPNN) prediction model. The results showed that compared with other forecasting models, the prediction accuracy of IL-HMMs model was improved significantly, the error was smaller and the robustness was better. The groundwater depth in 2018 was predicted by using the IL-HMMs model. The prediction results showed that in 2018 the average annual groundwater depth would be increased slightly and the groundwater depth would be increased in some areas. The IL-HMMs model of groundwater depth prediction had good stability and robustness, it can provide ideas and methods for the dynamic prediction of groundwater depth, and also provide an important basis for the development, utilization and protection of groundwater resources in the region. Tracking and monitoring the change of water level, preventing the groundwater level from falling continuously and making emergency response plan can be utilized to realize the sustainable development and utilization of water resources.