A large amount of continuous input data is used to estimate groundwater level (GWL) by using machine learning models. However, data collection is very difficult and costly in undeveloped countries. Therefore, obtaining a general model and using less input data is the key to popularizing the application of machine learning models for estimating groundwater levels. This study evaluated the potential of the kernel-based nonlinear extension of the Arps decline model (KNEA), long short-term memory network (LSTM) and gated recurrent unit (GRU) for accurately estimating GWL in the Hetao Irrigation District in China. All models were developed using monthly records from 143 monitoring wells between 1990 and 2015. Eight input combinations (including the one-month prior GWL, air temperature, global solar radiation, precipitation and amount of irrigation) were applied to explore the possibility of improving model accuracy using less input data. In addition, the general performance of the models was evaluated by cross validation. The results showed that the KNEA model was superior to the LSTM and GRU models for all input combinations using the local application. For cross-district application, the average statistical results indicated that the LSTM (RMSE = 0.45 m and R² = 0.78) and GRU (RMSE = 0.48 m and R² = 0.76) models performed better than the KNEA model (RMSE = 0.70 m and R² = 0.62), and the LSTM model achieved the highest accuracy and stability. For input data, these three models had difficulty obtaining satisfactory monthly GWLs using meteorological and irrigation data without pervious GWL data. Adding meteorological data on the basis of the historical GWL greatly improved the accuracy of the models. Compared with PREC and GSR, adding temperature input had the best improvement. However, adding large-scale average irrigation data did not significantly improve the accuracy of the models. In addition, the LSTM model and input data of the historical GWLs and temperature were recommended in arid and semiarid agricultural areas with limited data.