The coseismic groundwater-level change in Odawara well which located in Kanagawa Prefacture, south Japan, has been recoded from 2011. The sampling rate of the groundwater level is 1Hz. In the long time observation, some ‘abnormal’ change has been checked by artificial. In these ‘abnormal’ groundwater level change, same can be related to earthquake waves as coseimic groundwarer level change. To pick out the coseismic groundwater level changes, We applied a new method based on Recurrent Neural Network(RNN) process to automatically pick out the coseismic groundwater-level change. In the RNN model, we applied a simple geomodel to relate the atmospheric pressure and groundwater level as a weight parameter in the neural network. We use a whole year of 1Hz sampling data to train the RNN model. As the result of the method, the accuracy of the all validation is 0.949, the accuracy of the seismic groundwater level events is 0.966 and the accuracy of the normal groundwater level is 0.971. This method show high accuracy to pick out the short-term secimic groundwater-level change which the change time is less than 45 minutes.