Twitter-based public health surveillance systems have achieved many successes. Underlying this success, much useful information has been associated with tweets such as temporal and spatial information. For fine-grained investigation of disease propagation, this information is attributed a more important role. Unlike temporal information that is always available, spatial information is less available because of privacy concerns. To extend the availability of spatial information, many geographic identification systems have been developed. However, almost no origin of the user location can be identified, even if a human reads the tweet contents. This study estimates the geographic origin of tweets with reliability using a density estimation approach. Our method reveals how the model interprets the origin of user location according to the spread of estimated density.