Abstract In this article two new approaches are presented for time domain identification of base isolated buildings from recorded response during earthquakes: (1) a least squares technique with time segments is developed to identify the piece-wise linear system properties; and (2) an observer is used to estimate the unmeasured states and initial conditions of different time segments. In base isolated buildings changes in dynamic properties occur during earthquake response due to nonlinear behavior. Hence, a multi-input and multi-output technique using time segments is developed for piece-wise linear system identification. The primary advantage of the developed time segmented technique is that it can be applied to windows of time history instead of the entire duration of earthquake response. The developed technique (1) starts with identification using the entire duration of the earthquake response; (2) evaluation of time segments during which the identified response differs significantly from the recorded response to establish windows of time history during which refined identification is necessary; and (3) identification of the change in dynamic properties in the established windows using the observer based time segmented least squares approach. Only partial state measurements are usually available for identification. Hence, an observer is used to estimate the unmeasured states and initial conditions needed for different time segments. By comparing identified response with recorded response, of an actual base isolated building which experienced Northridge earthquake, it is shown that the change in dynamic system parameters, such as periods and damping ratios, due to nonlinear response, are reliably estimated using the presented technique.