Serious damaged structures can result in catastrophic disasters during earthquakes. However, visual inspection of damage in structures by human is an inefficient and unreliable approach. Alternatively, a more scientific approach should be exploited to rapidly and accurately localize damage of structures. In this study, two damage detection methods based on prediction errors using a bank of Kalman estimators are presented and compared including a) a centralized approach and b) a decentralized approach. In the centralized approach, a representative model of a building is first derived from a frequency-domain system identification method under ambient vibration prior to earthquake events. This model is then converted into a bank of Kalman estimators, and the estimation errors can be calculated and then turned into statistics. The damage location, level, and time of occurrence can be statistically determined and presented by the damage indices. Similarly, in the decentralized approach, the same system identification method is first applied to structural responses. To be more realistic, the monitoring system is decentralized into subsystems with some overlapped sensor measurements. Banks of Kalman estimators can be constructed using the subsystems. By normalizing the damage probability indices from prediction errors of each bank, the damage location, level, and time of occurrence can be identified. A numerical example is given to demonstrate the two damage detection methods. Moreover, the two methods are compared by a scaled twintower building using shake table testing. The results indicate that both methods are quite effective for seismic damage detection.