In this study, we present an unsupervised change detection method using multi-spectral and multi-temporal remotely sensed imageries. This method is a pre-classification approach based on a spectral rule-based per-pixel classifier (SRC) developed by Baraldi et al. (2006). SRC is purely based on spectral-domain prior knowledge, such that no training or supervision process is needed. To explore its capability to detect change, we applied it in the Zhoushan Islands, Zhejiang, China. First, images were classified by SRC, and change detection was performed by two separate methods. One was the comparison of the merged categories obtained by reclassifying the pre-classification types of SRC. The other was comparing bi-temporal pre-classification types directly. The classification accuracy of the merged categories based on SRC was compared to the Maximum Likelihood Classifier and Support Vector Machine. The accuracy of the change detection was assessed and compared to results processed by the common post-classification comparison and change vector analysis methods. Results show that the change detection by directly comparing pre-classification types of SRC had the highest accuracy (overall accuracy was 90%, kappa coefficient was 0.81) among these methods and that the method of comparing merged categories was the worst (overall accuracy was 73%, kappa coefficient was only 0.46).