Change detection between images is a procedure used in many applications of remote sensing data. Among these applications, the identification of damaged infrastructures in urban areas due to a large-scale disaster is a task that is crucial for distributing relief, quantifying losses, and rescue purposes. A crucial consideration for change detection is that the images must be co-registered precisely to avoid errors resulting from misalignments. An essential consideration is that some large-magnitude earthquakes produce very complex distortions of the ground surface; therefore, a pair of images recorded before and after a particular earthquake cannot be co-registered accurately. In this study, we intend to identify changes between images that are not co-registered. The proposed procedure is based on the use of phase correlation, which shows different patterns in changed and non-changed areas. A careful study of the properties of phase correlation suggests that it is robust against misalignments between images. However, previous studies showed that, in areas with no-changes, the signal power in the phase correlation is not concentrated in a single component, but rather in several components. Thus, we study the performance of the ℓ1-regularized logistic regression classifier to identify the relevant components of phase correlation and learn to detect non-changed and changes areas. An empirical evaluation consisting of identifying the changes between pre-event and post-event images corresponding to the 2018 Sulawesi Indonesia earthquake-tsunami was performed for this purpose. Pairs of visible and near-infrared (VNIR) spectral bands of medium-resolution were used to compute the phase correlation to set feature space. The phase correlation-based feature space consisted of 484 features. We evaluate the proposed procedure using a damage inventory performed from visual inspection of optical images of 0.5-m resolution. A third-party provided the referred inventory. Because of the limitation of medium-resolution imagery, the different damage levels in the damage inventory were merged into a binary class: “changed” and “non-changed”. The results demonstrate that the proposed procedure efficiently reproduced 85 ± 6% of the damage inventory. Furthermore, our results identified tsunami-affected areas that were not previously identified by visual inspection.