Foundation models have revolutionized artificial intelligence by performing generative tasks and enhancing task-specific capabilities that traditionally required substantial volumes of labeled data. One area that has been significantly hindered by supervised training methods is remote sensing-based change detection. While standard pixelwise change detection methods provide binary categorization, these techniques frequently exhibit sensitivity to innocuous or routine changes such as seasonal variations, shadows, and building parallax. Deep learning approaches have previously shown promise for improving change detection; however, many complexities persist. Change exists on a continuous spectrum, varying significantly in importance and scope on a site-by-site or task-by-task case. Subtle, semantic changes (e.g., hanger doors opening or closing) can often be more critical than more evident physical alterations. Manual annotation for these continuous changes remains prohibitively difficult, time-consuming, and subject to inconsistency due to labeling variability among annotators. Foundation models, which leverage self-supervised techniques to learn rich, generalizable feature spaces from diverse and unlabeled remote sensing imagery across various locations, perspectives, and resolutions, offer a promising solution. Utilizing the outputs of these rigorously trained encoders yields robust unsupervised change detection capabilities. In this work, we demonstrate how these foundation model embeddings can be effectively tuned to differentiate semantic from physical changes and categorize the objects or contexts contributing to them.