Generalized change detection is a critical capability to mitigate the need for massive human inspection of the rapidly expanding volume ofglobal overhead satellite imagery. Current optical change detection approaches focus on fully specified systems to detect a predefined set ofchanges, and effective approaches for generalized change detection have not yet been demonstrated. We propose to build a deep learningbasedchange detection system that operates at the level of "intermediate semantics" for satellite images. Intermediate semantics are animage representation above the level of pixels, which vary considerably from day to day, but below the semantics of objects. Our system willlearn intermediate image semantics by discovering latent representations that can recreate the consistent aspects of images, even in thepresence of superficial changes such as changes in lighting. Our system will also have a trained discriminator to determine which changes in theintermediate semantics suggest meaningful change.Our approach is self-supervised: only time and place are used to provide training signals, and we will not use human-labeled training data.This approach will allow us to find interesting and significant changes over time, cueing regions of interest for analysts without needing to trainfor specific changes to detect.