This paper studies the detection of spectral targets corrupted by a colored Gaussian background from noisy, incoherent projection measurements. Unlike many detection methods designed for incoherent projections, the proposed approach a) is computationally efficient, b) allows for spectral backgrounds behind potential targets, and c) yields theoretical guarantees on detector performance. In particular, the theoretical performance bounds highlight fundamental tradeoffs among the number of measurements collected, the spectral resolution of targets, the amount of background signal present, signal-tonoise ratio, and the similarity between potential targets in a dictionary.