Operational real time hyperspectral reconnaissance systems adaptively estimate multivariate background statistics. Parameter values derived from these estimates feed autonomous onboard detection systems. However, inadequate adaptation occurs whenever an airborne sensor encounters a physical boundary between spectrally distinct regions. The transition area generates excessive false alarms, because standard detection algorithms rely on quasistationary models of background statistics. Here we describe a two-mode stochastic mixture model aimed at solving the boundary problem. It exploits deployed signal processing modules to solve a generalized eigenvalue problem, making a threshold test for targets computationally feasible.