We consider the problem of segmentation in noisy, blurred astronomical hyperspectral images (HSI). Recent methods based on an hypothesis-testing framework handle the problem, but do not allow to use a prior on the result. This paper introduces a pairwise Markov field model, allowing the unsupervized Bayesian segmentation of faint sources in astronomical HSI. Results on synthetic images show that the segmentation methods outperform their state-of-the-art counterparts, and allow the detection at very low SNR. Besides, results on real images provide encouraging detections with respect to the application.