We present a new model for extracting tree crowns from aerial images, containing a near infrared channel. We addres s the problem as an image segmentation problem solved by a Markov Random Field (MRF) modeling embedded in a Bayesian framework. The prior model consists of a couple field defined by a three states MRF, associated to vegetation, background and center of trees, and a template field representing the tree shape. A template field allows to associate an ellipse to model the crown of each detected tree. The mod el is optimized by a simulated annealing scheme based on a Metropolis dynamics. We show some results on real images of manmade plantations and natural forests.