MRF models are widely used in remote-sensing image segmentation to take into account dependencies among neighboring pixels. Compared to non-contextual techniques, MRF-based techniques provide much smoother segmentation maps, as they are able to counter the effects of sensor noise. Because of finite resolution of sensors, however, many boundary pixels are mixed (comprise two different land covers) and are incorrectly classified as belonging to a third class. Here we propose an adaptive tree-structured MRF model, which largely reduces such classification errors and increases map smoothness without sacrificing classification fidelity.