By incorporating the local statistics of an image, a causal nonstationary autoregressive random field models can be extended to a nonstationary image. This nonstationary random field can provide a better description of the image texture than the stationary one. As a consequence of the model above as well as another model for labeling field, an image is better segmented. Due to low-order dependence in image for texture random field above, there are still false alarms. Therefore, high-order dependence as a new classification feature to improve the result of object detection is introduced. Entropy rate, depicting the feature, is estimated using random field model. The technique proposed above is applied to extract urban areas from landsat image.