In this paper, a model based method to recognize agricultural fields is presented and demonstrated. At first, the task of the recognition is formulated as the problem of cost minimization. The approach is implemented through an hypothesis-correction-improvement process which usually starts with an initial hypothesis of object feature, compares it with the measured one, and then a cost can be calculated to control the improvement process. In this method, firstly, the cost is a function of the object shape parameters, and its value indicates the difference between the predicted and measured feature of an object. Secondly, to drive the cost to its minimum, the method used geometrical and topological properties of objects to constrain the optimization procedure. This prior knowledge helps the method reach the global minimum instead of a local one. The object recognition experiments performed on the high noise images (SAR) and the comparison results between different search strategies are given.