In view of the problem of the approximation models with poor self-adaptability by uniform error threshold, a method is proposed in this paper. In this method, the terrain is divided into some different areas with various terrain features, and then the error function and the error threshold are adaptively determined according to these areas. Aimed at the character of big data in terrain model, a detailed hierarchy is constructed and its fast index method is described. In order to control the density of feature points in the primary step, the feature points are selected based on the convex terrain and its diffused points. It solves the problem of the feature selection for the gentle slope terrain. The method of searching for multi-resolution adjacent nodes is put forward to accelerate the area division process with coarse grain, and the matching function between the adjacent nodes is also described. Besides, a relief degree function for terrain surface is proposed to evaluate the terrain areas, so as to determine the division strategy of the approximation model. As a result, the terrain model is of higher adaptability and precision. The experiments on some real data show that the method is superior to the uniform error threshold method.