We present a three steps approach that aimed at simulating neutral agricultural landscape models: (1) we characterized the geometry of three real field patterns; (2) we generated simulated field patterns with two tessellation methods attempting to control the value of some of the observed characteristics and, (3) we evaluated the simulated field patterns. The first two steps were integrated to the GenExP-LandSiTes software that thus simulates two-dimensional agricultural landscapes. It is written in Java, and it is freely accessible through a Gnu Public Licence. For the third step, we considered that good simulated field patterns should capture characteristics of real landscapes that are important for the targeted agro-ecological process. Real landscapes and landscapes simulated using either a Voronoi or a rectangular tessellation were thus compared when used as input data within The MAPOD-maize gene flow model. The results showed that the Voronoi tessellation performed better than the rectangular tessellation. In our ongoing research we consider random line-based tessellations constrained by a probability distribution penalizing the extreme values of targeted features (for example too large variability of cell areas). We also propose an algorithm for simulating such tessellations. The probability distribution parameters can be fitted from observed landscapes. This should result in generating tessellations similar to real patterns.