Long-term soil moisture prediction is of significance for agriculture, transpiration, and other activities. By coupling models of regional climate, surface hydrology, and crop development, we have established a two-way feedback agroecosystem model that provides soil moisture and other hydrological variable variations in four months in advance. One of the problems with this model, however, is its persistent over-drying soil that even prevents seeds from germination in typical years. To solve this problem, we first ran a series sensitivity experiments diagnosing the error sources, determining whether the error comes from water input to soil (i.e., rainfall), soil property (e.g., bulk density, percentage triangle) or soil hydrology formulations. Of particular concern is the soil water retention curve used in the model. It is found that the soil water retention relation typically used in the atmospheric models tends to retain less water at a give water pressure than that typically used in soil science community. We will report the model improvement after adopting a new retention relation in the model by validating the model against in-situ and remote sensing soil data.