Currently environmental distribution maps, such as for soil fertility, rainfall and foliage, are widely used in the natural resource management and policy making. One typical example is to predict the grazing capacity in particular geographical regions. This paper uses a discovering approach to choose a prediction model for real-world environmental data. The approach consists of two steps: (1) model selection which determines the type of prediction model, such as linear or non-linear; (2) model optimisation which aims at using less environmental data for prediction but without any loss on accuracy. The latter step is achieved by automatically selecting non-redundant features without using specific models. Various experimental results on real-world data illustrate that using specific linear model can work pretty well and fewer environment distribution maps can quickly make better/comparable prediction with the benefit of lower cost of data collection and computation.