The estimation of the probability of exceeding critical thresholds for a pollutant, the optimization of a measurement network for agro-climatic variables or the mapping of soils are few aspects that are covered by the spatial stochastic analysis and the modelling of environmental variables. Traditional methods for spatial statistics that have been under developments during the last few decades have led to satisfactory results in some of these fields, but they are still inappropriate or deficient in many others. One of these limitations is their incapacity to jointly use on a sound theoretical basis several sources of information that are of different accuracy and nature. The new methods of Bayesian Maximum Entropy are bridging this gap and are under developments since a few years. Our research aims at elaborating the methodological and theoretical approach that are needed for using these methods, as well as evaluating their performance for real-case studies. We try to generalize the BME approach to incorporate and predict continuous and categorical variables, simultaneously.