Decision making for environmental issues increasingly relies on numerical simulations and various observational data. However, the numerical models are limited by strong uncertainties because of poor input data and inaccurate physical, chemical, biological or mathematical modeling. Moreover, measurement instruments do not allow for a complete observation of environmental systems, and they often acquire noisy observations. Nevertheless, there is a strong need to optimally and jointly exploit numerical simulations and field observations for an objective assessment of risks on present and future times. In this context, it is critical to quantify the uncertainties of all information sources (numerical models, empirical rules, fixed observations, mobile observations, qualitative observations) and to evaluate the best estimates that are derived from all the information. The final scientific products that may help decision-making are the probability distribution of the target quantities, confidence intervals or probabilistic forecasts. These various products can be derived from ensembles of simulations possibly combined with observations by the so-called data assimilation methods. The ensembles can be calibrated, including for the forecasts, in order to approximate the distribution of simulation errors. Such methods are for instance operationally applied for weather forecasting. The distribution of ensembles can be processed for a better quantification of the uncertainty. It is for instance possible to derive risk maps. Practical applications, like the protection of populations after a nuclear disaster as in Fukushima, can benefit from such risk maps, e.g., to determine an evacuation zone.