The state of our environment becomes a very important issue and especially people with health problems need more information and support in their daily life. This article presents an approach for forecasting values of several environmental-state variables as a basis for location-based services. We propose a system for making predictions for several spatial temporal variables using the Bayesian Network method as a machine learning technique. The system is based on a three-tier architecture, which assists the environmental data acquisition, processing and dissemination of predictions. To handle the missing values of data we use the structural expectation maximisation algorithm. The system's evaluation case study is based on real environmental data acquired from the Swiss national network. The data represents several environmental-state variables at different types of location, e.g. rural, urban, and at different times in a time span of a year.