Atmospheric pollution control is becoming a key public health issue for our 21st century. The growing diversification of sensors and the growing variety and number of data they produce, lean some important issues for qualifying and processing information. Belief functions theory and data fusion could be a part of the solution of these issues. By the way the fact that fusion take each sensor in consideration for global information is a threat to the global uncertainty of merged information. This paper aims at building a method using belief theory and deep learning to get a merged information from a known attribute data model, while reducing uncertainty. Beyond the method, this paper aims at creating performance indicators to post validate the model.