Deep Learning-based Refinement of the Emission Source Identification Results

A hybrid emission source identification algorithm combining sensitivity operator-based and deep learning approaches is presented. A distributed emission sources estimate disregarding the point-wise structure of the “true” emission source is first obtained with general sensitivity operator-based algorithm. The algorithm produces both the source estimate and the sensitivity operator of the inverse problem. With the sensitivity operator, the aggregate is constructed that can be used to predict the result of the source identification in the considered conditions given the “true” source function. This aggregate can be applied to train the neural network to refine the source identification results by incorporating the a priori information of the unknown sources belonging to a specific class of multiple point-wise sources. The algorithm was tested in a remote-sensing inverse modeling scenario for the Baikal region. In the numerical experiments the convolution-type neural network was able to refine the inverse modeling results by providing the solution in the form of the localized point- wise sources.