This work considers the task of automatic identification of discontinuities from high resolution 2D seismoacoustic data. The training of artificial intelligence in this work was performed exclusively on real multifrequency seismoacoustic data, which were marked by an expert. As a result of the work the main problems arising at the decision of a task of creation of the automated marking tool on the basis of the real data marked out by the person have been allocated. The importance of selecting a loss function and the possibility of applying radical data compression while keeping the result close to human markup were noted.