This paper primarily investigates a method for monitoring atmospheric ducts through the inversion of automatic identification system signals based on deep learning algorithms. First, it introduces the most typical paradigm in deep learning algorithms, the deep feedforward neural network, and outlines the process for atmospheric duct inversion using deep learning. Then, the Latin hypercube sampling method is employed to generate an inversion database for training the deep feedforward neural network. To obtain the optimal deep neural network structure, the tabu search algorithm is combined with the Adam algorithm for optimization. Finally, simulations are conducted to verify the feasibility and effectiveness of using deep learning algorithms for inverting combined ducts. From the simulation results, compared with the Lévy flight quantum-behaved particle swarm optimization algorithm, the deep learning algorithm achieves higher inversion accuracy.