Atmospheric point spread function (PSF) is an effective research and correction method for the adjacency effect of the optical remote sensing. Based on the atmospheric PSF acquired by Monte Carlo simulation, a two-layer feed-forward neural network which has enough hidden neurons with Sigmoid function and linear output neurons is designed and implemented. By means of Levenberg-Marquardt back-propagation algorithm, the relationship between the atmospheric PSF and its influence factors, such as atmosphere condition, spectral range and observation geometry is obtained. The results obtained show that our neural network can estimate the atmosphere PSF with 95% accuracy within relatively short time.