The discrimination of small earthquakes from explosions based on the seismic signal recorded at teleseismic distances is an important and difficult task. The characteristics of a seismic signal are related to the energy release of the event through a complex functional relationship. The elastic and an elastic response characteristics of the propagation paths, and the response characteristics of seismometers, are undesired effects that usually fade the inherent source properties of seismometers. This shows that preprocessing stage, that performs a transformation from data space into a feature space to remove redundancy from recorded signals, is a critical process. In this study, 26 earthquakes from Eurasian events, and 25 underground nuclear explosions, which have occurred in East Kazakhstan (Semipalatinsk) test site is selected.Vertical components of the short period teleseismic records of these events is used. In order to reduce the mentioned disturbing effects, the following preprocessing are performed: In the first step, instrumentation correction was performed on all of the input records. In the second step, in order to extract suitable features from the P-wave, each record was filtered through 0.5 to 4Hz, using a fourth -order band pass Butterworth filter, and in the last step, as spectral characteristics suffer from the attenuation effects of the propagation paths, the spectra of p-wave were corrected for the effect of seismic attenuation. Constant level Ω, spectral corner frequency fc, and high-frequency spectral slop s, are features extracted from corrected P-wave spectrum, in this research. Another feature, obtained from spectrum, which is used is P-coda/P spectral ratio from 1 to 5 Hz.Recent developments of the neural network classifiers indicate that they are useful for solving many difficult real-world problems in discriminate analysis and pattern classification with powerful theoretical supports. According to successful and satisfactory history of neural networks, and by using the mentioned preprocessing methods, the current study is going to used two powerful type of neural networks, multilayer perceptron and radial basis function, in order to develop a discrimination model to separate the earthquakes from the explosions.As a results, 4 different models is conducted in this study. Two types of neural networks, MLP and RBF, and for each network two different datasets is used, P-coda/P spectral ratio and the values of Ω-fc-s. According to the results, it can be concluded that the preprocessed method which applied to the data were very suitable, and by use of neural networks and these preprocessed data, discrimination of the earthquakes and the explosions can be done with high reliability and precision.