An approach incorporating spectral recomposition results for training a neural network is proposed. The initial procedure is based on an inversion scheme for reconstructing the seismic spectrum of wavelets associated with a reflection event. This allows one to estimate each wavelet's temporal position within a seismogram. Since each calculated wavelet's temporal position is identified, this information is integrated into the seismogram as an additional feature for a trained neural network. A Fully Convolutional Network is trained with models that have features commonly found in offshore hydrocarbon reservoirs. The training step is based on composing a set of an acoustic velocity model set with its seismograms and their position matrices at the time of each wavelet for each seismogram. Next, the network training with a specific number of sets is performed. This approach was able to predict acoustic velocity models more accurately, especially for predicting the shape of salt bodies and the interface of inclined layers, which makes the proposed approach an interesting mean for estimating specific features in offshore hydrocarbon reservoir structures.