نوع مقاله : مقاله پژوهشی
موضوعات
عنوان مقاله English
نویسندگان English
The convolutional model of the earth proposes that seismic data results from the combination of reflectivity and a source wavelet along with some random noise. As a result, enhancing seismic data resolution involves mitigating the impact of the seismic wavelet, an essential task in seismic data processing. Deconvolution of seismic data is a method utilized for this purpose, ultimately improving the interpretation of seismic sections by enhancing data resolution and bandwidth. This process can be applied to both single and multichannel seismic data. In single-channel deconvolution, it is assumed that multidimensional seismic data is a weighted combination of one-dimensional traces, and deconvolution is applied to each trace independently. On the other hand, multichannel deconvolution takes into account the time and spatial correlation between seismic channels, making it more advantageous compared to single-channel deconvolution methods. Deconvolution can be performed in nonblind or blind manners. In the case of nonblind deconvolution, information about the source wavelet is available. This type of deconvolution relies on recorded source wavelets or estimated wavelets obtained from sources such as well logs. Conversely, due to the often unavailability of precise information about seismic wavelets, blind deconvolution techniques are utilized to estimate the wavelets before or during the deconvolution process. Traditional seismic deconvolution methods commonly utilize optimization techniques to solve this issue, often heavily dependent on the accuracy of the optimization parameters and requiring human intervention for decision-making. To address these challenges, deep learning methods have been proposed. These methods automatically determine the parameters and reduce the need for human-computer interaction. In this paper, a deep learning-based deconvolution method is proposed for applying multichannel semi-blind deconvolution to two-dimensional (2D) post-stack seismic data sets. This approach employs 2D convolutional neural networks (CNNs) to generate high-resolution reflectivity images from seismic sections. CNN is a network where some of the hidden layers are convolutional, and the convolution layer combines the inputs as feature maps with convolution filters to form transformed feature maps. The trainable parameters are adjusted through backpropagation of errors using advanced optimization algorithms. We used the advantages of this network and proposed a CNN-based multichannel semi-blind deconvolution method. The method is referred to as semi-blind since it requires a wavelet estimation as the initial step to produce training data, but does not demand precise wavelet information, as a rough estimation suffices to achieve good results. The effectiveness of this proposed method has been confirmed using both synthetic and real seismic data examples incorporating complex structures and varying
signal-to-noise ratios. The experiments have demonstrated that the method can effectively handle diverse scenarios, highlighting a significant improvement in data deconvolution results through training with simple synthetic data.
کلیدواژهها English
