GNGTS 2019 - Atti del 38° Convegno Nazionale

GNGTS 2019 S essione 3.1 567 different statistics and a variety of missing traces distributions. Achieved results demonstrate that the proposed method is a promising strategy for seismic data pre-processing. Future work will investigate issues related to denoising of field data, exploring the feasibility of transfer learning by training convolutional neural networks on properly designed synthetic data and testing on field data. Moreover, investigations are needed for denoising of more challenging types of coherent noise (e.g., ground roll in land data) and different kinds of missing data (e.g., missing short offsets and crossline upsampling). Further studies on the network architecture and loss function could relax the need of pairs of gathers for the training set, thus helping in dealing with the problem of building a training dataset for denoising. Finally, we aim at exploiting the similarity between adjacent gathers to improve the reconstruction performances and extending of the procedure to 3D data. In the light of the promising results achieved with the proposed CNN, we believe this tool can pave the way towards even more efficient and accurate solutions. References Adamo, A., P. Mazzucchelli, and N. Bienati, 2015, Irregular interpolation of seismic data through low-rank tensor approximation : Geoscience and Remote Sensing Symposium, IEEE, 4292–4295. Fomel, S., 2003, Seismic re ection data interpolation with differential offset and shot continuation : Geophysics, 68, 733–744 Liu, G., X. Chen, J. Du, andK.Wu, 2012 , Randomnoise attenuation using f-x regularized nonstationary autoregression : Geophysics, 77, no. 2, V61–V69. Ronneberger, O., P. Fischer, and T. Brox, 2015, U-net: Convolutional networks for biomedical image segmentation: International Conference on Medical image computing and computer-assisted intervention , Springer, 234–241. Trickett, S., and L. Burroughs, 2009, Prestack rank-reducing noise suppression : Theory: 79th International Annual Meeting, SEG, Exp. Abstracts, 3332–3336. Zhu, L., E. Liu, and J. H. McClellan, 2015, Seismic data denoising through multiscale and sparsity-promoting dictionary learning : Geophysics, 80, no. 6, WD45–WD57. PROCESSING OF 3D SEISMIC REFLECTION DATA ACQUIRED FOR GEOTHERMAL EXPLORATION G. Miccichè, A. Tognarelli, A. Mazzotti Earth Sciences Department, University of Pisa, Italy Introduction. In this work we discuss the time processing sequence applied to 3D seismic reflection data acquired in the geothermal field of Larderello in southern Tuscany. The recorded seismograms are affected by poor signal-to-noise ratio (S/N), they are strongly contaminated by surface waves and ambient noise, and rough topography further complicate the problem. In addition, the investigated area is characterized by complex geology with strong lateral and vertical velocity variations and the occurrence of isotropic bodies and of altered rocks (Bertini et al. , 2006, Brogi and Liotta, 2006) seriously affects the reflectivity of the seismic events, so that the recordings of anomalous amplitudes, discontinuous reflections or absence of reflections in great portions of the seismograms are common. All of aforementioned issues make the estimation of a reliable stack volume a difficult task. We describe the experience of a careful processing employed to carry out a stack volume with an improved S/N ratio and where the most important reflections have been recovered. In the first part of this work we present the seismic data and in the second part we illustrate the processing sequence applied. Then we describe the processing steps that have paid a key role for the estimation of an improved stack. They are: the denoising step and, in particular the operations applied to attenuate the