GNGTS 2019 - Atti del 38° Convegno Nazionale

562 GNGTS 2019 S essione 3.1 These methods can be combined together to increase compression performances. Although this is a preliminary work, the results obtained on 2D propagation suggest that CNNs could represent a viable alternative for wavefield compression and open the way for further investigations. Future studies will involve the design of more specific CNN architectures for compression and interpolation of wavefield and the extension to 3D propagation, which should provide a sparser representation and larger compression ratios. Moreover, more sophisticated quantization strategies and lossless compression schemes for the encoded representation could improve the performances. Future work will also include the integration of CNN based compression techniques into FWI and RTM algorithms. References Anderson, J. E., L. Tan, and D. Wang, 2012, Time-reversal checkpointing methods for rtm and fwi: Geophysics, 77, S93–S103. Boehm, C., M. Hanzich, J. de la Puente, and A. Fichtner, 2016, Wavefield compression for adjoint methods in full- waveform inversion: Geophysics, 81, R385–R397. Dalmau, F. R., M. Hanzich, J. de la Puente, and N. Gutiérrez, 2014, Lossy data compression with dct transforms: Presented at the EAGE Workshop on High Performance Computing for Upstream. Goodfellow, I., A. Courville, and Y. Bengio, 2016, Deep learning: MIT press Cambridge, 1. Li, H., W. Yang, and X. Yong, 2018, Deep learning for ground-roll noise attenuation, in SEG Technical Program Expanded Abstracts 2018: Society of Exploration Geophysicists, 1981–1985. Luporini, F., M. Lange, M. Louboutin, N. Kukreja, J. Hückelheim, C. Yount, P. Witte, P. H. J. Kelly, G. J. Gorman, and F. J. Herrmann, 2018, Architecture and performance of devito, a system for automated stencil computation: CoRR, abs/1807.03032. Mandelli, S., F. Borra, V. Lipari, P. Bestagini, A. Sarti, and S. Tubaro, 2018, Seismic data interpolation through convolutional autoencoder, in SEG Technical Program Expanded Abstracts 2018: Society of Exploration Geophysicists, 4101–4105. DENOISING AND INTERPOLATION OF SEISMIC GATHERS VIA CONVOLUTIONAL NEURAL NETWORKS S. Mandelli, V. Lipari, P. Bestagini, S. Tubaro Dipartimento di Elettronica Informazione e Bioingegneria - Politecnico di Milano, Italy Introduction. Seismic exploration targets are more and more complex; therefore, the requirements for the quality of seismic data, both in terms of signal-to-noise ratio (S/N) and regularity and density of its spatial sampling, are increasing. However, various types of random and coherent noise corrupt seismic data sets. Furthermore, economic limitations, cable feathering, environmental constraints and badly acquired traces cause irregular space sampling in almost all seismic acquisitions. Most seismic processing algorithms, such as reverse-time migration, full-waveform inversion, and surface-related multiple elimination benefit from regularly sampled high-quality data. Therefore, most seismic processing workflows include denoising and trace interpolation as pre-processing steps. These topics have been widely investigated, either simultaneously or separately. We can roughly identify four families of algorithms: 1. Model-based algorithms implement an implicit migration-demigration pair (Fomel, 2003); their drawback is that performances are strongly affected by complex structural burden. 2. Prediction filters model seismic data as a local superposition of plane waves and are useful for both denoising and interpolation of regularly sampled data, which is a heavy limitation (Spitz, 1991).