Seismic facies classification using different deep convolutional neural networks

VeritasDGC. Summary We present a new approach to the simultaneous pre-stack inversion of PP and, optionally, PS angle gathers for the estimation of P-impedance, S-impedance and density. Our algorithm is based on three assumptions. The first is that the linearized approximation for reflectivity holds. The second is that PP and PS reflectivity as a function of angle can be given by the Aki-Richards equations (Aki and Richards, 2002). The third is that there is a linear relationship between the logarithm of P-impedance and both S-impedance and density. Given these three assumptions, we show how a final estimate of Pimpedance, S-impedance and density can be found by perturbing an initial P-impedance model. After a description of the algorithm, we then apply our method to both model and real data sets.

INTRODUCTION Controlled source audio-frequency magnetotellurics (CSAMT) is a frequency-domain electromagnetic sounding technique which uses a fixed grounded dipole or horizontal loop as an artificial signal source. CSAMT is similar to the natural-source magnetotellurics (MT) and audio-frequency magnetotellurics (AMT) techniques; the chief differences center around the use of the artificial CSAMT signal source at a finite distance. The source provides a stable, dependable signal, resulting in higher-precision and more economical measurements than are usually obtainable with natural-source measurements in the same spectral bands. However, the controlled source can also complicate interpretation by adding source effects, and by placing certain logistical restrictions on the survey. In most practical field situations these drawbacks are not serious, and the method has proven particularly effective in mapping the top 2 to 3 km of the earth's crust.

Summary Full waveform inversion has been successful in building high resolution velocity models for shallow layers. To be successful, it requires refracted waves or low frequencies in the reflection/refraction data. We revisit full waveform inversion theory in hopes of relaxing the dependence on low frequency reflections. We implement an approach allowing the updating of long wavelength components of the velocity model affecting the reflected arrivals even with absence of low frequency in the input data. Our tactic is based on a non-linear iterative relaxation approach where short and long wavelength components of the velocity model are updated alternatively. We study theoretically the associated Frechet derivatives and gradients and discussed how and why such a strategy improves the resolution that we can expect from full waveform inversion. The kernel of our approach is very similar to the algorithm of migration based travel time tomography proposed by Chavent et al. (1994). Finally we present a preliminary 2D application to a 2D Gulf of Mexico conventional streamer dataset.

The Earth Gravitational Model (EGM08) to degree 2160 is scheduled for completion at the end of April 2008. EGM08 is being developed using the best available terrestrial gravity from surface and airborne sources, gravity from satellite altimetry and marine sources over the oceans, and the latest GRACE‐derived satellite solutions. Critical to the success of this endeavour is the compilation of a complete and accurate 5′ × 5′ global gravity anomaly database that takes advantage of all the latest data and modeling for both land and marine areas worldwide. This paper will provide an overview of the data being used in the new model; describe the development of the EGM08 and show comparisons of the final model with independent truth data.

Migration with large apertures is known to be a powerful tool for imaging non‐specular (scattering) objects such as rough erosion surfaces, fault faces, fuzzy zones of increased fracturing and lithology replacements. Yet sometimes, the scatterers, even properly imaged, are masked by dominating specular reflections. To enhance the images of scatterers, we propose to create supplementary data cube where the specular component is attenuated relative to the scattered component. For this, we elaborate a special technique which is, basically, a Kirchhoff migration with weighting function modified by inclusion of an additional taper with controlled parameters. We illustrate through a field data example that the technique can essentially enhance seismic images of scattering objects.

A common issue of seismic data analysis consists in the lack of regular and densely sampled seismic traces. This problem is commonly tackled by rank optimization or statistical features learning algorithms, which allow interpolation and denoising of corrupted data. In this paper, we propose a completely novel approach for reconstructing missing traces of pre-stack seismic data, taking inspiration from computer vision and image processing latest developments. More specifically, we exploit a specific kind of convolutional neural networks known as convolutional autoencoder. We illustrate the advantages of using deep learning strategies with respect to state-of-the-art by comparing the achieved results over a well-known seismic dataset.

Seismic facies analyses as aid in regional gas hydrate assessments. Part-I: Classification analyses

By interpreting the full wavefield, full waveform inversion has the potential to become a key tool to interpret seismic data acquired in complex geological settings. However, its application requires low frequencies and large offsets to avoid ending up in a local minimum. This approach has been illustrated with marine data sets and acoustic full waveform inversion. Acoustic full waveform inversion can also be applied with land data sets when the acquisition and the pre-processing are planned correctly. To demonstrate the relevance of this dedicated approach, we invert a land data set acquired with low frequencies down to 1.5 Hz and 20 km offset. We show that the velocity retrieved by full waveform inversion supersedes the one derived by handpicked nmo-gather velocity analysis even when we start with a crude 1D model. With the acquisition of low frequencies and long offsets and a dedicated preprocessing, a high-resolution seismic migrated image could be obtained over this land area.

Summary Multiple reflections are commonly treated as noise in oneway imaging methods. High effort is put into research and data processing worldwide in an effort to suppress this source of noise. In a new perspective multiples are treated as valuable imaging information. Based on dual-sensor towed streamer measurement, we decompose the wavefield and apply up/down imaging of primary and multiple reflections. This approach is tested using shallow water synthetic data and finally applied on dual-sensor field data.

Delineating salt is an inherently difficult problem. Currently, there are several methods to detecting salt bodies in seismic. However, the result and volumes are often noisy, and are therefore hard to use for voxel based segmentation interpretation. We present a new method of using a 3D Sobel edge detector in combination with dip guiding to generate clearer images. We also introduce elements of weighting and normalization to achieve our results.

The goal of simultaneous shooting is to acquire better seismic data more quickly at lower total cost. Effective source deblending techniques provide us with one of the tools for accomplishing this goal. The use of compressive sensing theory gives us another tool by helping to increase the effective spatial bandwidth of our acquired data. Seismic surveys designed to collect both optimally sampled and blended data can reduce acquisition costs and significantly improve image quality. In this paper, we consider a joint deblending and reconstruction problem using the framework of a synthesis-based basis pursuit denoising model. The combination of a “deblending” operator together with a “restriction” operator leads to a joint inversion in which the data are both deblended and reconstructed at regular sampling intervals. Our inversion model can be further constrained by down-weighting the evanescent portion of the wavefield. We illustrate our method using both synthetic and real data examples simulating continuous-time recording under ocean bottom node (OBN) settings.

The recognition of paleokarst in subsurface carbonate reservoirs is not straightforward because conventional seismic interpretation alone is generally not sufficient to discriminate karstified areas from their surroundings. In the Loppa High (Norwegian Barents Sea), a protracted episode of subaerial exposure occurring between the late Paleozoic and mid-Triassic—Late Permian to Anisian—resulted in a significant overprinting of the previously deposited carbonate units. Here, we map the extension of the karstified areas using an integrated approach consisting of (1) a core study of critical paleokarst intervals, (2) a three-dimensional (3-D) seismic stratigraphic analysis, and (3) a 3-D multiattribute seismic facies (SF) classification. A core retrieved in the flat-topped Loppa High revealed breccia deposits at least 50 m (164 ft) thick, which probably resulted from cave collapses following the burial of the karst terrain. The SF classification was tested on a 3-D cube to (1) discriminate the respective SF related to the breccia deposits compared with other SF and (2) to estimate their spatial extent. Seismic-facies analysis suggests that breccias occupied the topmost area of the structural high, extending up to 12 km (7 mi) in width, 46 km (29 mi) in length, and tens of meters in thickness. The inference of such a large amount of breccia suggests that a significant part of this terrain was derived from the amalgamation of successive cave-development events—including periods of subaerial exposure and subsequent burial and collapse—resulting in a coalesced collapsed paleocave system. Previous observations from the Loppa High revealed the presence of karst plains associated with sinkholes, caves, and other dissolution phenomena associated with the breccia facies, further suggesting that a large volume of carbonate rocks in this area was affected by subaerial exposure and karstification. Our integrated approach and proposed karstification model could be applied to similar sedimentary basins that accommodate deeply buried carbonate successions affected by protracted episodes of subaerial exposure, where only few wells as well as 3-D seismic data are available.

An improved deep dilated convolutional neural network for seismic facies interpretation

Seismic facies analysis is to study the sedimentary environment of stratigraphic sequence and provides an important basis for reservoir prediction. Most of the existing analysis methods have low efficiency and heavily rely on manual experience, and therefore, it is difficult to interpret increasingly complex seismic data. Deep learning techniques can help to solve these problems and achieve automatic seismic facies classification. We regard seismic facies classification as a target segmentation problem and propose new method and training strategies. Our workflow primarily involves four sections. First, we process the manually annotated labels and seismic data with mirroring and cropping operations to ensure that network can accept input with arbitrary size and the model training is not limited to GPU memory. Second, data augmentation is applied to automatically generate massive training samples from the processed data. Third, we build two independent networks based on encoder–decoder architecture: one identifies all seismic facies simultaneously, and the other identifies single seismic facies in each model. However, both the results of the two networks have some drawbacks. Fourth, to overcome these drawbacks, we propose an ensemble learning method to get optimized model and test it on 3-D seismic data. The testing results manifest that the proposed method can improve the predictive ability of model, accurately describe the seismic facies, and can be applicable to entire seismic data volume.

Summary Deep learning have led to a certain extent breakthrough in seismic facies analysis(SFA). The greatest challenge in the SFA domain is how to cope with large seismic datasets and the limited amount of annotated samples, especially when employing supervised machine learning algorithms that require labelled data and larger training sample. we develop a supervised seismic facies analysis via an improved Auxiliary Classifier Generative Adversarial Network(ACGAN).To solve the problem of imbalanced seismic facies class distribution, and limited training data with labels, we used an improved ACGAN to generate synthetic seismic data sets, and then train Convolutional Neural Networks(CNN) using theses generated synthetic data. Fault data and seismic facies data test demonstrated that generated seismic data via the improved ACGAN can be used for synthetic data augmentation as for increased the performance of CNN and avoided overfitting compared to original seismic data training set. Therefore, we believe that the proposed method can effectively alleviate the problems of insufficient and imbalanced seismic data in SFA, and also can generalize to other seismic data classification applications.