Hydrocarbon exploration is strategic to the sustainability of economies of developing nations, with hydrocarbons still contributing to a major proportion of their energy basket. ( Ex : India , 30% — IEA 2020 ).
Understanding and characterisation of sub surface structures is paramount to evaluating new hydrocarbon ventures and undertaking exploration and development. Sub surface imaging is the de facto standard and an essential first step in building models of sub surface. Imaging is carried out through seismic surveys in which acoustic waves are sent through earth’s surface and reflections are recorded using precision surface equipment. Chaouch and Mari 2006 provides a great introduction to seismic acquisition terminology.
Discovery of potential hydrocarbon zones is largely dependent on the quality of these seismic images. However, seismic images are inherently noisy ( Low SNR ) and band limited due to acquisition constraints and instrument limitations.
Noise and other image artefacts introduced by the constraints and acoustic attenuation of earth degrade the quality of seismic images , thereby increasing the uncertainty in exploration decisions.The aforementioned quality problem is further exacerbated by the growing industry trend towards deeper and thinner exploration targets, whose corresponding imaging shows higher degradation , thereby increasing the risk of skipping potential reserves. Hence, enhancement of seismic images is a primary endeavour in the exploration workflow.
This article briefly explains a few facets of the problem of seismic image enhancement and focuses on the recent applied deep learning research undertaken in this area. I hope this article brings readers up to speed on the recent research and generates interest in application of deep learning to the problem area.
The rest of the article begins with a brief treatment of seismic images and seismic image quality and explains two sub-problems of seismic image enhancement - seismic resolution & noise contamination. This is followed by a summary of methodology , major contributions and performance metrics used in the recent research focused on applying deep learning to seismic image enhancement.
Seismic Images are spatial records of reflected acoustic energy from the boundaries of various sub surface layers of earth. These records are a primary source of information that feed into seismic processing workflows which estimate various structural , topographical and geological characteristics of the sub surface. They are very important in estimation of quantity and quality of hydrocarbon reserves. The product of seismic acquisition over a region is a grid of 2D / 3D slices of seismic images which are then approximated as a 3D seismic image cube of the respective region. As mentioned earlier, the quality of seismic images affects resource and capital intensive exploration decisions. There are two major factors that affect seismic image quality and seismic interpretation results — Seismic resolution and Noise contamination.
Seismic resolution is a measure of minimum separation between reflection events — spatially and temporally. The minimum separation determines how close two events could be so that they are sufficiently distinguishable from each other i.e resolved separately. Seismic resolution is dependent on the signal bandwidth of the seismic source, sensitivity of receivers, density of seismic acquisition (i.e the number and separation between sources and receivers). Vertical seismic resolution determines the thickness of the thinnest bed that is distinguishable whereas horizontal seismic resolution determines the minimum lateral dimension of a feature , for the feature to be distinguished as a single interface. A high resolution seismic image is one where thin edges and shapes can be discriminated from each other. The edge thickness and shapes , among others are important information used by experts to interpret ‘bright spots’ with hydrocarbon potential.
Li et al., 2021 describes two potential categories of techniques for improving seismic resolution. High Density Acquisition attempts to increase the horizontal seismic resolution by increasing the density of seismic image slices per area by using a larger number of sources and receivers per area. Broadband seismic technique increases the vertical seismic resolution by in- creasing the recording bandwidth. BroadSeis is a popular product and service of this category. Although these techniques have showed improved seismic resolution, they have drawbacks like — high capital and operational expenditure.
Noise is another facet of seismic image quality. Noise is any unwanted and un-interpretable energy found in seismic images. Seismic image quality is hugely affected by the amount of noise contamination during seismic surveys.Sources of noise could be environmental interference, human activity, power generation etc.Noise affecting seismic images is broadly classified into coherent and incoherent noise. Incoherent noises are random noises caused by environmental interferences - Ex: airwaves.These noises are assumed to be Gaussian or white.Coherent noises are caused by seismic sources. Several coherent noises exist. For example, linear noise or ground-roll noise, multiple reflections i.e. each receiver receiving multiple reflections from the same sub surface boundary.
Noise contamination causes blurring and other image artefacts like aliasing effects in seismic images. These effects prevent accurate processing and interpretation of seismic images. Noise attenuation in seismic images can lead to improvement in perceptual quality and hence seismic interpretation. Traditionally noise sources are modelled and filters are constructed to remove each type of noise. However, field estimation of noise which is needed for noise modelling , is found to be resource intensive and specific to each field. Previous noise attenuation methods based on filtering techniques differ for each types of noise. Custom filters are needed to be constructed based on the characterisation of noise contamination and parameters need to be tuned to achieve reasonable attenuation. Further, filtering based noise attenuation techniques need sequential processing. Hence, there is need for uniform framework for general noise attenuation which can perform noise attenuation without prior knowledge of noise.
The field of deep Learning has enabled substantial improvements in several computer vision tasks by learning hierarchical representations of data. A few extremely successful DL architectures include - CNNs , GANs. Image enhancement was previously perceived as a problem in the domain of digital image processing and solved by constructing custom filters i.e. Image enhancement was previously considered a filtering problem. However, image enhancement using filtering has the disadvantage of image degradation modelling and parameter tuning to fit specific degradation. With the success of deep learning in CV , it could be applied to Seismic image enhancement problems, by formulating them as image translation tasks , converting the filtering problem into a prediction problem. In particular , seismic image enhancement could be considered a problem of improving seismic resolution and attenuating seismic noise. The problems could be framed as Image Super resolution problem and Intelligent de-noising.
Super Resolution (SR) is the image processing procedure of approximating a High Resolution Image (HR) version of a given Low Resolution image (LR). The SR paradigm considers the LR image be a degraded version of an HR image, where the degradation characteristics are unknown. The degradation could be loss of detail, down sampling, corruption etc. Deep Learning solves the SISR problem by supervised learning of a model of the unknown degradation function that maps a LR image to HR image. The model is trained on a dataset of LR, HR image pairs. Such a trained model can be used to create HR images of previously un- seen LR images. A brief survey of Single Image Super Resolution which deals with generating a HR image from a single LR image , is provided by Yang et al., 2019
Among the architectures proposed for SISR, SRGAN proposed by Ledig et al., 2017 was hugely successful in preserving the perceptual quality of natural images at up-scaling factors of 2x & 4x, by defining a new objective function which accounted for perceptual similarity rather than similarity in pixel space. The authors employed a deep residual network (ResNet) with perceptual loss calculated using the feature maps of the pre-trained 19 layer VGGNet. The training was carried out in an adversarial manner. In addition to Peak SNR (PSNR) and MSE (Mean Squared Error), the authors utilised a different metric to measure the quality of resolution enhancement called Structural Similarity Index (SSIM), as proposed by Wang 2004.
Dutta et al., 2019 proposed a modified SRGAN to enhance and de-noise 2D & 3D seismic images. To augment the image enhancement, the authors used Conditional GANs, where low resolution images are concatenated with a one hot encoded lithology class label at different depths. The conditional information was added to the input images in the form of additional channels. Up-scaling layers previously proposed in SRGAN were not used. The network was trained separately on 3D, 2D and corresponding sets with conditional lithology introduced at different depths. Image degradation, to be learnt by the network, was simulated by passing HR images through 5Hz LPF filter and adding 50 % uniform random noise. Conditional GAN methodology showed a PSNR gain of 12.21% and SSIM gain of 19.84% over SRGAN , when trained on 2D seismic images and 10.9% PSNR & SSIM gain when trained on 3D seismic images.
Halpert, 2018 adopted DCGAN architecture , previously proposed in Radford et al., 2015 for de-noising and resolution enhancement of seismic images. The author reported increase in higher wavenumber spectra in enhanced images compared to ground truth and LR images and qualitative results in improvement of perceptual quality. Li et al., 2021 adopted CNN based architecture, a variant of U-net with sub pixel layer and residual blocks. The work enhanced synthetic seismic images with an emphasis on improvement of thin layers and small scale faults. The author trained the network for super resolution and de-noising using L1 loss coupled with MS-SSIM (Multi Scale - Structural Similarity Index). LR, HR pairs used were created by degradation emulated by adding random coloured noise with SNR ranging from 4-14. The work further provided evidence of image enhancement by showing improved performance on a fault detection task.
The task of de-noising seismic images has been successfully performed by both supervised methods as shown above and also unsupervised methods. A rigorous review of seismic image de-noising using deep learning has been provided by Yu et al., 2019. In particular, Chen et al., 2019 proposes a simple unsupervised auto encoder with a sparsity constraint which learns the representations of seismic signals from noisy observations. This method uses a simple three layer auto encoder and has been shown to be very effective in automatically removing spatially incoherent random noise. The key design insight of the above de-noising network is the introduction of a regularisation term based on KL divergence to the objective function of the hidden layer which helps the network to dropout non-trivial features that correspond to noise. Moreover, this work proposes the use of Local Similarity metric to quantitatively measure the signal damage due to de-noising. Such a measure of de-noising accuracy and reliability has not be used other recent studies. However, Chen et al., 2019 shows efficacy in de-noising only incoherent noise. The authors have also shown that quality of seismic de-noising is affected by the size of image patches used for training. The author explained this phenomenon by the fact that large patch size might impede the network from learning small scale features and small patch size would prevent the network from learning meaningful features. None of the other approaches have considered this phenomenon.
To summarise, this article breaks down the problem of seismic image enhancement into seismic resolution improvement and de-noising. It formulates these problems as prediction problems of image translation nature to be tackled by deep learning methods. It covers the relevant recent research in super resolution and de-noising. Note that the research reviewed is by no means exhaustive and I welcome any feedback and suggestions in improving my understanding of the subject.
I’m currently pursuing this problem area to write a master’s thesis. You can help by suggesting papers and methodologies discussing important ideas related to seismic image enhancement that I may have missed.
