Th ELI1 07 How to Teach a Neural Network to Identify Seismic Interference S. Rentsch* (Schlumberger), M.E. Holicki (formerly Schlumberger, now TU Delft), Y.I. Kamil (Schlumberger), J.O.A. Robertsson (ETH Zürich) & M. Vassallo (Schlumberger) SUMMARY In this paper, we approach automated seismic interference (SI) identification for towed-streamer acquisitions in a new way. We show how to teach a neural network to identify SI based on features the human brain would use for such a task. This includes describing geophysical attributes such as amplitudes and wavefield propagation directions with instantaneous and multishot statistical measures. The statistical measures are then passed to a multilayer perceptrons neural network, first for training and then validation using noise-free synthetic data. Following encouraging results on the synthetic data, we applied the neural network to identify SI on field data from the North Sea and Barents Sea without any further training. The SI in a North Sea case study was very similar to the SI used for training; whereas, the SI in a Barents Sea case study was very different in terms of moveout and amplitude. Even though the neural network was only trained on noise-free synthetics, it successfully detected the SI in both cases.
Introduction Marine seismic data are often acquired in areas where other exploration activities already take place, leading to contamination of the acquired data. This contamination is commonly referred to as seismic interference (SI). SI caused by rig noise and the noise of passing vessels is typically limited to the vicinity of these activities; whereas, SI from other seismic surveys can be observed tens of kilometers away. Jenkerson et al. (1999) demonstrated the importance of SI detection and noted that, in a case study, the interference caused a 32% increase in dry hole rate for amplitude-based interpretation, and a 28% increase in missed opportunities. An abundance of SI removal techniques can be found in literature ranging from amplitude downscaling (Lynn et al., 1987) to more complex methods such as predictive filtering (Gülünay and Pattberg, 2001) or SI source location-based filtering methods (e.g., Warner et al., 2004). The recent introduction of multimeasurement towed-streamer acquisitions has provided new powerful means for SI identification and removal (Cambois et al., 2009; Vassallo et al., 2012). In this paper we focus on the automated detection of SI using feed forward neural networks and multimeasurement towed-streamer data. The automated detection can then be used for instance to guide SI removal techniques as described in Vassallo et al. (2012). Neural network for SI detection The way a human and a machine identify SI on a sequence of seismic shot gathers is quite different. Seismic data analysts can often identify SI simply by looking at a few common-shot gathers and identifying events that do not correspond to the norm, such as events arriving too early. They then go on to further investigate these events by looking at their moveouts or their behavior in different domains. In this work, an attempt was made to quantify how humans identify SI in a mathematical context that can be passed to a neural network. For this reason, measures were chosen that describe the way a signal looks in the domain spanned by time, inline offset, and shot number. Traditionally, SI may be identified based on its variability in arrival time from shot to shot, its anomalous amplitudes, its temporal length, and its moveout and we will continue using these attributes. However, with the additional measurement of the particle velocity vector, it is possible to identify the propagation direction of the SI at a receiver, as demonstrated by Rentsch and Brouwer (2011). This allows separation of arrivals as in or out of plane, with SI typically arriving out of plane. Propagation directions can be estimated using a number of methods such as those described by Rentsch and Brouwer (2011) or Vassallo et al. (2009). Furthermore, the measured directional vector may now be transformed from Cartesian to spherical coordinates. The sole use of angles in detecting SI is, however, dangerous because seismic signal associated with subsurface geology can arrive from out of the plane. For this reason, it is important to jointly consider the new measurements and the traditional attributes mentioned above. Finally, to improve generalization of SI detection and to reduce the influence of noise, statistical measures of the above geophysical attributes were computed to quantify how SI modifies data in bulk instead of at every receiver. This has the advantage that, through the use of robust estimators, the influence of noise is reduced. Furthermore, this practice improves the generalization of a neural network (NN) by measuring SI indicative features of attributes instead of only using the attributes that may, themselves, be less indicative. For the statistical description of instantaneous (individual shot) attributes, we used 1 st and 3 rd quartiles, Trimean, and variance measures of amplitudes and propagation directions. For multishot attributes, we used the shot-to-shot difference of the instantaneous Trimean and variance attribute and shot-to-shot correlation measures. For added stability of the shot-to-shot difference, we estimated a background measure (insensitive to SI) and subtracted the instantaneous values from the background value. Overall, we designed the statistical attributes such that they highlight the presence of SI as anomalous values. We then designed a NN to detect SI based on the statistical attributes described. NN were a natural choice due to its ability to learn relationships from data examples that are otherwise poorly understood. In particular, we chose to use multilayer perceptrons (MLP) that aim at representing how neurons in our brain are interconnected and pass along signals from one another to form a solution. Note that MLPs require supervised learning. A simple schematic representation of the NN is shown in Figure 1.
By convention, the first layer in such a network is called the input layer and, in our case, used to pass the statistical measures described above to the hidden layer. In the hidden layer, the inputs are processed and passed to the output layer. The output layer then computes the network output based on the hidden layer outputs. A(x) are the layer activation functions. Also shown in Figure 1 is a schematic network node. The input side is connected either to a neuron or to some other form of sensor that generates input. The amount of input that a node receives depends on its connections to the outputs of other nodes and sensors. These received inputs are weighted, by weights w, and summed, after which a bias b is added before being moved along to the activation function. The weights are scalar quantities that measure the importance a node places on an input. Biases are intrinsic properties of a node that are added to the weighted sum of inputs in an attempt to shift the activation function such that it is sensitive to certain cases of the weighted sum of inputs. The activation function then processes the received sum of weighted inputs and the bias in some form, and the result is passed to the output side of the neuron from where it may be queried. The output of a node is represented mathematically as follows:. (1) Here O j is the output of the j th node, w i,j is the weight of the j th node for the output of the i th node or the i th input, b j is the bias of j th node, N j is the set of input connections of the j th node, and A j () is the activation function of the j th node. Common activation functions are sigmoidal or linear and as shown in Figure 1 we used the hyperbolic tangent. To train the MLP, we generated synthetic plane-wave SI events with propagation directions randomly sampling the entire range of possible angles (azimuth and incidence angles) and added them to noisefree shallow-water synthetic North Sea data. We kept the wavelet and frequency content of the SI similar to the synthetic data. A subset of SI-contaminated data was then used for supervised training. That meant to input not just the SI-contaminated synthetics to the MLPs, but also the desired outputs. The desired output was defined as the ratio between SI and signal, quantifying the amount of SI contained in data rather than a binary SI / no SI measure. Errors between network outputs and the desired outputs are back propagated into the network to update weights and biases in an effort to reduce the difference between network outputs and desired outputs. Results on synthetics After training the NN on a random subset of synthetic North Sea data, the remaining data (~ 40 shots of 6 s recording time, nine streamers of 4 km length) were used to validate the NN SI detection capabilities. The remaining data consisted of twelve SI realisations that differed in propagation angle and SI repetition time, but also had small variations in amplitude and frequency content. The pressure data of one shot and streamer for the twelve SI realisations are shown in Figure 2a. Figure 2b shows the reference solution, indicating the amount of SI as a ratio between SI and signal with hot colours corresponding to significant amounts of SI and cold colours corresponding to small or no SI contamination. Figure 2d is the NN prediction of SI, which matches the desired solution in Figure 2b remarkably well. If we average the errors of the NN for all shots and steamers and plot them as functions of azimuth and incidence angle as shown in Figure 2c, we can see that the MLP performs equally well for all SI arrival directions. Results on field data Encouraged by the results on noise-free synthetics, we further applied the NN to field data from the North Sea as well as from the Barents Sea. The North Sea data were acquired using six 500 m long multimeasurement streamers spaced 75 m apart and a single source fired every 25 m. Seismic interference was observed when another seismic Figure 1 Simplified sketch of our multilayer perceptron neural network.
Figure 3 Twelve different SI realizations of the synthetic North Sea data, arriving with different angles. Pressure data are shown in a), the true amount of SI in b) and the NN s SI contamination prediction in d). In c), the mean square error between b) and d) as a function of incidence angle and azimuth is shown. crew started shooting about 70 km away at the start of one sail line and came closer to about 45 to 50 km at the end of the sail line. The hydrophone data contaminated with SI can be seen in Figure 3a for the 70 km SI distance and 3b for the 45 to 50 km distance. The SI has very similar features to the SI in our synthetics. Among differences with respect to the training data was that the data were not noise free and that the SI duration was longer. This longer duration resulted in SI being present in the same time-space window over two or more shots, which the NN had not seen during training. Despite these two new challenges, the NN identified the amount of SI contamination rather well as shown in Figures 3c and 3d. Also note that the SI amplitudes increased at the end of the sail line due to the closer proximity of the interfering crew. However, the NN performed well in both cases. The NN was also tested on data from a multiclient survey in the Barents Sea. The data were acquired Figure 2 North Sea pressure data contaminated with SI of linear moveout originating from a) 70 km away and b) about 45 km away. MLP predictions of SI contamination are shown in c) and d). using ten multimeasurement streamers, this time approximately 7 km long with a streamer spacing of 100 m and flip-flop sources fired 18.75 m apart. Seismic interference was observed when another seismic crew started shooting about 70 km away. The data contaminated with SI is shown in Figure 4a. We note that the SI was much weaker than in the previous example and also much weaker than in any MLP training data. Furthermore, the SI seemed to have a dominant frequency above 120 Hz, while the desired signal bandwidth was from 0-150 Hz. In Figure 4b, the SI contamination prediction from the NN is shown. Considering that the NN was trained on synthetics that contained only noisefree data with plane-wave SI, the result is quite encouraging. A prediction like this can now be used to
drive SI removal techniques only where it is required and, hence, not just save processing time but also protect SI-free parts of the record from unnecessary and potentially damaging filtering. Figure 4 a) Barents Sea pressure data (15 shots) contaminated with SI of hyperpolic moveout. b) MLP predictions of SI contamination. Conclusion We presented a method to quantify how humans identify SI in a mathematical context that can be passed to a NN. We described geophysical attributes such as amplitudes and wavefield propagation directions with instantaneous and multishot statistical measures. The statistical measures are then passed to a MLP, NN first for training and validation on synthetic data and secondly to field data. The SI in a North Sea case study was very similar to the SI used for training, while the SI in a Barents Sea case study was very different in terms of moveout and amplitude. Even though it was only trained on noise-free synthetics, the NN successfully detected the SI in both cases. Acknowledgement We thank Schlumberger for allowing us to publish this work and Statoil for permission to use the synthetic North Sea data. References Cambois, G., Jones, C., Lesnes, M., Sollner, W., Tabti, H.and Day, A. [2009]. Dual-sensor streamer data: calibration, acquisition QC and attenuation of seismic interferences and other noises. 71 st EAGE Conference & Exhibition, Extended Abstracts, X033. Gülünay, N. and Pattberg, D. [2001] Seismic Crew Interference and Prestack Random Noise Attenuation on 3D Marine Seismic Data. 63 rd EAGE Conference & Exhibition, Extended Abstracts, A-13. Jenkerson, M., Clark, H., Houck, R., Seyb, S., and Walsh, M. [1999] Effects of seismic interference on 3D data. 69 th Annual International Meeting, SEG, Expanded Abstracts, 1208-1211. Lynn, W., Doyle, M., Larner, K., and Marschall, R. [1987]. Experimental investigation of interference from other seismic crews. Geophysics, 52(11), 1501 1524. Rentsch, S. and Brouwer, W.G. [2011]. Methods to Process Seismic Data Contaminated By Coherent Energy Radiated From More Than One Source. Patent pending US20110096625 A1. Vassallo, M., Kostov, C. and Gonzalez, A. [2009]. Estimating and Using Slowness Vector Attributes in Connection with a Multi-Component Seismic Gather. Patent pending US 20090003132 A1. Vassallo, M., Eggenberger, K., van Manen, D., Rentsch, S., Brouwer, W., and Özbek, A. [2012] Effective seismic interference elimination enabled by multi-component data from marine acquisitions. 82 nd Annual International Meeting, SEG, Expanded Abstracts, 1-5. Warner, C., Brittan, J., van Borselen, R. and Fookes, G.P.C. [2004] Attenuation of Interference Noise on Marine Seismic Data. 66 th EAGE Conference & Exhibition, Extended Abstracts, Z-99.