• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    The invention relates to a method for the automatic detection of geological objects belonging to a given type of geological object in a seismic image, based on a priori probabilities of belonging to the type of geological object assigned to each of the samples of the geological object. the image to interpret. According to the invention, a transformation of the image is first applied to a plurality of seismic attributes and then a classification method. Then, for each of the classes thus determined, a posterior probability of belonging for each of the samples of this class is determined to the type of geological object as a function of the prior probabilities of membership, of the class, and of a parameter α describing a confidence on the probabilities a priori of belonging. Then, with each sample of the image, from the class of this sample, the posterior probability of belonging to the type of geological object determined for the samples of this class is attributed. Then, the geological objects belonging to the type of geological object are detected from the posterior probabilities of belonging to the type of geological object thus determined for each of the samples of the image to be interpreted. - Application in particular to oil exploration and exploitation.
    公开号:FR3081232A1
    申请号:FR1854190
    申请日:2018-05-18
    公开日:2019-11-22
    发明作者:Pauline LE BOUTEILLER；Jean CHARLETY
    申请人:Centre National de la Recherche Scientifique CNRS；IFP Energies Nouvelles IFPEN；Sorbonne Universite；
    IPC主号:
    专利说明:

    The present invention relates to the field of exploration and exploitation of a fluid contained in an underground formation.
    In particular, the present invention relates to the field of exploration and exploitation of hydrocarbons present in an underground formation, or even the field of exploration of an underground formation with a view to geological storage of a fluid such as CO 2 , or the field of monitoring a geological storage site for a fluid.
    Oil exploration consists of finding hydrocarbon deposits within a sedimentary basin. Understanding the principles of the genesis of hydrocarbons and their links with the geological history of the subsoil has enabled the development of methods for assessing the petroleum potential of a sedimentary basin. The general approach to evaluating the petroleum potential of a sedimentary basin involves going back and forth between a prediction of the petroleum potential of the sedimentary basin, carried out on the basis of measured information relating to the basin studied (analysis of outcrops, seismic surveys, drilling, for example), and exploration drilling in the various zones with the best potential, in order to confirm or deny the potential predicted beforehand, and to acquire new information making it possible to clarify the predictions of petroleum potential in the basin studied.
    The petroleum exploitation of a deposit consists, from information gathered during the petroleum exploration phase, in selecting the zones of the deposit having the best petroleum potential, in defining exploitation schemes for these zones (for example at using a reservoir simulation, in order to define the number and positions of exploitation wells allowing optimal recovery of hydrocarbons), to drill exploitation wells and, in general, to set up the infrastructures of production necessary for the development of the deposit.
    A technique widely used in the field of oil exploration and exploitation is seismic prospecting. Seismic prospecting generally consists of three stages: the acquisition of seismic data, the processing of this seismic data, and finally the interpretation of the processed seismic data, then called seismic image.
    Interpreting a seismic image involves finding geological objects (also known as "geobodies") of interest within the seismic image. In general, we can distinguish within a seismic image:
    - large-scale geological objects, such as roofs and bases of geological layers, erosion surfaces, faults, etc. This type of object is conventionally interpreted by a structural geologist or a geophysicist.
    - smaller-scale geological objects, such as fluvial channels, dunes, turbidite deposits, mass transport deposits, commonly called MTD ("mass-transport deposits" in English). Typically, small-scale geological objects are interpreted by a sedimentologist geologist. These objects are characterized by particular patterns, formed by their internal reflectors.
    In addition, it should be noted that a seismic image can also include seismic artefacts, which do not correspond to any concrete geological object. These artifacts may come from the seismic processing that led to the seismic image. The geophysicist is particularly able to recognize such seismic artifacts.
    Thus, from the interpretation of a seismic image, one can deduce information on the underground formation studied, and in particular the presence, the shape and the type of geological objects present in the formation. The seismic amplitudes associated with these objects also provide information on the petro-physical properties of the formation. The quality of the information resulting from the interpretation of a seismic image is therefore essential. Indeed, from this information are constructed representations of the formation studied, called geological models, which make it possible to determine many technical parameters relating to the research, study or exploitation of a reservoir, of hydrocarbons for example.
    However, very often, the interpretation of geological objects in a seismic image is based on the experience of an interpreter, who knows, from experience, that the object in question has this or that organization of characteristic reflectors or in other words such " seismic facies >> particular.
    Methods exist to go towards an automation of the identification of certain types of geological objects in a seismic image. These methods considerably speed up the interpretation of seismic images (which is advantageous in the case of large volumes of seismic data to be interpreted), and also make rendering of this interpretation less dependent on the experience of the interpreter. These methods are currently relatively effective with regard to large-scale geological objects as defined above, but their performance is more limited in the case of smaller-scale geological objects.
    In particular, certain types of geological objects appear in the seismic image as regions which combine several different aspects, or in other words several seismic facies. For this type of object, visual detection is, in some cases, possible for an experienced interpreter, but more difficult to implement automatically, which is harmful for the interpretation of large volumes of seismic data.
    State of the art
    The following documents will be cited in the following description:
    Adams, R. and Bishof, L. (1994). Seeded region growing. IEEE Trans on Pattern Analysis and Machine Intelligence, vol. 16, pp. 641-647.
    Bishop, C. M., Svensén, M. and Williams, C.K.I. [1998] GTM: The Generative Topographic Mapping. Neural computation, 10 (1), 215-234.
    Carrillat, A., Randen, T., Sonneland, L. and Elvebakk, G. [2002] Seismic stratigraphic mapping of carbonate mounds using 3D texture attributes. 64th EAGE Conference & Exhibition, Extended Abstracts, G-41.
    Haralick, Robert M., Shanmugam, K. and Dinstein, I. "Textural Features for Image Classification >>. IEEE Transactions on Systems, Man, and Cybernetics. Flight. SMC-3 (6), 1973, pp. 610-621.
    Hashemi, H., De Beukelaar, P., Beiranvand, B. and Seiedali, M. [2017] Clustering seismic datasets for optimized faciès analysis using a SSCSOM technique. 79th EAGE Conference & Exhibition, Extended Abstracts, Tu B4 10.
    Kohonen, T. [1986] Learning vector quantization for pattern recognition. Technical report TKK-F-1601, Helsinki University of Technology.
    Lloyd, Stuart P. “Least Squares Quantization in PCM.” IEEE Transactions on Information Theory. Flight. 28, 1982, pp. 129-137.
    Matos, M. C., Malleswar (Moe) Y., Sipuikinene M. A. and Marfurt K. J. [2011] Integrated seismic texture segmentation and cluster analysis applied to channel delineation and chert reservoir characterization. Geophysics, 76 (5), P11-P21.
    Mumford, D. and Shah, J. (1989). Optimal approximation by piecewise smooth functions and associated variational problems. Communications on Pure Applied Mathematics, vol. 42, pp. 577-685.
    Zhao, T., Zhang, J., Li, F. and Marfurt, K. J. [2016] Characterizing a turbidite system in Canterbury Basin, New Zealand, using seismic attributes and distance-preserving selforganizing maps. Interpretation, 4 (1), SB79-SB89.
    Zhao, T., Li, F. and Marfurt, K. [2016b] Advanced self-organizing map faciès analysis with stratigraphic constraint. SEG Technical Program Expanded Abstracts, 1666-1670.
    Wang, Z., Hegazy, T., Long, Z. and AIRegib, G. [2015] Noise-robust detection and tracking of salt domes in postmigrated volumes using texture, tensors, and subspace learning. Geophysics, 80 (6), WD101-WD116.
    West, B.P. and May, S.R., inventors; Exxonmobil Upstream Research Company [2002] Method for seismic facies interpretation using textural analysis and neural networks. U.S. Patent 6,438,493, issued August 20, 2002.
    Classically, a distinction can be made between methods aimed at detecting geological objects having homogeneous seismic properties (cf. section 1 below) and methods aiming at detecting geological objects having heterogeneous seismic properties (cf. section 2 below). below).
    1. Objects with homogeneous seismic properties
    Among the existing methods aimed at detecting objects with homogeneous internal seismic properties, there are known so-called "supervised" approaches for processing the seismic image. For this type of method, a representation of the image is generally used in place of the original amplitude image. We can for example represent the image by image filtering (with different filters), or statistical calculations in each sample of the image with a predefined neighborhood (statistics of order 1, 2 or more). For the interpretation of seismic reflection data, we call "attribute" a quantity which can be calculated on the seismic image, which itself has a value per sample of the image. An attribute, or a combination of attributes (which can be reduced to the calculation in several stages of a single final attribute), is chosen for its representativeness of the seismic facies of the object sought. This attribute must be sufficiently discriminating for the facies in question, i.e. it must take very different values in the regions comprising the facies in question compared to the other regions of the seismic image.
    Then, for example by thresholding the values of this attribute, a binary image is determined in which the regions corresponding to the facies of interest are highlighted with respect to the others. It is also possible to use one or more post-processing operations more than a simple thresholding for the detection of regions of interest. We know in particular the region growth algorithms, which make it possible, starting from at least one source point in the object sought, to extend more and more, in the vicinity of this source point, the region considered to be “the object ”, until a predefined stopping criterion, relating to the contour obtained or to the values taken by the attributes within this contour, is reached.
    We know for example the document (Wang et al., 2015) which implements a method for the detection of a salt body in a seismic image, the salt bodies having in known manner homogeneous seismic properties. This document describes in particular the use of a textural attribute of the “texture gradient” type which characterizes in each pixel of the seismic image the variation of the texture in the vicinity of this pixel. Then, an algorithm for growing regions from a source point placed in the target object (the salt body) is used. Thus, detection is based on initially fixed properties, thanks to the homogeneous nature of the salt body (both physically and in its seismic imagery).
    The patent EP2659291 B1 is also known, which describes a method making it possible to take into account several attributes not by combining them in another, but by applying a post-processing by algorithm of detection of contour (and not of growth of region) for each attribute image separately. Contour detection makes it possible to introduce constraints a priori on the topology of the objects to be detected, because these are then represented by their contours (and it is these contours which undergo modifications during the application of the method) . The constraints introduced in the detection of contours make it possible to obtain topologically relevant objects despite the noise or the incomplete representation of the object in the raw data. The process is carried out either sequentially (by calculating the contour first only for one attribute image, then for another, etc.), or simultaneously (by calculating the contour by including multiple constraints coming from the different attribute images ). Thus, despite the introduction of a priori constraints on the topologies of objects, and despite the use of several different attributes, the objects must have the same ranges of attribute values, i.e. be homogeneous.
    Thus, in general, these approaches are characterized by the presence of a "manual" (as opposed to "automatic") step of choosing reference facies characterizing the object to be detected. To emphasize, or use, the resemblance between the facies of a region of the data and this reference aspect, these methods pass by attributes. Typical ranges of values for these attributes are chosen to define the baseline. Whatever the post-processing (region growth or detection and application of constraints on contours), a thresholding on the attributes is therefore used to distinguish the regions corresponding to candidate objects (then refined by the post-processing). This approach is interesting because it is object-based (for example to detect saline bodies). However, it is not suitable for objects or regions characterized by heterogeneous seismic properties.
    2. Objects with heterogeneous seismic properties
    In order to distinguish objects with heterogeneous seismic properties, it is necessary to use an approach allowing to bring out the differences between regions of the seismic image in a non-binary way. A distinction is made between supervised and unsupervised approaches.
    2.1. Supervised approaches
    In order to represent the variability of facies present in a seismic data set, one can consider classifying all the samples of the seismic image. Supervised approaches carrying out this classification are typically based on a selection of several reference facies, supposed to represent the typical facies found in the dataset. This selection is made through a manual study on the seismic data.
    In the document (Carrillat et al., 2002), the user chooses seismic attributes that he deems consistent to discriminate the different facets of the data set. The user then chooses thresholds (several thresholds on one or more attributes), which then makes it possible to classify each pixel of the data according to the predefined classes, by calculating the chosen attributes.
    Patent EP 1334379 A1 (US 6438493 B1) is also known, which describes a method according to which the user chooses zones of the seismic image to be interpreted which he considers as references for different facies; it also chooses seismic attributes it deems relevant. A neural network with back propagation will then learn (in a training phase) the thresholds to apply on each attribute to allow to classify all of the information according to the reference facies chosen initially.
    However, the selection of reference facies by the user has two major drawbacks:
    - It limits the expression of the variability contained in the seismic image. Indeed, the set of reference facies chosen by the user may not completely cover the variety of facies present in the seismic image. This can lead to cases where regions of the seismic image are assigned, during classification, to a facies A, by default only, when in reality their own appearance is very different from the facies A itself visually.
    - The set of reference facies defined by the user can give a representation of this variability which is unbalanced. Indeed, among these reference facies, certain facies (A and B for example) may be more similar to each other than others (A and C for example). This causes a bias in the probabilities that a region is classified in a particular facies, at the end of the classification. For example, if the user has defined a lot of classes (i.e. a lot of reference facies) for facies which are in fact relatively similar (ex: facies A, B1, B2, B3, B4, B5 ), and only one class for the remaining facies (ex: facies C), the reference set badly represents the variability of the data. For a region R, we could for example have probabilities of belonging to each reference facies distributed in this way: A (10%), B1 (10%), B2 (10%), B3 (10%), B4 (10%), B5 (10%), C (40%). Consequently, region R will be assigned to facies C, while its probability of belonging to one of the aspects A, B1, B2, B3, B4 or B5, is 60%.
    Thus, the variability of the facies present in a seismic image therefore risks being poorly represented due to the manual choice of the reference facies.
    2.2. Unsupervised approaches
    Unsupervised methods have the advantage of eliminating this potential bias introduced by the user. Indeed, for this type of approach, the definition of the reference aspects is carried out automatically, from the information contained in the data.
    In the field of seismic interpretation, the classifications in seismic facies based on unsupervised classification algorithms are often based on self-organizing maps ("Self-Organizing Map" in English, also known by the acronym "SOM" ; cf. for example (Kohonen 1986)) which are based on a neural network, or even an alternative algorithm to SOM methods, based on a probabilistic approach as described in the document (Bishop et al. 1998), and known as name of generating topographic maps (“Generative Topography Mapping”, and noted “GTM”).
    For these two algorithms, the clusters that are produced are specifically organized ("rows"), and continuous in their characteristics along this "ranking". In particular, (1) this makes the visualizations of the classification more useful for interpretation, since the close colors correspond to close and therefore similar clusters; and (2) increasing the number of clusters sought does not change the overall appearance of a visualization with a good color scale, since only more classes are then added within the same color scale.
    These advantages make facies interpretations easier for the user who then manually post-processes the data with its classification. Examples are provided in the document (Matos et al., 2011) which, after using an SOM type method, interprets the classification to determine the "delineations" of channels and reservoirs. We also know the document (Zhao et al., 2016), which uses a method called "DistancePreserving SOM" or DPSOM, which is an improvement on the SOM approach. The DPSOM is an improvement in the calculation of the SOM: the calculation of the position of the nodes in the latent space is adjusted so that the distances in the data space are preserved. This leads to a more discriminating representation (in latent space) between the different types of facies (more representative of the differences between attribute values in the data space). However, in general, these unsupervised classification algorithms require a manual step, by an interpreter, of class interpretation. In fact, unsupervised classification methods assign labels to all samples of a seismic image, but do not indicate which labels are relevant to the geological object sought. In addition, this type of method does not indicate whether groupings of labels (of classes) are often found in the seismic image. This study remains the responsibility of an interpreter, who must in particular be trained to understand the results of an unsupervised classification. In addition, this type of approach does not allow integrating a priori information.
    So-called semi-supervised methods are also known, which improve unsupervised algorithms by introducing constraints based on information of a type other than the seismic type.
    Such an approach is notably described in the document (Zhao et al., 2016b), which teaches an SOM algorithm constrained by a VMD (“Variational Model Decomposition”) decomposition, which is a decomposition of the seismic signal into several components and which provides information on possible stratigraphic cycles in a given seismic trace. This VMD information corresponds to the a priori information in this semi-supervised approach. It constrains the SOM method to classify in close (similar) classes the regions of the same stratigraphic layer. This a priori information has the advantage of being neutral, objective, since it comes directly from seismic information, which is a definite advantage compared to other a priori information defined by users (as in the supervised methods of detection of homogeneous objects, or supervised facies classification methods, described above). However, the approach is not intended to search for objects belonging to a particular type (such as BAT or saline bodies), and no post-treatment is provided in this sense.
    We also know the document (Hashemi et al., 2017) which uses facies maps from a well log analysis as a constraint. It is therefore a spatial constraint. However, as in the previous paragraph, the method does not make it possible to find objects belonging to a particular type (such as BAT or saline bodies).
    The interpretation performed after a classification is illustrated in US patent 9008972 B2. Once the seismic dataset has been classified into regions, each bearing a class label (in other words a class identifier), a new classification ("ranking") is given to the different classes themselves, according to their potential to validate a certain objective. . This "ranking" is created by an interpreter from a predefined criterion; the interpreter will (re) use a seismic attribute (or a combination of attributes), or a non-physical attribute but corresponding to his own interpretation of the classes. This will give a quantitative "ranking" of the regions of the seismic image, created according to a particular objective (for example, ranking the areas most likely to represent oil traps in decreasing order of probability). Thus, the interpreter must be expert on the attributes and algorithms used for the classification of the seismic image. In addition, the regions are analyzed one by one; an object which would be represented by a grouping of different classes could not be detected in this approach.
    Thus, these unsupervised or semi-supervised clustering algorithms have a strong potential to classify large datasets and to highlight all the internal variability of a dataset without omitting rarer seismic (or seismic facies) properties. than others (unlike supervised methods). However, they do not include the detection of geological objects. No particular strategy is given for the delineation of geological objects after having fully classified the samples of the image to be processed, even if constraints may have been added during processing (case of semi-supervised methods). A manual step to gather the clusters into groups of interest for the interpretation is for the moment necessary, as a postprocessing, to study / detect objects in the data. This step presents the interpreter with the risk of lacking in objectivity, and of requiring a lot of time and knowledge (concerning the functioning of the classification algorithm, etc.). In any case, in particular to guide interpretation towards obtaining delimited geological objects, this manual step is necessary in existing methods.
    Thus, fully automated methods of interpreting seismic images are currently limited to objects with homogeneous seismic properties. Methods exist to represent heterogeneous seismic facies of a seismic dataset, but not to detect objects made up of heterogeneous facies in an automatic and objective manner.
    The present invention aims to overcome these drawbacks by allowing automation of the detection of geological objects, in particular geological objects having heterogeneous seismic facies. In addition, the present invention can be configured on a seismic training image, the parameters determined on the training image then being able to be applied to another seismic image, for example including the seismic training image or a seismic image. from a neighboring seismic campaign.
    The method according to the invention
    The invention relates to a computer-implemented method for the automatic detection of at least one geological object belonging to a type of geological object in at least one seismic image of an underground formation, from a priori probabilities of belonging to said type of geological object attributed to each of the samples of said image.
    The method according to the invention comprises at least the following steps:
    A. defining a transformation of said image into a plurality of seismic attributes and applying said transformation to determine a plurality of seismic attributes of said image;
    B. An unsupervised classification is applied to said seismic attributes of said image and a class is determined for each of said samples of said image;
    C. For each of said classes, an a posteriori probability of belonging of said samples of said class to said type of geological object is determined as a function of said a priori probabilities of said samples of said image belonging to said class, of said class, and of a parameter a describing a confidence in said a priori probabilities;
    D. For each sample of said image, from said class of said sample, said sample is assigned said a posteriori probability of belonging to said type of geological object determined for said samples of said class.
    Then, according to the invention, said geological object belonging to said type of geological object is detected from said a posteriori probabilities of belonging to said type of geological object determined for each of said samples of said image.
    According to an implementation of the invention, said posterior probability yt posterior probability of belonging of said samples of said class / to said type of geological object can be determined as follows:
    L if PU = ,, i r | 1 it. if pi t <we determine qt = max | ke 1: n such that J] pj t <
    a · if Eyei ^ Pq = ^, then = s q .
    b. if ΣίΕΐ-.qPj.i then
    - If q = n, yt = s n
    - If q <n, y t = s q + 1 .
    or
    X a vector containing a value of said a priori probability of belonging for each sample of said seismic image;
    L a vector containing for each of said samples of said seismic image an identifier of said class to which said sample belongs;
    - S = {s 1 , s 2 , ... s n } = [s k , kel: n} with 0 = s 1 <s 2 <... <s n = 1, all the values taken by the elements of said vector X;
    C b ieL said class of identifier i;
    Pk, î = P (. S k c d> the proportion of points in Q having a priori probability value s k ;
    Advantageously, said seismic attributes can be textural attributes.
    Preferably, a dimension reduction can be applied to said attributes determined by said transformation.
    Very preferably, said reduction in size can be carried out by means of a principal component analysis.
    Alternatively, said reduction of dimension can be carried out by means of an analysis of an attribute selection method.
    According to an implementation of the invention, said unsupervised classification method can be a method of generating topographic maps.
    According to a variant implementation of the invention, at least said steps A) to D) can be applied to a first seismic image, then at least the following steps can be applied to a second seismic image:
    I. applying said transformation into seismic attributes to said second image to determine a plurality of seismic attributes of said second image;
    II. applying said unsupervised classification to said attributes of said second image and determining a class for each of said samples of said second image;
    III. for each of said classes of said second image and for each of said samples of said second image belonging to said class, said probability of belonging to said type of geological object determined for said class on said first image is assigned.
    According to a variant implementation of the invention, said geological objects belonging to said type of geological object can be detected in said at least one image by means of a thresholding method applied to the values of said a posteriori probabilities of belonging to said type of geological object determined for each of said samples of said at least one image.
    The invention also relates to a method of operating an underground formation comprising hydrocarbons, by means of at least one seismic image relating to said formation, in which:
    - Detecting in said seismic image geological objects belonging to at least one type of geological object by means of the method as described above;
    - We define a geological model representative of said formation from at least said detected geological objects and we define an exploitation scheme of said hydrocarbons of said formation from at least said geological model;
    - said hydrocarbons of said formation are exploited according to said exploitation scheme.
    The invention further relates to a computer program product downloadable from a communication network and / or recorded on a computer-readable medium and / or executable by a processor, comprising program code instructions for implementing the method. as described above, when said program is executed on a computer.
    Other characteristics and advantages of the method according to the invention will appear on reading the following description of nonlimiting examples of embodiments, with reference to the appended figures and described below.
    Brief presentation of the Figures
    - Figure 1 presents, on a theoretical example, the distribution of the classes determined by a classification method, with respect to a geological object to be detected.
    - Figures 2a and 2b present the samples of class C1 having an a priori probability of belonging to the geological object to be detected of 0 and 0.75 respectively.
    Figure 2c shows schematically the determination of the posterior probability associated with class C1.
    - Figure 3a shows the distribution of samples belonging to class C2, in relation to the geological object to be detected.
    - Figure 3b shows schematically the determination of the posterior probability associated with class C2.
    Figure 4a shows the distribution of samples belonging to classes C4, C5 and C6, in relation to the geological object to be detected.
    Figure 4b shows schematically the determination of the posterior probability associated with the class classes C4, C5 and C6.
    - Figure 5a shows the distribution of samples belonging to class C3, in relation to the geological object to be detected.
    - Figure 5b shows schematically the determination of the posterior probability associated with class C3.
    - Figures 6 and 7 respectively present three seismic learning sections selected from a global seismic volume and the a priori probabilities of belonging to a geological object of BAT type for each of the samples in these sections.
    - Figure 8 shows the distribution of the samples of the three learning images of Figure 6 according to 49 classes.
    - Figure 9 shows the posterior probabilities of belonging to a BAT type geological object for each of the three learning images of Figure 6, these posterior probabilities having been determined according to the method according to the invention.
    - Figure 10 shows the result of the propagation of the posterior probabilities of belonging to a geological object of the BAT type, determined for each of the three learning images of Figure 6, with 72 sections in the global seismic volume to be interpreted.
    - Figure 11 shows the samples of the global seismic volume to be interpreted having been detected as belonging to a geological object of PTD type, using the method according to the invention.
    Detailed description of the process
    In general, one of the objects of the invention relates to a computer-implemented method for the automatic detection of samples of at least one seismic image of an underground formation which belong to a predefined type of geological object. The method according to the invention can be advantageously applied to the detection of a type of geological object characterized by heterogeneous seismic properties, or in other words to a type of geological object characterized by heterogeneous seismic facies. Such heterogeneous seismic facies include, without limitation, mass transport deposits, commonly called mass-transport deposits (MTD), fluvial channels with heterogeneous seismic properties, or geological bevels (such as a beveled contact between inclined geological layers and an erosion surface). However, the method can also be applied for the detection of geological objects known to have homogeneous seismic properties, such as salt bodies, in particular when the internal facies of these objects vary slightly.
    Thus the method according to the invention is implemented for a given type of geological object, and is advantageously repeated if the detection of another type of geological object in the same seismic image is desired.
    The seismic image is indifferently a two-dimensional (2D) or three-dimensional (3D) image, where one of the axes would be the time axis (seismic image in the time domain) or the depth axis (seismic image in the depth domain), and the other axis (s) would correspond to the horizontal axes of a geographic space. The method according to the invention can also be implemented on a plurality of seismic, 2D and / or 3D images. This or these seismic images may have been obtained by means of seismic measurements (generally, we speak of “seismic acquisition campaign”) and of processing of these seismic measurements (generally, we speak of “processing seismic ”). The specialist is fully aware of the steps to be taken to obtain a seismic image of an underground formation.
    The method according to the invention requires a priori probabilities of belonging to the type of geological object sought, these probabilities being attributed to each of the samples of the seismic image considered. In other words, each sample (a pixel in the case of a two-dimensional seismic image, a voxel in the case of a three-dimensional seismic image) of the seismic image considered must have a value representative of the probability that this sample belongs to a geological object of the type sought.
    According to an embodiment of the invention without limitation, the method can include the following steps:
    1. Determination of a priori probabilities in the seismic image
    2. Determination of seismic attributes on the seismic image
    3. Classification of the seismic image in the attribute space
    4. Determination of a posteriori probability of belonging
    5. Propagation of posterior probabilities to another seismic image
    6. Automatic detection of geological objects
    7. Exploitation of the underground formation
    However, in general, the method according to the invention comprises at least steps 2 to 4 and step 6.
    Thus, the means of obtaining a priori membership probabilities, necessary for the implementation of the invention, can be obtained by the implementation of step 1 which is optional.
    According to an advantageous implementation of the invention when large volumes of seismic data are to be interpreted, steps 2 to 4 are applied to at least one seismic image which is a seismic learning image, and the process parameters determined on this seismic learning image are then applied, as described in step 5, to at least one second seismic image, for example of a larger dimension than the seismic learning image. Thus, in the case of a 2D seismic image to be interpreted, the learning seismic image can for example correspond to a window taken in the overall seismic image, of larger dimension. In the case of a 3D seismic image to be interpreted, learning can be carried out on a series of 2D seismic sections chosen from the 3D volume image, or else one or more 3D sub-volumes taken from this 3D seismic volume.
    Step 7, also optional, concerns the exploitation of the underground formation for which a seismic image has been interpreted by means of at least steps 2 to 4.
    The steps of the nonlimiting implementation of the method according to the invention described above are detailed below, steps 1.5 and 7 being optional.
    1. Determination of a priori probabilities in the seismic image
    This step is optional, because probabilities of belonging a priori to the type of geological object that one seeks to detect may have been previously determined and are then provided as input to the method according to the invention.
    During this step, it is a question of delimiting regions within the studied seismic image (which can also be a learning seismic image) and of assigning to these regions a probability of belonging to the object. sought geological, or in other words a probability of belonging a priori to the type of geological object that one seeks to detect.
    According to an implementation of the invention, an automatic delimitation method of geological objects is used, an alternative method to the method according to the invention, consisting for example of (i) calculating a seismic attribute chosen to be (at least partially) discriminating for at least some of the objects sought; (ii) normalize the values of the attribute between 0 and 1; (iii) assign this result to each region of the seismic image as a prior probability.
    According to another implementation of the invention, it is also possible to determine probabilities of belonging a priori to the type of geological object sought using other types of information than seismic information, such as for example log measurements carried out in at least one well crossing the formation studied. According to an implementation of the invention, it is possible for example: (i) for each well in which a log measurement has been carried out, determine a log of geological facies from a conventional processing of log measurements and deduce therefrom proportion of presence of the object to be detected as a function of the depth along the well studied, (ii) propagate this determined information to the various wells in all or part of the space covered by the seismic image to be interpreted, by means of '' interpolation or through a geostatistical method; (iii) assign this result, normalized between 0 and 1, to each region of the seismic image, as a priori probability.
    According to an implementation of the invention, regions are delimited in the seismic image considered (which may possibly be a learning image) by means of a manual and / or automatic interpretation of the geological objects of the seismic image considered and a probability of belonging to the geological object sought is assigned to each of these regions. According to an implementation of the invention:
    - we assign an a priori probability of belonging value of 1 for the samples of the seismic image belonging to regions for which it is believed that the delimited region corresponds to the type of geological object sought;
    - a value of probability of belonging a priori of 0 is assigned for the samples of the seismic image belonging to regions for which it is believed that the delimited region does not certainly correspond to the type of geological object sought;
    - we assign intermediate probability values (for example values of 0.25, 0.5, 0.75) for the other regions, according to the likelihood of their belonging to the type of geological object sought.
    2. Determination of seismic attributes on the seismic image
    During this step, a transformation of the seismic image (which can also be a seismic learning image) is defined into a plurality of seismic attributes and the transformation thus defined is applied to determine a plurality of associated seismic attributes. to the seismic image considered. Advantageously, at least four seismic attributes are determined, which potentially makes it possible to distinguish 2 4 = 16 different types of seismic facies, and thus to characterize geological objects having particularly heterogeneous seismic properties. Advantageously, attributes are chosen so that they represent the variability of the seismic facies of the seismic image considered.
    Preferably, the determined seismic attributes comprise textural attributes, which are attributes whose value in each sample characterizes the texture of the image in the vicinity of this sample. The value in each sample is generally a statistical magnitude of the distribution of the values of the seismic amplitudes in a lateral and vertical neighborhood of the sample of the seismic image (for example taking into account second order statistics in the vicinity of each sample of a seismic image). Thus, this type of attribute is particularly suitable for distinguishing different types of seismic facies in a seismic image because they provide information on the local organization of the reflectors in a fine manner in an image. Among the textual attributes, we know the attributes from the gray-level co-occurrence matrix (“Gray-Level Co-occurrence Matrix” in English, or GLCM). The GLCM attributes were initially introduced (see for example the document (Haralick, 1973)) for image processing in general, and are now widely used in seismic interpretation. The calculation of an attribute is done on a grayscale image, for a theta orientation and a predefined inter-pixel distance d. This calculation is done in two stages: first the construction of the co-occurrence matrix M (r, c), which indicates, for each pair (r, c) of gray levels, the number of pairs of pixels in the neighborhood window of a pixel (or sample of the seismic image) which are separated by a vector of norm d and of angle theta and having for gray levels r and c; then the calculation of statistical quantities on this matrix. In general, the textural attributes also have the advantage of being calculated according to the same calculation scheme, so that by calculating them all, we 'scan' all the possible patterns (for example, with the GLCM attributes, we calculates an attribute by orientation, all calculated in the same way).
    According to another implementation of the invention, the determined seismic attributes include spectral attributes, this type of attributes also carrying information on the size scales. According to another implementation of the invention, it is also possible to use conventional seismic attributes in seismic interpretation such as coherence, or even auto-correlation.
    According to an implementation of the invention, if the values of the determined seismic attributes are not of the same order of magnitude (for example if there is a ratio of 1 in 10 between two attributes), the values of the attributes are normalized thus determined.
    According to an implementation of the invention, if it turns out that the attributes thus determined are correlated with each other, an operation is performed to reduce the dimensions of the determined attributes. This can be done by applying a feature extraction method to attributes, such as a Principal Component Analysis (ACP), or a Feature selection method. According to an advantageous implementation of the invention in which the ACP method is applied, only the attributes projected are retained in the reference frame of the eigen vectors from the PCA whose sum of normalized eigenvalues is greater than 0.95. According to an implementation of the invention according to which an attribute selection method is applied, a selection of the initial attributes is kept, the selection being defined so that the selected attributes represent the total variance (or a percentage of 0.95 for example) of the point cloud and are as least correlated as possible (optimizing the selection of attributes therefore consists in minimizing the correlation of the selected game while maximizing the variance represented).
    3. Classification of the seismic image in the attribute space
    During this step, an unsupervised type classification method is applied to the seismic attributes determined at the end of step 2 described above, i.e. possibly to the attributes after dimension reduction. Thus, during this step, the seismic image in the multi-attribute space is considered, and no longer the seismic image expressed in seismic amplitude.
    According to the invention, the classification of the attributes resulting from the seismic image is carried out by means of a Self-Organizing Maps (SOM) algorithm as described for example in (Kohonen 1989), or else a Generative Topography Mapping algorithm. (GTM) as described for example in (Bishop, 1998), or even a K-means algorithm as described for example in the document (Lloyd, 1982). In the case of an SOM or GTM method, the classification consists in creating a non-linear representation of the data by a variety of small dimension (typically 2), in which there is a grid (potentially irregular) containing the class centers. During the implementation of the SOM or GTM algorithm, the positions of the class centers in the variety and in the data space are optimized (and in the case of GTM, the variety itself is parameterized). For this type of approach, the number of classes is predefined by the specialist in advance. The specialist is fully aware of how to define a number of ad hoc classes for the implementation of an unsupervised type classification method.
    At the end of this classification step, a class is assigned to each sample (a pixel in a 2D image, or a voxel in a 3D seismic image) of the seismic image considered. In other words, each sample of the seismic image considered has a class "label", or an identifier of the class to which it belongs. Typically, the class identifier for each sample in the seismic image is a class number.
    4. Determination of a posteriori probability of belonging
    According to the invention, it is a question of assigning, to each class obtained at the end of the preceding stage, an a posteriori probability of belonging to the type of geological object sought. This posterior probability is estimated based on:
    - a priori probabilities of belonging to the geological object sought, determined for example as described in step 1; and
    - the partitioning into classes determined at the end of stage 3; and
    - a parameter a, predefined (in particular by the specialist), which makes it possible to qualify confidence in the quality of the a priori interpretation of the geological objects of the seismic image (which can also be a learning seismic image) . According to an implementation of the invention, the parameter a for a given seismic image is between 0 and 1, the value 0 being assigned in the case of an a priori interpretation of geological objects very unreliable, and 1 being at contrary attributed in a case where confidence in interpretation is strong. The parameter a can also be used in order to give a certain "flexibility" to the detection algorithm according to the invention (0 for great flexibility and 1 for no flexibility). The parameter a can also make it possible to take into account the potential absence of annotations of certain objects, for example if we think that geological objects, corresponding to the type of object to be detected, have not been interpreted a priori in the seismic image (possibly learning).
    Thereafter, we note:
    X the vector containing a value of probability of belonging a priori for each sample of the seismic image considered;
    L the vector containing for sample the seismic image considered the identifier of the class to which the sample belongs;
    S = {sp s 2 , ... s n } = {s fe , ke 1: n} with 0 = s 1 <s 2 <... <s n = 1, the set of values taken by the elements of X;
    C b ie L the class number i of the classification described by L;
    Pk.i = P (. S k c d> the proportion of points of Q having a priori probability value s k ;
    Y the vector containing for each sample of the image considered the posterior probability.
    According to a preferred implementation of the invention, the posterior probability yi of belonging to the type of geological object sought for each class number i of the classification described by L is determined in the manner described below:
    • if Pi.i> ri = • If pi, <——, we determine:
    a + 1 qi = max 1 k G l: n such that £ Pj.î.
    (yei: fc
    at. siZ jÊ i: q Pj, i = - ^, Yi = s q ..
    b- ^ ljEi-. q Pj, i
    - if q = n, y t = s n
    - if q <n, Yi = s q + 1 .
    Thus, the method according to the invention makes it possible to exploit both a priori probability of belonging as well as information on the structure of the data set, represented by the result of the classification, while taking into account a degree of confidence in the a priori interpretation predefined by the specialist.
    At the end of this step, we obtain a unique posterior probability value by class, which approximates the arithmetic mean of the a priori probabilities of all the samples of the class (in particular we have equality with the arithmetic mean in the case where the parameter a is worth 1) but which can include a certain flexibility, via the parameter a, according to the confidence in the interpretation carried out a priori.
    Then, each sample of the seismic image considered is assigned an a posteriori probability of belonging to the geological object sought, according to the class (for example identified by a class identifier) to which the sample considered belongs.
    5. Propagation of posterior probabilities to another seismic image
    This step is optional. It applies in the case where the a priori probabilities of belonging to the geological object sought have been predefined on a seismic learning image, as defined above, and the automatic detection of geological objects of the type considered is desired on another seismic image, hereinafter called "global seismic image". It is obvious to the specialist that the learning seismic image and the overall seismic image preferably have, as far as possible, similar seismic parameters. For example, the seismic data at the origin of these two seismic images must preferably have been recorded with the same acquisition parameters (in particular the sampling steps) or have been corrected so as to reduce to the same parameters. acquisition (this type of correction is conventionally done in 4D seismic processing for example). Also, the seismic data at the origin of these two seismic images must preferably have undergone the same seismic processing or have been corrected so as to reduce to the same seismic processing (this type of correction is conventionally done in 4D seismic processing for example ); it is indeed important that the seismic amplitudes of the training images and of the global image to be interpreted are in similar ranges. The specialist is fully aware of these constraints and the means to achieve them, by routine seismic processing methods.
    At the end of step 4, one obtains in particular a posteriori probabilities of belonging to the geological object sought, associated with each of the classes determined on the seismic learning image. During this step, it is a question of "propagating" these posterior probabilities, determined on the learning seismic image, to the global seismic image.
    According to the invention, this propagation is carried out as follows:
    I. we apply the transformation into seismic attributes, as defined in step 2 and applied to the learning seismic image, to the global seismic image. A plurality of seismic attributes of the global seismic image are then determined, attributes which are the same (in type and number) as those determined on the training image. Preferably, the postprocessings (normalization, reduction of dimension; cf. step 2 above) applied to the attributes determined for the seismic learning image are advantageously also applied to the attributes of the global seismic image, using the same parameters of post-processing operations as those used for the seismic learning image.
    II. we apply the unsupervised classification method, as defined in step 3 and applied to the attributes of the learning seismic image, to the attributes of the overall seismic image. We apply here the same classification method as that applied in step 3 on the learning image, configured in the same way, and in particular with the same number of classes. We thus obtain for example a class identifier for each sample of the global seismic image.
    III. for each of the classes of the global seismic image and for each of the samples of this global seismic image belonging to this class (for example identified by a class identifier), the a posteriori probability of membership determined for this class is assigned to l seismic learning image.
    Thus, at the end of this step, a posteriori probabilities of belonging to the geological object sought for a global seismic image are obtained, from a posteriori probabilities of belonging to the geological object sought determined on an image. seismic learning. This notably allows a significant saving of time when large volumes of seismic data are to be interpreted since the determination of the a posteriori probabilities of membership is made finely only on the seismic learning image. This implementation of the invention also makes it possible to limit the size of the computer memory used for determining the probabilities of membership a posteriori.
    6. Automatic detection of geological objects
    During this step, it is a matter of automatically detecting, from the posterior probabilities of belonging to the geological object sought determined in step 4 and / or 5, geological objects of the type sought in the seismic image of interest.
    This detection can be applied with any type of known method for the automatic detection of geological objects having homogeneous internal properties, since the posterior probability values which have thus been determined can then be used as attributes for the implementation. of this type of method. In other words, the method according to the invention makes it possible, from very heterogeneous and complex seismic information, to determine an image having more homogeneous properties, making the application of methods for detecting geological objects having homogeneous internal properties possible.
    According to an implementation of the invention, it is possible, on the basis of the posterior probability of membership values for each sample of the seismic image considered, to automatically detect the geological objects of this image of the type of that sought by the following methods:
    we perform a thresholding on the posterior probability values: for example, we only keep the samples of the seismic image having posterior probability values greater than 0.8; and / or a region growth algorithm is used, implemented with a predefined criterion relating to the expected posterior membership probability values inside the objects of the type of the object to be detected. Reference may be made to the document (Adams and Bishof, 1994) which describes this type of method. In general, a regions growth algorithm consists in defining the starting points (“seeds” in English) of the regions, then in growing these regions via a measure of similarity between samples, until a criterion stop is reached. It is well known that these methods are effective when the starting points are chosen in homogeneous areas of the image to be interpreted.
    we use an edge detection algorithm, with explicit or implicit formulation. Reference may be made to the document (Mumford and Shah, 1989) which describes this type of method in general.
    7. Exploitation of the underground formation
    From at least the geological objects detected by the implementation of at least steps 2 to 4 and 6 described above for at least one type of geological object to be detected, we have precious information relating to the formation underground which we seek to exploit for example hydrocarbons.
    Indeed, it is notably classic to construct a meshed representation of the formation studied, a representation known as a geo-model, with a view to the exploitation of this formation. Each mesh of this mesh representation is filled by one or more petro-physical properties, such as porosity, permeability, type of sedimentary deposit etc. This meshed representation notably reflects the geometry of the geological objects encountered in the formation studied. It is thus most often structured in layers, that is to say that a group of meshes is assigned to each geological layer of the modeled basin. However, it advantageously reflects the presence of smaller geological objects (such as BAT, fluvial channels, etc.), by meshes having petro-physical properties distinct from the layers in which these objects are collected. In general, the construction of this model is based on data acquired during seismic surveys, measurements in wells, cores, etc. Thus, the information obtained by implementing the method according to the invention, by improving knowledge of the geological objects present in the formation, will contribute to building a more precise geological model of the formation studied.
    From such a meshed representation, the specialist can in particular select the zones of interest of the underground formation, and in particular the geological reservoirs likely to generate a high production of hydrocarbons (on the basis of criteria such as high porosity, high permeability, the presence of a cover rock, etc.). From such a meshed representation, the specialist can also predict the fluid movements in the training studied and plan its future development, by determining the operating patterns of the training. In particular, the determination of an operating scheme for an underground formation includes the definition of a number, a geometry and a location (position and spacing) of injector and producer wells, the determination of a type of enhanced recovery (by injection of water, surfactants, etc.), etc. An operating plan for a hydrocarbon reservoir must, for example, allow a high rate of recovery of hydrocarbons trapped in the identified geological reservoir, over a long operating period, requiring a limited number of wells and / or infrastructures . Conventionally, the determination of a hydrocarbon exploitation diagram is carried out using a digital flow simulator. An example of a flow simulator is the PumaFlow® software (IFP Energies nouvelles, France).
    Then, once an exploitation scheme has been defined, the hydrocarbons trapped in the reservoir are exploited according to this exploitation scheme, in particular by drilling the injection wells and / or producers of the exploitation scheme thus determined, and by installing the production infrastructure necessary for the development of the deposit.
    Computer program equipment and product
    The method according to the invention is implemented by means of equipment (for example a computer workstation) comprising data processing means (a processor) and data storage means (a memory, in particular a hard drive), as well as an input and output interface for entering data and rendering the results of the process.
    The data processing means are in particular configured to carry out the following steps:
    - a plurality of seismic attributes of the seismic image are determined;
    - an unsupervised classification method is applied to the seismic attributes;
    - we determine an a posteriori probability of belonging to the type of geological object sought for the samples of each of the classes;
    - we determine an a posteriori probability of belonging to the type of geological object sought for each of the samples of the seismic image;
    - and at least one geological object of the type sought is detected from the a posteriori probabilities of membership for each sample of the seismic image.
    Furthermore, the invention relates to a computer program product downloadable from a communication network and / or recorded on a computer-readable medium and / or executable by a processor, comprising program code instructions for implementing the method as described above, when said program is executed on a computer.
    Variant: obtaining a seismic image beforehand
    The method according to the invention is implemented on at least one seismic image of an underground formation. According to an alternative embodiment of the invention, a seismic image is determined prior to the implementation of the method according to the invention. Obtaining such a seismic image comprises a step of seismic measurements, then a step of seismic processing of these seismic measurements.
    The seismic measurement step is carried out using a seismic measurement device. This seismic measurement device comprises means (such as an explosive or a vibrator in terrestrial seismic, an air cannon or a water cannon in marine seismic) for emitting one or more seismic waves in the underground formation of interest, as well as means (such as acceleration (seismometer), vibration (geophone), pressure (hydrophone) sensors, or by a combination of elementary sensors of the above types (for example multi-component sensors)) the waves thus emitted and having been at least partially reflected in the formation studied. The specialist is fully aware of the means for acquiring seismic measurements in order to produce a seismic image of an underground formation.
    The seismic processing step is applied to the seismic measurements thus carried out on the underground formation studied. Indeed, the recorded seismic measurements as described above are very often unusable. Conventionally, the seismic processing can include a step of correcting the recorded seismic amplitudes, deconvolution, static corrections, noise filtering (random or coherent), NMO correction ("Normal Move Out" in English, "Normal Curvature" »In French), a stack (or« sommation »in French), and a migration (depth or time, before summation or after summation). These processing steps, often requiring complex and long calculations, are carried out on a computer. The resulting seismic data is then called a seismic image. These seismic images are most often represented on a computer, by a mesh or grid, each mesh corresponding to a lateral and vertical position (the vertical direction corresponding to time or depth depending on whether the processing has resulted in a time image or depth image) within the formation studied, and being characterized by a seismic amplitude. The specialist is fully aware of the means for applying an ad hoc seismic treatment to seismic measurements in order to produce a seismic image of an underground formation, an image intended for seismic interpretation.
    Examples of realization
    The characteristics and advantages of the process according to the invention will appear more clearly on reading the application examples below.
    First example
    In this example, it is a question of illustrating, on a purely theoretical example, the principle of the method according to the invention.
    Figure 1 shows a geological object (outline in bold) which has been delimited by a specialist manually in a seismic image, as well as the classes (classes C1 to C6) resulting from the application of an applied classification method. The prior probabilities are discretized as follows [s - ^, $ 3, $ 4, ^ 5] = [0, 0.25, 0.5, 0.75, 1]. The probability s 4 was assigned inside the contoured area, and the probability Si was assigned to the samples outside this outline. Thus, for this example, only two probability values are represented in the global distribution of a priori probabilities.
    In this example, the risk of forgetting to interpret a geological object of the type sought in the seismic image is considered to be around 40 percent. According to the embodiment defined above, this amounts to setting a = 0.6 (as a reminder, a = 1 corresponds to the case where the interpreter is sure to have spotted all the objects).
    The posterior probabilities by class are examined below • Class 1:
    Class 1 corresponds to pattern C1 in Figure 1. The probability that the previously created contour corresponds to points of the object sought has been estimated at 0.75. Outside this area, the probability is 0. We must therefore now estimate the proportion of points of class 1 in each of the a priori probability zones, ie the proportion p 4; 1 = p (s 4 | C 1 ) of pixels of Q which are in the a priori probability region s 4 = 0.75, and the proportion pOJQ) of pixels of Q which are in the a priori probability region s x = 0. The other probability values a a priori are not represented in Q, therefore the proportions p (s 2 l c i), pOsIQl) and pOsICl) are all zero. For this example, the area of the segmented surface having an a priori probability of 0 is small (cf. hatched portion in Figure 2a). The area of the segmented surface having an a priori probability 0.75 is larger (cf. hatched portion in Figure 2b). The calculation of y15 posterior probability associated with class 1, takes into account this distribution of values and is illustrated schematically in Figure 2c. For each a priori probability value sk, the proportion of class 1 points having this value sk is represented by a stick surmounted by a gray disc. The cumulative value of these proportions (Z 7el: fc Py, i) is represented by the dashed line, and the black disc represents the first point of the cumulative distribution curve greater than Indeed, the value determines the threshold for research of q as defined in section 4 above. Since p ± <we determine q, (in this case q = 3) and the posterior probability y ± of class C ± is therefore Sg +1 = 0.75.
    • Class 2
    The implementation of the method according to the invention for class 2 (motif C2 in Figure 1) is identical to that presented for class 1. Figure 3a illustrates the distribution of the image samples belonging to class 2 ( see pattern C2 in Figure 1), in relation to the previously interpreted object (delimited by the bold outline). The calculation of y 2 , posterior probability associated with class 2, takes into account this distribution of values and is illustrated schematically in Figure 3b. The threshold value is compared to the cumulative distribution curve (Z 7el: fc Py, i) and the posterior probability y 2 of class C 2 is therefore Sq +1 = 0.75.
    • Classes 4, 5, and 6
    The method according to the invention is applied to classes 4, 5 and 6 (patterns C4, C5 and C6 in Figure 1). Figure 4a illustrates the distribution of the image samples belonging to these classes (cf. patterns C4, C5 and C6 in Figure 1). In these three cases, the intersection between the zones for which the a priori probability of the presence of an object of interest and the zones resulting from the classification step is empty. Thus, p ± > which implies that 74,5,6 = ° In other words, the first point of the cumulative distribution curve being above the threshold all the points of these zones take as posterior probability l 'abscissa of this first point, that is Si = 0, as illustrated in Figure 4b.
    • Class 3
    Finally, class 3 (see reason C3 in Figure 1) only intersects a single a priori probability surface, as presented in Figure 5a, but, unlike classes 4, 5 and 6, all of its points are found at the a priori probability s 4 = 0.75. The cumulative distribution curve (see Figure 5b) leaves zero only at the level of 0.75. Consequently, s q = 0.5 and the posterior probability retained for class 3 is Sq +1 = 0.75.
    Second example
    For this example, the geological objects are to be delimited in a 3D seismic image (a 3D seismic cube after summation, or even “post-stack” in English) and correspond to sedimentary deposits of the MTD type (“mass transport deposits” or "mass transport depots" in French). These objects are recognizable in seismic by an interpreter warned by the different seismic facies they contain and their arrangement. The seismic image has a vertical sampling of 4 ms, and an inter-trace distance of 25 m in both directions (crossline, inline). The BATs that we are trying to detect have a size of several tens of kilometers in lateral extension, and at most 100 ms in thickness. We use three 2D sections of the cube as seismic learning images (see Figure 6 which presents these three seismic learning images A, B, and C). A pixel in a seismic image therefore has a size of 4 ms x 25 m. These three seismic sections were interpreted by an interpreter, who delimited zones with very high a priori probability of belonging P to a geological object of the MTD type (zones in white in Figure 7 for the three seismic learning images A , B and C), a lower probability (areas in dark gray in Figure 7 for the three sections A, B and C) and areas do not include BAT (areas in black in Figure 7 for the three seismic images d learning A, B and C).
    For this illustrative example, we use textual attributes, of GLCM type as defined above. In particular, the textural attributes conventionally called "Contrast", "Correlation", "Energy" and "Homogeneity" are used (see for example the documents (Haralick, 1976; West et al., 2002)). We choose a study window (the neighborhood around the pixel, in which the co-occurrence matrix will be calculated) of 11 x 11 pixels. Two different scales are used: distances of d = 1 and d = 2 pixels. For each scale, we use all the possible theta orientations (4 possible orientations for scale 1 and 8 for scale 2). For the calculation, it is necessary to limit the number of possible levels for the values of seismic amplitude (gray levels). This number is often 256, or 64, or 8 at least; in this case, only two gray levels are chosen. The computations are thus shorter, because the computation time of the GLCM attributes increases with the square of the number of gray levels. To compensate for the loss of amplitude information, an attribute from the envelope of the seismic signal is added to the GLCM set of attributes: for a pixel, this is the result of the Gaussian filtering of the image of l seismic envelope, with a 2D Gaussian nucleus, of size close to the window of 11 x 11 pixels. The GLCM attributes, as well as this envelope-based attribute, are calculated for all the pixels in the training image. The attributes are then normalized.
    For this implementation of the invention, the dimension of the attribute space is reduced. For this, we use the Principal Component Analysis (PCA), which we apply separately to the attributes from the scales d = 1 and d = 2. The attribute from the envelope is left separate from the CPA. For each of the two scales, only the eigenvectors are kept whose cumulative sum of normalized eigenvalues is greater than 0.97. We then have three new attributes for the scale d = 1 and 6 for the scale d = 2. We impose weights of 0.5 on the attributes of scale 2 (to ensure an equal representation of scales 1 and 2), and we add to this set of nine new attributes the one from the envelope, with a weight 3.
    Thus, the basis of the new attribute space is made up of these 10 eigenvectors: the first three eigenvectors from the PCA on the scale 1, the following 6 (with a weight 0.5) from the PCA on the scale 2, and the last from the envelope, with a weight 3.
    We then apply an unsupervised classification algorithm of GTM type (“Generative Topography Mapping >>, and noted“ GTM >>) as defined above, in its 2D version (the variety created will be of dimension 2). We thus determine for each pixel of each learning image, a class label among 49 classes (cf. Figure 8 which presents the distribution of the 49 classes on each of the seismic learning images A, B, and C).
    The preferred implementation of the method according to the invention is then applied, as described in step 4, to assign to each class an a posteriori probability of the presence of the objects. The parameter a is fixed at 0.8 (degree of confidence in the interpretation). We then determine, for each class, the associated posterior probability. Figure 9 presents the posterior probabilities P ’thus obtained for each of the seismic learning images A, B and C.
    Then the posterior probabilities thus determined for each of the seismic learning images are "propagated" to a series of 72 parallel 2D sections extracted from the 3D seismic cube, which are advantageously parallel to the learning images. The sections are extracted with a sampling of one section every 250 m (1 section out of 10), which is rather thin in relation to the size of the objects sought. Figure 10 represents a part of these sections dressed in posterior probabilities P ’.
    For illustrative purposes and in order to determine a complete seismic volume, each section is replicated 10 times, to represent the overall 3D seismic volume. An interpolation of the posterior probability values between each section could have been advantageously carried out. One then applies a postprocessing by thresholding on the values of the probabilities. More precisely, for this example, only the pixels whose posterior probabilities are equal to 1 are kept, followed by a thresholding on the size of the objects and on the distribution of the class numbers inside the objects.
    We obtain the delimitation of the desired BAT presented in Figure 11. It can be observed that the method according to the invention has thus made it possible to (i) automatically delimit these objects without omitting any part due to the difference in facies, and (ii) to keep the information relating to these various facies within the delimited object itself, which is useful later on for the interpretation of the properties of the object.
    权利要求:
    Claims (11)
    [1" id="c-fr-0001]
    1. Process implemented by computer for the automatic detection of at least one geological object belonging to a type of geological object in at least one seismic image of an underground formation, from a priori probabilities of belonging to said type geological object assigned to each of the samples of said image, characterized in that at least the following steps are carried out:
    A. defining a transformation of said image into a plurality of seismic attributes and applying said transformation to determine a plurality of seismic attributes of said image;
    B. An unsupervised classification is applied to said seismic attributes of said image and a class is determined for each of said samples of said image;
    C. For each of said classes, an a posteriori probability of belonging of said samples of said class to said type of geological object is determined as a function of said a priori probabilities of said samples of said image belonging to said class, of said class, and of a parameter a describing a confidence in said a priori probabilities;
    D. For each sample of said image, from said class of said sample, said sample is assigned said a posteriori probability of belonging to said type of geological object determined for said samples of said class;
    and in that said geological object belonging to said type of geological object is detected from said a posteriori probabilities of belonging to said type of geological object determined for each of said samples of said image.
    [2" id="c-fr-0002]
    2. Method according to claim 1, in which said posterior probability y t posterior probability of belonging of said samples of said class i to said type of geological object is determined as follows:
    i · if Pu > - ^, yt = if.
    1 z · i 1) ii. if Pu <we determine qi = max | ke 1: n such that Σ Pj, t - and a- if Σ 7 · Ε ι :( pj'i = then yt = s q .
    or
    b. siljei-.qPj.i <777 ' a if q = n, y t = s n if q <n, y t = s q + 1 .
    X a vector containing a value of said a priori probability of belonging for each sample of said seismic image;
    L a vector containing for each of said samples of said seismic image an identifier of said class to which said sample belongs;
    - S = {sp s 2 , ... = {s fe , ke 1: n} with 0 = s 1 <s 2 <... <s n = 1, the set of values taken by the elements of said vector X;
    C b ie L said identifier class i;
    Pk.t = p (. S k c ù, the proportion of points in C t having a priori probability value s k ;
    [3" id="c-fr-0003]
    3. Method according to one of the preceding claims, wherein said seismic attributes are textural attributes.
    [4" id="c-fr-0004]
    4. Method according to one of the preceding claims, in which a dimension reduction is applied to said attributes determined by said transformation.
    [5" id="c-fr-0005]
    5. Method according to claim 4, wherein said reduction of dimension is carried out by means of a principal component analysis.
    [6" id="c-fr-0006]
    6. The method of claim 4, wherein said reduction of dimension is carried out by means of an analysis of an attribute selection method.
    [7" id="c-fr-0007]
    7. Method according to one of the preceding claims, in which said unsupervised classification method is a method of generating topographic maps.
    [8" id="c-fr-0008]
    8. Method according to one of the preceding claims, in which at least said steps A) to D) are applied to a first seismic image, and in which, at least the following steps are applied to a second seismic image:
    I. applying said transformation into seismic attributes to said second image to determine a plurality of seismic attributes of said second image;
    II. applying said unsupervised classification to said attributes of said second image and determining a class for each of said samples of said second image;
    III. for each of said classes of said second image and for each of said samples of said second image belonging to said class, said probability of belonging to said type of geological object determined for said class on said first image is assigned.
    [9" id="c-fr-0009]
    9. Method according to one of the preceding claims, in which said geological objects belonging to said type of geological object are detected in said at least one image by means of a thresholding method applied to the values of said posterior probabilities of belonging to said type of geological object determined for each of said samples of said at least one image.
    [10" id="c-fr-0010]
    10. Method for operating an underground formation comprising hydrocarbons, by means of at least one seismic image relating to said formation, in which:
    - geological objects belonging to at least one type of geological object are detected in said seismic image by the method according to one of claims 1 to 9,
    - We define a geological model representative of said formation from at least said detected geological objects and we define an exploitation scheme of said hydrocarbons of said formation from at least said geological model;
    - said hydrocarbons of said formation are exploited according to said exploitation scheme.
    [11" id="c-fr-0011]
    11. Computer program product downloadable from a communication network and / or recorded on a computer-readable medium and / or executable by a processor, comprising program code instructions for implementing the method according to one of the previous claims, when said program is executed on a computer.
    类似技术:
    公开号 | 公开日 | 专利标题
    Wang et al.2018|Successful leveraging of image processing and machine learning in seismic structural interpretation: A review
    CN101932954B|2013-04-10|Subsurface prediction method and system
    EP2761329B1|2016-01-13|Method for validating a training image for the multipoint geostatistical modeling of the subsoil
    EP2846176A1|2015-03-11|Method for the exploitation of a subsurface deposit including at least one outcrop by means of photogrammetry
    FR3069330A1|2019-01-25|DEEP MODELING BASED ON DEEP LEARNING
    EP3570074A1|2019-11-20|Method for detecting geological objects in an image
    EP2356493A2|2011-08-17|Method for geologically modeling seismic data by trace correlation
    FR3027688A1|2016-04-29|
    Bruna et al.2019|A new methodology to train fracture network simulation using multiple-point statistics
    EP2246714A2|2010-11-03|Method for modelling an underground heterogeneous medium using multipoint statistics
    EP2873989B1|2018-07-04|Method for building a grid representing the distribution of a physical property of a subsurface formation by conditional multi-point statistical simulation
    Mattéo et al.2021|Automatic fault mapping in remote optical images and topographic data with deep learning
    Gillfeather-Clark et al.2021|A comparative study of neural network methods for first break detection using seismic refraction data over a detrital iron ore deposit
    US20200348433A1|2020-11-05|System and method for lateral statistical estimation of rock and fluid properties in a subsurface formation
    FR3032532A1|2016-08-12|
    EP3140677B1|2018-06-27|Method for processing seismic images
    EP3835835A1|2021-06-16|Method for updating a stratigraphic model of a sedimentary basin according to measurements
    US20220035069A1|2022-02-03|Systems and methods for identifying subsurface features as a function of position in a subsurface volume of interest
    WO2022043051A1|2022-03-03|Method for determining an electrofacies interpretation of measurements carried out in a well
    Cassel2018|Machine Learning and the Construction of a Seismic Attribute-seismic Facies Analysis Data Base
    CN112731522A|2021-04-30|Intelligent recognition method, device and equipment for seismic stratum and storage medium
    Deighton2006|3D texture analysis in seismic data
    Shukla et al.0|SEMANTIC SEGMENTATION TO RECOGNIZE SALT DEPOSITS
    同族专利:
    公开号 | 公开日
    EP3570074A1|2019-11-20|
    FR3081232B1|2020-10-02|
    US11226424B2|2022-01-18|
    US20190353811A1|2019-11-21|
    CA3043212A1|2019-11-18|
    BR102019010073A2|2019-12-10|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    US6438493B1|2000-09-29|2002-08-20|Exxonmobil Upstream Research Co.|Method for seismic facies interpretation using textural analysis and neural networks|
    EP1334379A1|2000-09-29|2003-08-13|ExxonMobil Upstream Research Company|Method for seismic facies interpretation using textural analysis and neural networks|
    US9008972B2|2009-07-06|2015-04-14|Exxonmobil Upstream Research Company|Method for seismic interpretation using seismic texture attributes|
    EP2659291B1|2010-12-31|2015-02-18|Foster Findlay Associates Limited|Improvements in 3d object delineation|
    US6226596B1|1999-10-27|2001-05-01|Marathon Oil Company|Method for analyzing and classifying three dimensional seismic information|
    US8688616B2|2010-06-14|2014-04-01|Blue Prism Technologies Pte. Ltd.|High-dimensional data analysis|
    EP3121622B1|2015-07-24|2021-06-16|Bergen Teknologioverføring AS|Method of predicting parameters of a geological formation|
    AU2017343749A1|2016-10-14|2019-04-11|Chevron U.S.A. Inc.|System and method for seismic facies identification using machine learning|
    US10990882B2|2017-07-28|2021-04-27|International Business Machines Corporation|Stratigraphic layer identification from seismic and well data with stratigraphic knowledge base|EP3928131A1|2019-02-20|2021-12-29|Saudi Arabian Oil Company|Method for fast calculation of seismic attributes using artificial intelligence|
    CN111025394A|2019-12-31|2020-04-17|淮南矿业有限责任公司|Depth domain-based seismic data fine fault detection method and device|
    法律状态:
    2019-05-28| PLFP| Fee payment|Year of fee payment: 2 |
    2019-11-22| PLSC| Publication of the preliminary search report|Effective date: 20191122 |
    2020-05-28| PLFP| Fee payment|Year of fee payment: 3 |
    2021-05-26| PLFP| Fee payment|Year of fee payment: 4 |
    优先权:
    申请号 | 申请日 | 专利标题
    FR1854190|2018-05-18|
    FR1854190A|FR3081232B1|2018-05-18|2018-05-18|METHOD FOR DETECTION OF GEOLOGICAL OBJECTS IN A SEISMIC IMAGE|FR1854190A| FR3081232B1|2018-05-18|2018-05-18|METHOD FOR DETECTION OF GEOLOGICAL OBJECTS IN A SEISMIC IMAGE|
    EP19173720.4A| EP3570074A1|2018-05-18|2019-05-10|Method for detecting geological objects in an image|
    CA3043212A| CA3043212A1|2018-05-18|2019-05-13|Method for detecting geological objects in a seismic image|
    US16/412,664| US11226424B2|2018-05-18|2019-05-15|Method for detecting geological objects in a seismic image|
    BR102019010073A| BR102019010073A2|2018-05-18|2019-05-17|process for detecting geological objects in a seismic image|
    [返回顶部]