The Deep Convolutional Neural Networks As A Geological .

2y ago
22 Views
2 Downloads
4.86 MB
6 Pages
Last View : 1m ago
Last Download : 2m ago
Upload by : Matteo Vollmer
Transcription

The Sedimentary RecordDeep convolutional neuralnetworks as a geological imageclassification toolRafael Pires de Lima1,2, Alicia Bonar1, David Duarte Coronado1, Kurt Marfurt1, Charles Nicholson31School of Geology and Geophysics, The University of Oklahoma, 100 East Boyd Street, RM 710, Norman, Oklahoma, 73019, USA2The Geological Survey of Brazil – CPRM, 55 Rua Costa, São Paulo, São Paulo, Brazil3School of Industrial and Systems Engineering, The University of Oklahoma, 202 West Boyd Street, RM 124, Norman, Oklahoma, 73019, USAKeywordsconvolutional neural networks, transfer learning, automatization, microfossil identification, petrographyABSTRACTA convolutional neural network (CNN) is a deep learning(DL) method that has been widely and successfully applied tocomputer vision tasks including object localization, detection,and image classification. DL for supervised learning tasks is amethod that uses the raw data to determine the classificationfeatures, in contrast to other machine learning (ML)techniques that require pre-selection of the input features(or attributes). In the geosciences, we hypothesize that deeplearning will facilitate the analysis of uninterpreted images thathave been neglected due to a limited number of experts, suchas fossil images, slabbed cores, or petrographic thin sections.We use transfer learning, which employs previously trainedmodels to shorten the development time for subsequentmodels, to address a suite of geologic interpretation tasksthat may benefit from ML. Using two different base models,MobileNet V2 and Inception V3, we illustrate the successfulclassification of microfossils, core images, petrographicphotomicrographs, and rock and mineral hand sample images.ML does not replace the expert geoscientist. The expertdefines the labels (interpretations) needed to train thealgorithm and also monitors the results to address incorrector ambiguous classifications. ML techniques provide a meansto apply the expertise of skilled geoscientists to much largervolumes of data.INTRODUCTIONMachine learning (ML) techniques have beensuccessfully applied, with considerable success, in thegeosciences for almost two decades. Applications of ML bythe geoscientific community include many examples suchas seismic-facies classification (Meldahl et al., 2001; Westet al., 2002; de Matos et al., 2011; Roy et al., 2014; Qi etal., 2016; Hu et al., 2017; Zhao et al., 2017), electrofacies4 June 2019classification (Allen and Pranter, 2016), and analysis ofseismicity (Kortström et al., 2016; DeVries et al., 2018;Perol et al., 2018; Sinha et al., 2018), and classificationof volcanic ash (Shoji et al., 2018), among others.Conventionally, ML applications rely on a set of attributes(or features) selected or designed by an expert. Featuresare specific characteristics of an object that can be usedto study patterns or predict outcomes. In classificationmodeling, these features are chosen with the goal ofdistinguishing one object from another.Typically, feature selection is problem dependent.For example, a clastic sedimentary rock is most broadlyclassified by its grain size; therefore a general classificationfor a rock sample (data) is sandstone if its grain sizes(features) lie from 0.06 mm to 2.0 mm following theWentworth size class. In this example, a single featureis used to classify the sample, but more complex and/ordetailed classification often requires analysis of multiplefeatures exhibited by the sample. An inefficiency oftraditional ML approaches is that many features maybe constructed while only a subset of them are actuallyneeded for the classification.The use of explicitly designed features to classifydata was the traditional approach in ML applicationswithin the geosciences as in many other research areas.This classification approach works well when humaninterpreters know and can quantify the features thatdistinguish one object from another. However, sometimesan interpreter will subconsciously classify featuresand have difficulty describing what the distinguishingfeatures might be, relying on “I’ll know what the objectis when I see it”. In contrast to feature-driven MLclassification algorithms, deep learning (DL) modelsextract information directly from the raw unstructureddata rather than the data being manually transformed.

The Sedimentary RecordFigure 1: Examples of the data used in this study. A) Three of the seven Fusulinids groups (Beedeina (1), Fusulinella (2), and Parafusulina (3)).B) Three of the five lithofacies (bioturbated mudstone-wackestone (1), chert breccia (2), and shale (3)). C) Reservoir quality classes (high (1),intermediate (2), and low (3)) D) Three of the six rock sample groups (basalt (1), garnet schist (2), and granite (3)). Samples were interpreted byprofessionals working with each separate dataset.Because of their greater complexity(and resulting flexibility and power)convolutional neural networks(CNN) usually requires more trainingdata than traditional ML processes.However, when expert-labeled dataare provided, non-experts can usethe CNN models to generate highlyaccurate results (e.g. TGS SaltIdentification Challenge Kaggle,2019).DL applications in the geosciencesrequire experts to first definethe labels used to construct thenecessary data sets as well asidentify and address any ambiguousresults and anomalies. In orderto bring awareness and providebasic information regarding CNNmodels, DL techniques, and thenecessity of expert-level knowledgeJune 2019 5

The Sedimentary Recordneeded to utilize these advancements,we applied these methods to fourdifferent geologic tasks. Figure 1 showssamples of different types of data thatcan be interpreted and labeled byexperienced geologists. We use suchinterpretations to train our models. Inthis manuscript, we show how CNNcan aid geoscientists with microfossilidentification, core descriptions,petrographic analyses, and as apotential tool for education andoutreach by creating a simple handspecimen identification application.CONVOLUTIONALNEURAL NETWORKS ANDTRANSFER LEARNINGRecent CNN research hasyielded significant improvementsand unprecedented accuracy (theratio between correct classificationsand the total number of samplesclassified) in image classificationand are recognized as leadingmethods for large-scale visualrecognition problems, such as theannual ImageNet Large Scale VisualRecognition Challenge (ILSVRC,Russakovsky et al. (2015)). SpecificCNN architectures have been theleading approach for several years now(e.g., Szegedy et al., 2014; Chollet,2016; He et al., 2016; Huang etal., 2016; Sandler et al., 2018).Researchers noted that the parameterslearned by the layers in many CNNmodels trained on images exhibit acommon behavior – layers closer tothe input data tend to learn generalfeatures, such as edge detecting/enhancing filters or color blobs,then there is a transition to morespecific dataset features, such as faces,feathers, or object parts (Yosinski etal., 2014; Yin et al., 2017). Thesegeneral-specific CNN layer propertiesare important points to be consideredfor the implementation of transferlearning (Caruana, 1995; Bengio,2012; Yosinski et al., 2014). Intransfer learning, first a CNN modelis trained on a base dataset for aspecific task. The learned features(model parameters) are repurposed, ortransferred, to a second target CNNto be trained on a different datasetand task (Yosinski et al., 2014).New DL applications often requirelarge volumes of data, however thecombination of CNNs and transferlearning allows the reuse of existingDL models to novel classificationproblems with limited data, as hasbeen demonstrated in diverse fields,such as botany (Carranza-Rojas et al.,2017), cancer classification (Estevaet al., 2017), and aircraft detection(Chen et al., 2018). Analyzingmedical image data, Tajbakhshet al. (2016) and Qayyum et al.(2017) found that transfer learningachieved comparable or better resultsthan training a CNN model withrandomly initialized parameters.As an example, training the entireInceptionV3 (Szegedy et al., 2015)with 1000 images (five classes, 50original images for each class, fourcopies of each original image) withrandomly initialized parameters canbe 10 times slower than the transferlearning process (11 minutes vs 1minute on average for five executions)using a Nvidia Quadro M2000 (768CUDA Cores). On a CPU (3.60GHz clock speed), training the entiremodel can take up to 2 hours whereastransfer learning can be completedwithin a few minutes. We also noticedthat transfer learning is easier totrain. During the speed comparisontest, transfer learning achieved highaccuracies (close to 1.0) within 5epochs (note the dataset is very simplewith most of the samples being copiesof each other). Successful applicationsof computer vision technologiesin different fields suggest that MLmodels could be extremely beneficialfor geologic applications, especiallythose in the category of imageclassification problems.For the examples we presentin this paper (Figure 1), we relyon the use of transfer learning(Yosinski et al., 2014) using theMobileNetV2 (Sandler et al.,2018) and InceptionV3 as our baseCNN models. Both MobileNetV2and InceptionV3 were trained onILSVRC. Therefore, the CNNmodels we used were constructedbased on inputs of 3-channels (RGB)of 2D photographic images. Werandomly select part of the data tobe used as a test set maintainingthe same proportion of samples perclass as in the training set. The datain the test set is not used during theTable 1: Summary of test accuracy for the examples in this hicthin-sectionsRock samples6 June 2019Number oftraining samplesNumber oftest samplesNumber ofoutput 80.97

The Sedimentary RecordFigure 2: An example of the classification process. In this example, a thin-section image that should fit one of the seven fusulinid genera is analyzedby the model. The model outputs the probability assigned to each of the possible classes (all probabilities summing to 1.0). The term “classes” here isused in the ML sense rather than the biological one. In the example provided, our model provided a high probability for the same class as the humanexpert. Note that in the implementation we use the model will classify any image as one of the seven learned classes – even if the image is clearly not afossil. This highlights the importance of a domain expert intervention.computational process for modeltraining; rather, it is used to evaluatethe quality and robustness of thefinal model. Due to limited space, werefrained showing the CNN mistakesand many of the steps necessary fordata preparation.CNN-ASSISTED FOSSILANALYSISBiostratigraphy has become aless common focus of study in thediscipline of paleontology (Farleyand Armentrout, 2000, 2002), butthe applications of biostratigraphyare necessary for understandingage-constraints for rocks that cannotbe radiometrically dated. Accessto a specific taxonomic expert toaccurately analyze fossils at thespecies-level can be as challengingas data acquisition and preparation.Using labeled data from theUniversity of Oklahoma Sam NobleMuseum and iDigBio portal, wefound that fusulinids (index fossils forthe Late Paleozoic) can be accuratelyclassified with the use of transferlearning. Accurate identification of afusulinid depends on characteristicsthat must be observed and exposedalong the long axis of the (prolatespheroid-shaped) fusulinid. We useda dataset of 1850 qualified imagesincluding seven different fusulinidgenera. After retraining the CNNmodel, we obtained an accuracy forthe test set (10% of the data) of 1.0for both retrained MobileNetV2and InceptionV3 (Table 1). Figure2 shows a schematic view of theclassification process.CNN-ASSISTED COREDESCRIPTIONMiles of drilled cores are stored inboxes in enormous warehouses, manyof which have either been neglectedfor years or never digitally described.Core-based rock-type descriptionsare important for understanding thelithology and structure of subsurfacegeology. Using several hundred feetof labeled core from a Mississippianlimestone in Oklahoma (data fromSuriamin and Pranter, 2018 and Piresde Lima et al., 2019), we selected asmall sample of 285 images from fivedistinct lithofacies to be classified bythe retrained CNN models. Pires deLima et al. (2019) describes how asliding window is used to generateCNN input data, cropping smallsections from a standard core image.We used 10% of the data as the testset and achieved an accuracy of 1.0using the retrained MobileNetV2and an accuracy of 0.97 using theretrained InceptionV3 (Table 1).CNN-ASSISTED RESERVOIRQUALITY CLASSIFICATIONUSING PETROGRAPHICTHIN SECTIONSPetrography focuses on themicroscopic description andclassification of rocks and is one ofthe most important techniques insedimentary and diagenetic studies.Potential information gained fromthin section analysis compared tohand specimen descriptions includemineral distribution and percentage,pore space analysis, and cementcomposition. Petrographic analysescan be laborious even for experiencedgeologists. Using a total of 161photomicrographs of parallel Nicolpolarization of thin sections from theSycamore Formation shale resourceplay in Oklahoma, we classifiedthese images as representatives ofhigh, intermediate, and low reservoirquality depending on the percent ofcalcite cement and pore space. Weused 20% of the images in the testset and obtained a test set accuracyof 0.81 for both the retrainedJune 2019 7

The Sedimentary RecordMobileNetV2 and the retrainedInceptionV3 (Table 1).CNN-ASSISTED ROCKSAMPLE ANALYSISBy creating a simple website,the general population couldhave immediate access to a rockidentification tool using transferlearning technology. For this workin progress, we used smartphones toacquire 1521 pictures of six differentrock types, using five different handsamples for each one of the rocktypes. We took pictures with differentbackgrounds, as visually depicted inFigure 1, however all pictures weretaken in the same classroom. Afterretraining the CNN models, weobtained an accuracy for the test set(10% of original data) of 0.98 usingthe retrained MobileNetV2 and 0.97using the retrained InceptionV3(Table 1). We note that our modeldoes not perform well with nobackground images (i.e., pictures inwhich the rock sample is edited andseems to be within a white or blackcanvas) as such images were not usedin training.CONCLUSIONS ANDFUTURE WORKAlthough gaining popularity andbecoming established as robusttechnologies in other scientificfields, transfer learning and CNNmodels are still novel with respectto application within the geosciencecommunity. In this paper, we usedCNN and transfer learning to addressfour potential applications thatcould improve data management,organization, and interpretation indifferent segments of our community.We predict that the versatiletransfer learning and deep learningtechnologies will play a role in publiceducation and community outreach,allowing the public to identify rocksamples much as they currently can8 June 2019use smart phone apps to identifyvisitors to their bird feeder. Suchpublic engagement will increasegeological awareness and providelearning opportunities for elementaryschools, outdoor organizations, andfamilies.For all of our examples, we wereable to achieve high levels of accuracy(greater than 0.81) by repurposingtwo different CNN models originallyassembled for generic computer visiontasks. We note that the examples andapplications demonstrated here arecurated, and therefore we expectedhighly accurate results. We presenteddemonstrations with limited classesand relatively well-controlled inputimages, so near perfect accuraciescannot necessarily be expected inan open, free-range deploymentscenario. Regardless, the ability tocreate distinctive models for specificsets of images allows for a versatileapplication.The techniques we have showncould greatly improve the speed ofmonotonous tasks such as describingmiles of core data with very similarcharacteristics or looking at hundredsof thin sections from the samegeologic formation. While the tasksare performed by the computer, thegeoscience expert is still the mostimportant element in every analysis inorder to create the necessary datasetsand provide quality control of thegenerated results. In the end, theexpert validates the correctness ofthe results and looks for anomaliesthat are poorly represented bythe target classes. We believe MLcan help maintain consistency ininterpretations and even providea resource for less commonobservations and data variations,such as previously overlooked fossilsubspecies and unique mineralogicalassemblages in small communitiesand private collections, therebybuilding and reconciling a morecomplete international database. Bycombing expert knowledge and timeefficient technology, ML methodscan accelerate many data analysisprocesses for geologic research.ACKNOWLEDGEMENTSWe thank the iDigBio initiative forproviding access to the communityfor biodiversity collections data.Rafael acknowledges CNPq (grant203589/2014-9) for the financialsupport and CPRM for granting theleave of absence allowing the pursuitof his Ph.D. studies. We thank RogerJ. Burkhalter from the Universityof Oklahoma Sam Noble Museumof Natural History for providingthe Fusulinids images used in thismanuscript.REFERENCESALLEN, D. B., AND M. J. PRANTER, 2016,Geologically constrained electrofaciesclassification of fluvial deposits: An examplefrom the Cretaceous Mesaverde Group, Uintaand Piceance Basins: AAPG Bulletin, v. 100,no. 12, p. 1775–1801,doi:10.1306/05131614229.BENGIO, Y., 2012, Deep Learning ofRepresentations for Unsupervised and TransferLearning, in I. Guyon, G. Dror, V. Lemaire,G. Taylor, and D. Silver, eds., Proceedings ofICML Workshop on Unsupervised and TransferLearning: PMLR, Proceedings of MachineLearning Research, p. 17–36.CARRANZA-ROJAS, J., H. GOEAU, P.BONNET, E. MATA-MONTERO, ANDA. JOLY, 2017, Going deeper in the automatedidentification of Herbarium specimens: BMCEvolutionary Biology, v. 17, no. 1, p. 181,doi:10.1186/s12862-017-1014-z.CARUANA, R., 1995, Learning Many RelatedTasks at the Same Time with Backpropagation,in G. Tesauro, D. S. Touretzky, and T. K. Leen,eds., Advances in Neural Information ProcessingSystems 7: MIT Press, p. 657–664.CHEN, Z., T. ZHANG, C. OUYANG, Z.CHEN, T. ZHANG, AND C. OUYANG, 2018,End-to-End Airplane Detection Using TransferLearning in Remote Sensing Images: RemoteSensing, v. 10, no. 1, p. 139,doi:10.3390/rs10010139.CHOLLET, F., 2016, Xception: Deep Learningwith Depthwise Separable Convolutions:CoRR, v. abs/1610.0.DEVRIES, P. M. R., F. VIÉGAS, M.WATTENBERG, AND B. J. MEADE, 2018,

The Sedimentary RecordDeep learning of aftershock patterns followinglarge earthquakes: Nature, v. 560, no. 7720,p. 632–634, doi:10.1038/s41586-018-0438-y.ESTEVA, A., B. KUPREL, R. A. NOVOA,J. KO, S. M. SWETTER, H. M. BLAU, ANDS. THRUN, 2017, Dermatologist-levelclassification of skin cancer with deep neuralnetworks: Nature, v. 542, no. 7639, p. 115–118,doi:10.1038/nature21056.FARLEY, M. B., AND J. M. ARMENTROUT,2000, Fossils in the Oil Patch: Geotimes,p. 14–17.FARLEY, M. B., AND J. M. ARMENTROUT,2002, Tools, Biostratigraphy becoming lost artin rush to find new exploration: Offshore,p. 94–95.HE, K., X. ZHANG, S. REN, AND J. SUN,2016, Deep Residual Learning for ImageRecognition, in 2016 IEEE Conference onComputer Vision and Pattern Recognition(CVPR): IEEE, p. 770–778,doi:10.1109/CVPR.2016.90.HU, S., W. ZHAO, Z. XU, H. ZENG, Q. FU,L. JIANG, S. SHI, Z. WANG, AND W. LIU,2017, Applying principal component analysisto seismic attributes for interpretation ofevaporite facies: Lower Triassic JialingjiangFormation, Sichuan Basin, China:Interpretation, v. 5, no. 4, p. T461–T475,doi:10.1190/INT-2017-0004.1.HUANG, G., Z. LIU, AND K. Q.WEINBERGER, 2016, Densely ConnectedConvolutional Networks: CoRR, v. abs/1608

convolutional neural networks, transfer learning, automatization, microfossil identification, petrography ABSTRACT A convolutional neural network (CNN) is a deep learning (DL) method that has been widely and successfully applied to computer vision tasks including object localization, detection, and image classification.

Related Documents:

May 02, 2018 · D. Program Evaluation ͟The organization has provided a description of the framework for how each program will be evaluated. The framework should include all the elements below: ͟The evaluation methods are cost-effective for the organization ͟Quantitative and qualitative data is being collected (at Basics tier, data collection must have begun)

Silat is a combative art of self-defense and survival rooted from Matay archipelago. It was traced at thé early of Langkasuka Kingdom (2nd century CE) till thé reign of Melaka (Malaysia) Sultanate era (13th century). Silat has now evolved to become part of social culture and tradition with thé appearance of a fine physical and spiritual .

Learning a Deep Convolutional Network for Image Super-Resolution . a deep convolutional neural network (CNN) [15] that takes the low- . Convolutional Neural Networks. Convolutional neural networks (CNN) date back decades [15] and have recently shown an explosive popularity par-

On an exceptional basis, Member States may request UNESCO to provide thé candidates with access to thé platform so they can complète thé form by themselves. Thèse requests must be addressed to esd rize unesco. or by 15 A ril 2021 UNESCO will provide thé nomineewith accessto thé platform via their émail address.

̶The leading indicator of employee engagement is based on the quality of the relationship between employee and supervisor Empower your managers! ̶Help them understand the impact on the organization ̶Share important changes, plan options, tasks, and deadlines ̶Provide key messages and talking points ̶Prepare them to answer employee questions

Dr. Sunita Bharatwal** Dr. Pawan Garga*** Abstract Customer satisfaction is derived from thè functionalities and values, a product or Service can provide. The current study aims to segregate thè dimensions of ordine Service quality and gather insights on its impact on web shopping. The trends of purchases have

Deep Neural Networks Convolutional Neural Networks (CNNs) Convolutional Neural Networks (CNN, ConvNet, DCN) CNN a multi‐layer neural network with – Local connectivity: Neurons in a layer are only connected to a small region of the layer before it – Share weight parameters across spatial positions:

Video Super-Resolution With Convolutional Neural Networks Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K. Katsaggelos, Fellow, IEEE Abstract—Convolutional neural networks (CNN) are a special type of deep neural networks (DNN). They have so far been suc-cessfully applied to image super-resolution (SR) as well as other image .