AI-BOOSTED GEOLOGICAL FACIES ANALYSIS FROM HIGHRESOLUTION BOREHOLE IMAGES

摘要

Borehole images provide many different texture featuresfor facies analysis and natural fracture identification.However, classification of most of these features isachieved manually. The workflow proposed here is toimplement geological “facial recognition” from boreholeimages and other petrophysical measurements. Theimage segmentation is the first step to split the geological“facies” from continuous borehole images. Then theclustering based on texture similarity and petrophysicalmeasurements is the second step to major faciescategories. The major facies categories are labeledmanually, and a deep learning model is trained torecognize geological facies on new borehole images inthe same reservoir.A borehole image can be visually recognized as acomposition of successive zones; different zones havedifferent statistical properties, which can be used tocharacterize the image and generate the zonation. Thecontinuous histogram and variogram derived from imagedata are used for image segmentation. From the highresolutionborehole images, the segments obtained arenumerous enough to perform what is known asunsupervised classification. Among various methods ofunsupervised classification, we choose to use the meanshift algorithm for the automatic clustering. It is adeterministic process, which is suitable in determiningthe number of clusters. The segments are assigned asfacies with a local geological setting, then structured andformatted to build a library of multimodal data (imagedata and petrophysical log data) for a given facies. Adeep learning model is trained to associate multimodaldata to a given facies. This model is used to identifyautomatically the image features of another borehole, forcontinuous facies analysis in similar depositionalenvironments.We demonstrated this workflow in different depositionalenvironments. Twenty-four facies were recognized froma water-based mud image in a braided river environment from China compared with 14 with core descriptionTwelve facies identified from an oil-based mud image ina lacustrine system from the United States were thenapplied to another water-based mud image in the samereservoir with the deep learning model. The results fromthis approach were verified after comparing with amanual interpretation from cores.