[S-CG50] Driving Solid Earth Science through Machine Learning

[SCG50-P13] On the application of a suite of computer vision algorithms for natural fracture detection in borehole images

Daniil Kuprin², *Nikita Dubinya^1,2 (1.Moscow Institute of Physics and Technology, 2.Schmidt Institute of Physics of the Earth)

Keywords:Image processing, Well logs, Natural fractures, Computer vision

This study explores the application of a suite of computer vision algorithms to detect natural fractures and obtain their spatial orientations in wells using borehole images. As a testbed for the suite, we have used data from the IODP’s 314^th expedition in the region of the Nankai Trough, especially, data from the C-0001D well. The suite employed consists of several widely used image processing and computer vision techniques: binarization and subsequent skeletonization via several different approaches (basic, Otsu’s binarization and triangle binarization); and edge-detection methods (Scharr, Sobel and Canny edge-detection). The suite implemented is employing a sliding window algorithm which goes over the borehole image from top to bottom. Data containing information about fracture geometry is gained and validated via cross-correlation and filtering of each of the computer vision methods outputs from each iteration of the sliding window algorithm. Such data is represented as a sinewave on a borehole image. After the initial ridding of incorrect fracture detections, the data gathered is grouped by depths and spatial orientations of detected fractures. Then statistical analysis is performed on the grouped data to acquire the expected value and dispersion of each dataset, which in turn represent, respectively, the probable location and characteristics of a fracture and the probable error of its identification in that location. The results gathered at this point in development of the suite are mostly reliable with the acquisition of fracture locations but not so with the spatial characteristics of fractures. Our findings indicate the need for employment of a more adequate image preprocessing algorithm since the incorrect identifications of spatial characteristics are most likely associated with the noisiness of the borehole image, or an employment of an image recognition neural network as another mean of validation of correct fracture detection.