GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 3.1 GNGTS 2023 In this work, an experiment of application of a DL method, hereinafter referred to as Neuropick, on LWD density image logs is discussed for segmentation purposes. In particular, we compare different training strategies of a convolutional neural network architecture, known as U-Net, to compare the results in terms of quality and prediction time: the Curriculum Learning (CL) and the single step strategy. U-Net is a convolutional network architecture in which upsampling layers are added to the normal downsampling path, replacing pooling layers. For localization purposes, high-resolution features from the contraction path are combined with upsampling outputs in the expansive path, through concatenation. Compared to generic convolutional networks, which are known to be effective in image classification, segmentation via U-Net additionally offers the possibility of satisfying the need for localization of a target within the image, as it performs pixel-wise classification. The training was possible after generating representative synthetic data as a very high number of examples is required to train a neural network to perform such a task, where extreme variability of real data is assumed, in order to obtain good generalization of the model. From a purely geological point of view, a model can be considered well generalized when it is applicable and yields good results in a large number of different geological situations, which is not an easy task, as geological situations are highly variable from case to case or may differ substantially from apparently comparable situations. The segmentation results are discussed in this paper for comparison purposes. Future goals include using the U-Net segmentation results to train a further DL-based model to extrapolate planar approximations of discontinuities, as is standard in the interpretation of this type of data. Data&Methods For the purposes of this paper, the training data used correspond to 1’000’000 synthetic images logs composed of 16 sampling channels and 20 depth samples. These data are generated with a Python computer program. The corresponding target segmentation maps were generated as 2D binary arrays, with ones located at the sharpest change in numerical series simulating density series of drilled formation, and zeros elsewhere. In the synthetic data, density contrasts along each channel simulating a measuring LWD tool sector were simulated as sinusoidal curves (plus various pre- and post-convolution noise thresholds) with a single phase, as the horizontal dimension of the image represents the entire 360° of the borehole wall. These sinusoids serve, in turn, as planar approximations of geological discontinuities. This approximation is necessary in order to subsequently estimate the orientation of the geological planes in space in terms of azimuth and dip direction, by means of the characteristics of the sinusoid itself (minimum coordinates, amplitude). As we are operating in a supervised classification framework, the methodology consists of a Deep Neural Network architecture by the name of U-Net, known for its effectiveness in image segmentation (Ronneberger et al., 2015), trained to reconstruct the main features of synthetic LWD images to trace the distribution of pixels belonging to a measured parameter discontinuity. The operating principle of such a network is essentially based on the downscaling and upscaling of the input images, through the convolution that occurs at each layer of the