Classificação supervisionada de elementos arquiteturais do modelo 3d de La Jardinera, Bacia de Neuquén, Argentina

Abstract

The manual interpretation of 3D models in geosciences is highly interpreterdependent, time-consuming, non-reproducible, and does not allow the calculation of associated uncertainties. The La Jardinera outcrop, Neuquén Basin, Argentina, exposes sedimentary rocks deposited in turbiditic depositional systems and has been widely studied by several authors. All published works on the sedimentology and stratigraphy of La Jardinera strata were carried out analogically and manually, following traditional methods of faciological field surveys. Within the scope of the project "Virtual Outcrop Models: Construction, Extraction, and Integration of Stratigraphic, Sedimentological, Diagenetic, and Structural Information in 3D", 719 images were acquired in the field with a remotely piloted aerial system for the construction of a 3D model of the La Gardener. The manual interpretation of 3D outcrop models is inherent to the same difficulties and limitations of interpreting 3D models in geosciences. To automate the extraction of stratigraphic information from the 3D model of the La Jardinera outcrop and contribute to the methodological development, a convolutional neural network was used, a supervised classification technique within the scope of artificial intelligence tools (deep learning). The general objective is to obtain the point cloud classified by facies associations of the turbiditic deposits of the La Jardineira outcrop. Of the 719 images captured in the field, 69 were interpreted by facies association. The labeling consists of generating polygons in each image and assigning a class to each polygon, among the two facies associations that occur in the 3D model, lobes-interlobes and lobe fringe. Polygons were generated with guided tools (magic wand or polygonal lasso) available in Photoshop software. After generating the labels, half of the labeled images were used for neural network training and half of the images for the classification accuracy test; the overall accuracy was 85%. The accuracy per architectural element was 69% for interlobe lobes and 74% for lobe fringe. The architecture of the neural network used was U-net, used as the main structure of the segmentation model. After this first test, the neural network was used to classify the 719 images and obtain labels for the entire dataset. Statistical maps of entropy in the images and data validation tests were prepared to visualize the assertiveness of the predictions. The 719 images were processed by the SfM-MVS workflow to generate the point cloud classified by architectural elements. All images were aligned together, and 3-point clouds were constructed: (i) with all images; (ii) only with the parts of the images labeled as lobes-interlobes; and (iii) only with the parts of the images labeled as lobe fringe. From the point cloud with all images, 5.24% of the points were classified as lobe-interlobes and 17.78% as lobe fringe. Obtaining the point cloud classified by facies associations was carried out, and the results obtained in this study demonstrate the effectiveness of using deep learning techniques for interpretation in geosciences.