Student: Davi Bortolotti Batista
Title: Semantic Segmentation for Automatic Interpretation of Linear Geological Structures
Advisors: Diego Barreto Haddad (advisor) and Gabriel Matos Araujo (CEFET/RJ) (co-advisor).
Committee: Diego Barreto Haddad (president), Gabriel Matos Araujo (CEFET/RJ), Douglas de Oliveira Cardoso (CEFET-RJ), Kenji Nose Filho (UFABC), Milena Faria Pinto (CEFET-RJ) and GIlson Antonio Giraldi (LNCC).
Day/Time: December 15, 2021 / 14h.
Room: Link Teams
Automatic interpretation of geological structures may speed up the field work stage necessary in geotechnics, civil engineering and in the exploration of natural resources, such as oil, water and ore, providing geoscientists with a larger volume of data from rock outcrops. Geological parameters enrich and improve predictive capacity of geological numeric and statistic models. The geological fractures are of great interest, for they indicate past and presently ocurring stress regimes in the Earth’s litosphere, beside forming preferential ducts of economically important fluids.
This work proposes a new methodology combining deep learning, semantic segmentation and classic computer vision algorithms for the extraction of linear geological structures from UAV (Unmanned Aerial Vehicles) imagery. Results show Intersection over Union (IoU) metrics of up to 74% for trained model predictions before post-processing, and up to 78% after. The predicted segmentation was used as a mask for binarization and line detection to extract fractures accurately. The comparison of strike directions originated from geological interpretation and from automatically extracted structures exhibited very similar trends and behaviours. The results were then compared with other historically used techniques in the geoprocessing area and with the manually executed annotations from geoscientists, demonstrating larger precision and advantages.