Automatic mapping of fractures and faults in optical images and Digital Surface Models with deep learning
Abstract
Faults form dense, complex multi-scale networks over millimeters to thousands of kilometers of length. Those networks generally include a master fault and myriads of smaller-scale faults and fractures that intensely dissect the rocks encasing the master fault. Quantification of the architecture of these complex networks is critical to understand fault and earthquake mechanics, reservoir cap rock integrity, fluid flow pathways in the crust, etc. Commonly, faults are mapped manually from optical images and topographic data, through the recognition of the specific traces they form at the ground or outcrop surfaces. However, manual mapping is time-consuming, which limits our capacity to produce complete representations and measurements of the fault networks. To overcome this problem, we have adopted a machine learning approach to automate the identification and mapping of fractures and faults in optical images and DSMs (Digital Surface Models). Our approach is based on an adaptation of a U-Net Convolutional Neural Network from Tasar et al. (arXiv:1810.12448, 2018) for optical satellite images. In addition to common encoder-decoder blocks, U-net combines low level and high level features via skip connections to increase the semantic segmentation performance. We trained deep learning models with a database (ground truth) of manually created fracture and fault maps extracted from three types of optical images and DSMs: multispectral data from Pleiades satellite (10 km scale, 70 cm resolution), RGB drone photogrammetric data (km scale, cm resolution), and standard camera photogrammetric data (meter scale, mm resolution). Manual fault traces are drawn lines of a pixel size, whereas actual fault traces have a multi-pixel thickness. We thus applied a Gaussian filter to the manual fault traces, and we qualified our confidence in the mapping. We then applied the trained model to four fault zones in USA (Granite Dells, Valley of Fire, Bishop Tuff and Water Pocket), of which some have previously been mapped manually. The fault and fracture networks extracted by the model show good agreement with previous mapping, and the model correctly identifies fractures and faults observed in unmapped areas. The proposed method thus seems a viable tool for fast and accurate extraction of fault and fracture networks from image and DSM data.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.S54A..06M
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1910 Data assimilation;
- integration and fusion;
- INFORMATICS;
- 1914 Data mining;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS