Details of the Abstract
| Title of paper | Use of machine learning algorithms as a tool for interpretation of helicopter-borne electromagnetic data |
| List of authors | Cortés Arroyo, O., Schönfeldt, E., Siemon, B., Fahle, M., Janetz, S. |
| Affiliation(s) |
Federal Institute for Geosciences and Natural Resources (BGR) Sub-department Research and Development Centre for Post-Mining Areas (FEZB), Federal Institute for Geosciences and Natural Resources (BGR) Sub-department Research and Development Centre for Post-Mining Areas (FEZB), Federal Institute for Geosciences and Natural Resources (BGR) Sub-department Geophysical Exploration – Technical Mineralogy, Federal Institute for Geosciences and Natural Resources (BGR) Sub-department Research and Development Centre for Post-Mining Areas (FEZB),Federal Institute for Geosciences and Natural Resources (BGR) Sub-department Research and Development Centre for Post-Mining Areas (FEZB). |
| Summary |
To rehabilitate former open-cast, lignite-mining areas, information on aquifers is required. Helicopter-borne electromagnetics (HEM) may provide meaningful results, but processing, analysis and interpretation of such data must be done very carefully. Machine learning (ML) libraries may be useful due to their abilities to extract information from large, high-dimensionality data. Here we present results of two related projects: “D-AERO-Finsterwalde” and “FINA”. “D-AERO-Finsterwalde” consisted of an HEM survey of a post-mining area in the german federal state of Brandenburg. Data was recorded every 4 meters in 6 different frequencies, using a 10-meter-long flight probe, towed at about 50 m above ground. Data was acquired on approximately 1,300 km-length of flight lines, within a survey area of about 250 km². The HEM-Dataset was converted into 1D-resistivity models using a Levenberg-Marquardt procedure. Results show good agreement with local geology and sensitivity to local underground-water and mineralization processes. As part of project “FINA”, we make use of these HEM-results to test three ML-algorithms: K-Means (K-M), Self-Organized Maps (SOM) and Random Forest (RF). K-M and SOM provides an automatic clustering and classification for all the resistivity models, showing the presence of two separate regions in the area, each related to local physical and geological properties. RF algorithm predicts porosity three-dimensionally to gain information about the distribution of aquifers/aquitards. The model showing a 90% accuracy, using 70% of the data for training and 30% for test. Despite the large amount of information, results are obtained in a few minutes using a desktop computer. This encourage us to further explore the application of ML in our HEM processing scheme. |
| Session Keyword | 6.0 Marine and airbone EM |
| File upload |
6.0_use_of_machine_learning_a_cortes_arroyo.pdf
|