When using airborne geophysical measurements in e.g. groundwater mapping, an overwhelming amount of data is collected. Increasingly larger survey areas, denser data collection and limited resources, combines to an increasing problem of building geological models that use all the available data in a manner that is consistent with the geologists knowledge about the geology of the survey area.
In the ERGO project, funded by The Danish National Advanced Technology Foundation, we address this problem, by developing new, usable tools, enabling the geologist utilize her geological knowledge directly in the interpretation of the AEM data, and thereby handle the large amount of data,
In the project we have developed the mathematical basis for capturing geological expertise in a statistical model. Based on this, we have implemented new algorithms that have been operationalized and embedded in user friendly software. In this software, the machine learning algorithm, Smart Interpretation, enables the geologist to use the system as an assistant in the geological modelling process. As the software ‘learns’ the geology from the geologist, the system suggest from new features in the data.
In this presentation we demonstrate the application of the results from the ERGO project, including the proposed modelling workflow utilized on a data example from Gotland, Sweden.
Botany krishna series 2nd semester Only Mcq type questions
Fast modelling of Airborne EM data using "Smart Interpretation"
1. …by I•GIS
Presented at the 2015 AGU meeting in San Fransisco
Smart Interpretation
– application of machine learning in geological interpretation of AEM data
Torben Bach 1, Rikke Jakobsen1, Tom Martlev Pallesen1, Mats Lundh Gulbrandsen2, Thomas Mejer Hansen2, Anne-
Sophie Høyer3, Flemming Jørgensen3
1. GeoScene3D Team, I-GIS, Risskov, Denmark
2. Niels-Bohr Institute, Computational Geoscience, University of Copenhagen, Denmark
3. Geological Survey of Denmark and Greenland (GEUS), Denmark
The ERGO project: Effective High-Resolution Geological Modeling
2. …by I•GIS
Outline
Presentation outline
• Motivation behind and Introduction to “Smart Interpretation”
• Workflow when modelling with “Smart Interpretation”
• Case Example, Gotland, Sweden
• Summary and outlook
Introduction Workflow Test Case Summing Up
3. …by I•GIS
Motivation
Motivation for Smart Interpretation (SI)
• Observations:
• Large AEM surveys - enormous amount of data.
• One the one hand - manual interpretation is time consuming
• On the other hand - geophysical resistivity is not necessarily linked to geological formation or
lithology
• A Geological expert is needed.
• Inspiration: Seismic Auto-picker, used daily as a standard part of modelling of seismic data in O&G
• Goal: Develop a practical and usable tool for assisting the Geologist
Introduction Workflow Test Case Summing Up
Autumn Spring
20 50
ohmm
Sand and Clay have overlapping resistivitiesSeasonal variation is reflected in resistivities
4. …by I•GIS
SI - Theory
Steps
• Infer a statistical model h(d|M)
• Solve the problem: d = f (M).
• Perform predictions dpred with uncertainty
Mpred
dpred
f(M)
h(dpred|Mpred)
+/- 1 std.
M
d
Our Toolbox
• Standard Gaussian based inversion theory – with a twist…**
Benefits compared to other Machine Learning techniques:
• Tools for analysing parametric covariances and interdependencies
• A measure of uncertainty on the estimates
• Very fast !
**See ”Smart Interpretation - Automatic geological interpretations based on supervised statistical models” by
Gulbrandsen, Cordua , Bach and Hansen, currently subitted and in review for ”Computational Geosciences”
Introduction Workflow Test Case Summing Up
5. …by I•GIS
SI - Theory
M
Geophysical Data
(M)
Introduction Workflow Test Case Summing Up
6. …by I•GIS
SI - Theory
M d
Geophysical Data
(M)
Geological
Knowledge (d)
Introduction Workflow Test Case Summing Up
7. …by I•GIS
SI - Theory
M d
h(d,M)
Geophysical Data
(M)
Statistical Model
h(d,M)
Geological
Knowledge (d)
Introduction Workflow Test Case Summing Up
8. …by I•GIS
SI - Theory
M d
h(d,M)
Mpred
dpred
Geophysical Data
(M)
Statistical Model
h(d,M)
Geophysical Data
Elsewhere
Mpred
Predicted Geology
with uncertainty
h(dpred|Mpred)
Geological
Knowledge (d)
Introduction Workflow Test Case Summing Up
10. …by I•GIS
Groundwater mapping on the Island of Gotland
Courtesy Peter Dahlquist, SGU
Test Case
Introduction Workflow Test Case Summing Up
11. …by I•GIS
Test Setup
Introduction Workflow Test Case Summing Up
The Geologists
• Geologist 1: Using normal manual modelling
• Geologist 2: Using SI assisted manual modelling
Limestone
Marlstone
Clay- and marlstone
The Geology
Sharp boundary
Diffuse Zone
The Test
• Compare ”Manual Model” to ”Model generated using 10% as input to SI”
• Compare ”Manual Model” to ”SI assisted Model”
13. …by I•GIS
Test: Manual Model
Introduction Workflow Test Case Summing Up
Surface 2Surface 1
Geologist 1 – a standard manual model
• Evenly distributed mesh of manual interpretation points
• Surfaces dipping trend towards the south-east
• Abrupt high in north-west
14. …by I•GIS
Test: Manual Model
Introduction Workflow Test Case Summing Up
The Geologist avoids couplings and artifacts in data
Difuse Zone
Interpreted
The Geologist models the ”pinch out” of the ”diffuse” layer
Geologist 1 – a standard manual model
16. …by I•GIS
Test: SI using 10% of Manual Model
Introduction Workflow Test Case Summing Up
• General trend in surfaces is reproduced
• Higher small scale variation due to the increased amount of interpretation points
Surface 2Surface 1
ManualManual
MANUAL
17. …by I•GIS
Test: SI using 10% of Manual Model
Introduction Workflow Test Case Summing Up
• General trend in surfaces is reproduced
• Higher small scale variation due to the increased amount of interpretation points
Surface 2Surface 1
Manual10% of manual points,
1688 SI points generated
Manual10% of manual points,
1653 SI points generated
Smart Interpretation
18. …by I•GIS
Test: Reduced Model 10%
Introduction Workflow Test Case Summing Up
Geologist 1 Remove 90% of interpretation points – and run SI
10% Manual + SI
26 man.points, 1653 SI.points
Difference
Surface 1
264 points
343 points
Surface 2
Manual Model
+/- 10 m
26 man.points, 1688 SI.points
19. …by I•GIS
Manual
Test: SI using 10% of Manual Model
Introduction Workflow Test Case Summing Up
Manual
MANUAL
20. …by I•GIS
Manual
Test: SI using 10% of Manual Model
Introduction Workflow Test Case Summing Up
10% of manual points
Manual10% of manual points
Couplings only partly managed
Difuse Zone
Is managed
Pinch Out is managed
Smart Interpretation
22. …by I•GIS
Test: SI Assisted Model
Introduction Workflow Test Case Summing Up
• General trend in surfaces is reproduced
• Higher small scale variation due to the increased amount of interpretation points
Surface 2Surface 1
Manual Model Manual Model
MANUAL
23. …by I•GIS
Test: SI Assisted Model
Introduction Workflow Test Case Summing Up
• General trend in surfaces is reproduced
• Higher small scale variation due to the increased amount of interpretation points
Surface 2Surface 1
Manual Model Manual ModelSI Assisted Model SI Assisted Model
Smart Interpretation
25. …by I•GIS
Manual
Manual
Test: SI Assisted Model
Introduction Workflow Test Case Summing Up
SI Assisted Model
SI Assisted Model
Couplings are managed
Difuse Zone
Is managed
Pinch Out is managed
Smart Interpretation
26. …by I•GIS
Test: SI Assisted Model
Introduction Workflow Test Case Summing Up
Summary
• The theoretical framework derived from Gaussian based inversion techniques
• It is very fast
• calculation uncertainty
• Test case shows ability to map couplings and diffuse geological boundaries
• More interpretation points -> more variation in the generated surfaces
• Implemented in production software GeoScene3D
Looking ahead…
• Currently underway
• developments toward looking for “structures” in data
• other attribute types, e.g. coherency
• other datatypes included in SI
Come and join us
An enormous amount of information are available to the geologist when modeling large AEM surveys.
One the one hand - manual interpretation is both time consuming and prone to errors, when incorporate all information.
On the other hand - geophysical resistivity is not necessarily directly linked to geological formation or lithology – a Geological expert is needed.
We will develop practical tools that assist the Geologist in the modelling procedure, enabling geophysical results are used in a manner consistent with the Geological expert knowledge provided to the system.
The procedure is itterative and on this slide we can follow the workflow.
First the geologist add the manual interpretation
Second the algoritm is traind in a local area until the geolgist i satisfied
Third, the algorithm is applied on the whole dataset
And then the geologist evaluate and makes a quality controll of the results.
If the results look perculiar in some way or area, the geologist can go back and make another interpretation or change the area where the alorithm works well.
If everything looks fine the geologist can go on with another area or geological layer
The case study area is situated on an Island a few km outside of the Swedish mainland
The geology can be simplified as shown in this illustration, a none existing to thin soil cover on flatlying carbonates and marlstone
The main aquifers for drinking water are situated in fractured and karst aquifers.