Short Courses

Course description

Compositional data are vectors that show the relative importance of the parts of a whole. Typical examples are data in mol/liter, percentages, ppm, ppb, or similar, common in many fields of science, particularly in the geosciences. The classical statistical analysis of this type of data suffers from multiple problems, among them the one of spurious correlation. As a solution to these problems, J. Aitchison introduced the logratio approach in the 1980s. Since then, progress has been made in understanding the geometry of the sample space, the simplex of D parts.

Course objectives and learning outcomes

The course aims to introduce attendees to the principles and basic methods of compositional data analysis; how to apply them with Codapack; and how to interpret the results obtained. The course combines theoretical classes with practical data analysis.

Content

• The Aitchison geometry in the simplex and coordinate representation (2 hours).
• Exploratory analysis with CoDaPack: variation matrix, biplot and coordinates (2 hours).
• The CoDa-dendrogram. Orthonormal coordinates (ilr-olr) with CoDaPack.(1.5 hours).
• Compositional statistics. Regression. (1.5 hours).

Free software

• CoDaPack 2.02.04 downloadable from www.compositionaldata.com
• Compositions (in R)
• rob-Compositions (in R)
• zCompositions (in R)
• R Project for Statistical Computing: www.r-project.org

Content

Geochemical data usually consist of either element concentrations or isotope ratios. For this type of data, their inherent properties limit the application of conventional statistical methods. There are two coexisting approaches to analyzing such data: Classical geochemistry has developed individual graphs and normalization rules for each type of problem to minimize common artifacts, but so far lacks general-purpose tools that would find wide application in geochemistry. Modern compositional data analysis, on the other hand, offers comprehensive and general solutions for the mathematically correct analysis of concentration and ratio data, but has so far failed to provide methods for many typical tasks and challenges in geochemistry. This course bridges the gap between the two approaches and provides compositionally coherent statistical methods for geochemistry. The discrepancy is addressed in three ways: - We learn which classical geochemical tools are already compositionally coherent, why they are coherent, and how to incorporate them into a fully compositional data analysis. - We learn how certain compositionally incoherent methods can be replaced by coherent methods for the same questions and how the two classes of methods are related. - We use an interactive class structure where participants are strongly encouraged to bring their typical methods and challenges, and we search for artifacts and construct bridging solutions together.

Who should take the course?

The course is intended for geoscientists who are interested in or have experience with statistical analysis of geochemical data. For individuals who have no prior experience with compositional data analysis, we strongly recommend combining this short course with the pre-conference short course "Introduction to Compositional Data Analysis".

Free software

• R Project for Statistical Computing: www.r-project.org
• Compositions (in R)

Course Content

Both parts are independent of one another, but mutually supportive. Participants can attend either one or both days, each of which will be structured as follows:

• Introduction to the theoretical and mathematical concepts of the methods
• Presentation of geoscientific applications and previous work
• Live demonstrations and hands-on exercises for one or two use cases using different datasets
• Discussions

Part 1: Data analysis using self-organizing maps

This session introduces the GisSOM software, which applies the self-organizing maps (SOM) and k-means methods in performing exploration and clustering of multivariate data. SOM is used to arrange multivariate data onto a grid in such a way, that similar data points are close to one another. It also reduces the amount of data points by generating a large number of protoclusters. Often a 2D grid is used since it is powerful considering the SOM visualization, which is an essential part of the usefulness of the method. SOM is efficient in identifying patterns in multivariate data and helps to understand the information content of the dataset. By reducing the complexity of the dataset, SOM facilitates further analysis of the data. GisSOM is designed considering especially spatial data but can be applied to non-spatial data as

Being an unsupervised machine learning method, SOM can be implemented independent of the availability of labelled data (i.e., training or ground truth data). However, if labelled data is available, its distribution on SOM provides information on the distribution of the different label categories in the multivariate data space. Using SOM for labelled data can, thus, be used for classification of data to known categories.

During the course we shall use different types of geoscientific data to demonstrate the application of SOM and GisSOM:

• Spatial data
• ◦ Analysing regional geochemical till data and airborne geophysical data
• ◦ Analysing hyperspectral drill core data
• Non-spatial data
• ◦ Classifying spectral data of mineral sample preparates

Part 2: Mineral prospectivity analysis using probabilistic weights of evidence method

In this part of the course, we will introduce the principles of GIS-based mineral prospectivity analysis Mineral prospectivity analysis aims at distinguishing areas with a high potential for hosting a mineral deposit from those with a low potential using different statistical and machine-learning based data integration methods. In this course we will demonstrate end-to-end mineral prospectivity mapping using the data-driven weights of evidence (WofE) method in the open-source GIS platform - QGIS.

The WofE is a Bayesian method to estimate the probability of a hypothesis (H) based on the knowledge of occurrence of certain evidential events (E). Applied to mineral prospectivity analysis the hypothesis to be predicted is the probability of existence of the targeted mineral deposit and the evidential events are the geoscientific datasets representing geological features such as lithology, structures, whole rock geochemistry etc. Implementation using geospatial datasets involves quantification of spatial associations (i.e., the weights) between mineral deposits and the geospatial evidential layers and subsequent calculations of the posterior probabilities for potential of existence of a mineral deposit.

During the course the implementation of the WofE model to mapping the potential of mineral deposits will be accomplished using the WofE plugin in QGIS. All the steps involved in the modelling workflow will be demonstrated using a real-world case study and hands-on exercises will be designed for the participants to follow along.

Background information and short articles about tools and methods

• GisSOM for Clustering Multivariate Data ‒ Multivariate Clustering and a Glance to Self-organizing Maps | GTK
• GisSOM for Clustering Multivariate Data – Self-organizing Maps | GTK
• GisSOM for Clustering Multivariate Data ‒ GisSOM | GTK
• Science Blog: GisSOM for Mineral Deposits-related Pattern Recognition and Gold Prospectivity Modelling in the Rajapalot Project Area in Finland | GTK
• GTK Open File Research Report 52/2021. GisSOM for clustering multivariate data
• GitHub - chudasama-bijal/QGIS-Plugin-Weights-of-Evidence-Model: Weights-of-evidence based mineral prospectivity modelling in QGIS

Course concept

Advances in engineering equipment that can deliver massive in-situ data at runtime open the possibility to employ data analysis and data-driven modeling to ensure proactive risk management and enhance process optimization in geoengineering. However, obtained multivariate observational site-specific datasets are often incomplete and potentially corrupted and therefore require special techniques to be applied during the preprocessing step to ensure high-quality results from data analysis and data-driven modeling. The course will elaborate on methods and techniques for data quality checks, preprocessing, integration, and feature engineering.

This course is designed for audiences with a major in civil engineering and/or geosciences interested in data mining and does not require any particular expertise in programming. The course aims to show how the accuracy of data analysis depends on selecting the correct preprocessing strategy and will illustrate the applicability and limitations of main preprocessing steps using relevant examples from geoengineering and geotechnics. The real datasets (e.g., data from cone penetration test or tunnel boring machine) will be used throughout the course, bridging the gap between theoretical concepts and their applications.

Course objectives

• demonstrating the importance of data quality assurance and data preprocessing for the success of data-driven modeling
• explaining the limitations of analytical methods in geo-datasets
• showing how various techniques can be used to overcome the limitations and improve the precision of modeling

Learning outcomes

• review the knowledge discovery, data management, and the data life-cycle concepts
• refresh the concepts of data sparsity and dimensionality
• familiarize themselves with modern representation and formats for data storage
• understand the importance of data quality assessment
• learn the limitations of analytical methods in data-driven modeling when applied to geotechnical data
• get “ready-to-apply” solutions for overcoming analytical methods’ limitations with data preprocessing and data engineering techniques
• understand how the decisions made during data acquisition and processing affect the accuracy of data-driven modeling.

Course content

1. Introduction to Data Science. Why do we care: errors, accuracy, and decision support.
2. Structures and data types, with practical examples from geo-datasets. Data quality, storage, integration, and security. Data labeling and perception in labels.
3. Data formats, precision, and structures. Data sparsity.
4. Statistical and mathematical foundations of Data Science. Data-driven modeling: concepts, methods, applicability, and limitations. Machine learning: concepts, algorithms, limitations. Examples from geoengineering.
5. Data preprocessing and feature engineering. Evaluation and validation metrics in Data Science. Data integration, rebalancing, and dimension reduction. Synthetic data: pros and cons.
6. Sample workflows for quality assurance, preprocessing, and accuracy check for regression and classification tasks in geoengineering. Correlational analysis in preprocessing.

Course objectives and concept

Course objectives: In this course, the participants are guided through steps on how to interpret several surface and downhole geochemistry datasets using a combination of visual exploratory data analysis and data science tools.

While ioGAS is used as a tool for the course, it will not be the prime focus of the course. Prior experience in using ioGAS is recommended, to facilitate the delivery of the course content more effectively. Some prior experience in using Python, especially with the scikit-bio library for compositional data analysis (CoDa) will also be recommended.

The contents that we will cover are the following:

• Compositional data analysis: visual exploration and methods of handling compositional data using Python and ioGAS.
• Modeling the behavior of geochemical, lithological, and alteration data as well as interpretation of an integrated dataset; Combining lab-based geochemistry data with pXRF geochemistry, spectral mineralogy, structural geology, and downhole geophysics data.
• Guided workflow on how multivariate statistics and machine-learning-based data classification processes can be used to characterize a geoscience dataset. The following topics will be covered in the course: K-means clustering, principal components analysis, discrimination projection analysis, self-organizing maps, T-SNE, U-MAP, and Wavelet Tessellation Algorithm.

Learning Outcomes

By the end of the course, the delegates will learn a number of workflows that will allow them to perform the following:

• Context-driven geochemical data interpretation, where interpretation of geochemistry data using EDA and data science tools is driven by the context of the local geology
• Standardized workflows of interpreting a set of geochemistry data, from data import to generating products that can be used for geological model generation or resource estimation model
• Through context-driven workflow, estimation of parameters that can be mission-critical during resource production and mine rehabilitation stage of a mine life cycle

Course description and contents

A course about geostatistical methods, how they relate to geological thinking and modeling, and how tensions between the two can be resolved in practice.
Part 1: The need for geostatistical modeling
We must make decisions under uncertainty. In some contexts, probabilistic estimation and uncertainty quantification based on statistical models is the right tool for the job. What are these contexts? And what are we hoping to get out of our probabilistic modeling efforts?
Part 2: Approaches to geostatistical modeling
Phenomena can be represented in many ways. In geostatistical modeling we often face a choice between several modeling approaches. Different geological settings call for different models. Approaches vary in terms of both theoretical rigor and practical usefulness.
Part 3: Modeling and the real world
There is always a gap between theory and observation. Using the methods of geostatistics on real data means bridging that gap. What do observations from boreholes tell us about the rocks in the subsurface? How do we get this information into our model?
Part 4: Strategies for data conditioning
Making conditioning work in practice requires compromise, but this does not mean giving up theoretical consistency. By trading off approximation error, Monte Carlo noise and computational complexity against each other, we can tailor a solution suited to the conditioning problem at hand.

Who should take this course?

This course is for anyone who is curious about practical considerations involved in reconciling geological modeling needs with geostatistical methods. The list of prerequisites is short. Extensive experience with the subject is not needed to benefit from this course.

Course objectives and learning outcomes

• appreciate when and why geostatistical modeling is needed
• gain an overview of some widely used geostatistical modeling approaches
• understand the centrality of learning from data
• get a valuable perspective on consistent data conditioning