Created by Tom Schaap.

This blog post is a summarisation of Schaap et al., 2024. Computer vision tools for geophysicists. Presented at the 1st ASEG DISCOVER Symposium 2024: 15 -18 October 2024, Hobart, Australia.

 

Geophysical imagery has long been a cornerstone of mineral exploration, providing vital insights into the Earth’s subsurface. It offers a non-invasive way to explore covered terrains, helping geologists identify promising areas for further investigation. As exploration extends into more challenging environments, the volume and complexity of geophysical data have increased dramatically. With vast datasets now available of various types, such as magnetics, gravity surveys, and satellite imagery, the challenge for geophysicists is how to efficiently process, integrate and interpret this data.

This is where computer vision (CV) can  come into play. CV provides a powerful suite of techniques to automatically analyse and extract meaningful information from imagery. It allows geophysicists to identify hidden insights from large datasets, turning raw images into actionable intelligence. But what exactly is computer vision, and how can it be applied to geophysical data? In this blog, we’ll explore the role of CV in geophysics, the different approaches available, and how they are helping to push the boundaries of exploration.

 

What is Computer Vision?

At its core, computer vision enables machines to interpret and analyse visual data. In geophysics, this typically means analysing complex imagery – such as magnetics and gravity grids or hyperspectral imagery – to detect patterns, structures, and anomalies that are not immediately visible to the human eye.

CV doesn’t always require machine learning. For example, grey-level co-occurrence matrices quantify texture in an image by simply measuring the relationships between adjacent pixels. Haralick features summarise these to capture essential details about the spatial structure of the image that can be described in intuitive terms: contrast, correlation, energy and homogeneity are examples. These are examples of a computer vision tool for turning images into quantified outputs which can be used tofurther explore geophysical data. However, in many cases these are not enough to fully capture the complexities of our imagery, so more powerful tools are required.

Figure 1: Demonstration of Haralick features applied to VRTP magnetics grid of Australia. Each feature describes a different aspect of the texture in the imagery. We can view these spatially (top), and we can also interpret each feature by looking at locations with high and low values of each (bottom).

Moving Beyond Simple Feature Extraction with CNNs and ViTs

Convolutional neural networks (CNNs) are a class of machine learning models that have revolutionised image analysis. They are designed to detect patterns and structures in images by learning from the spatial relationships between pixels. The key aspect of CNNs is the convolutions. Just like a geophysicist applying filters to their imagery, the CNN applies hundreds of convolutional filters to the input data to extract textural information. This gives them the power to recognise geological features such as faults, folds, or mineralized zones, providing a fast and consistent way to analyse vast geophysical datasets. 

Vision transformers (ViTs) are a newer type of model that excels in understanding spatial relationships. Unlike CNNs, which focus on local patterns in the image, ViTs analyse the image as a whole by dividing it into smaller patches and learning the connections between them. This makes them particularly powerful for processing geophysical imagery where both local and global structures are important. 

The Challenge: Lack of Labelled Data in Geophysics

One of the main challenges in applying machine learning to geophysical data is the lack of labelled training data. In many cases, we simply don’t have enough well-defined labels to train a supervised model. For most of the features we’re looking for, we only have a handful of known examples, and we often have no way of knowing where the negative cases are; for example, in a given area we might only have a handful of locations of known porphyry copper deposits, and we don’t necessarily know where they are not.

To overcome this limitation, we can use unsupervised or self-supervised learning (SSL) techniques. SSL allows models to learn from the data itself, without the need for explicit labels. Instead, the model is trained to recognize patterns and generate numeric features, just like Haralick features but in far greater volume and complexity. SSL enables the model to learn robust feature representations, even in the absence of labelled data.

Turning Images into Numbers

With self-supervised learning, we can transform geophysical images into numerical representations that capture their essential characteristics. These numerical data, or feature vectors, can then be used for further analysis, such as clustering, classification, or similarity mapping.

Two powerful SSL techniques that Datarock has adapted to geophysical data are Swapping Assignments between Views (SWAV) and Masked Autoencoders (MAE). SWAV trains a CNN model by creating multiple altered versions of an image (such as cropping or rotating) and ensuring that the features extracted from these variations are similar. This makes the model robust to noise and artefacts in the data, which are common in geophysical surveys.

MAE takes a slightly different approach by masking portions of the image and training a vision transformer model to reconstruct the missing parts. This forces the model to learn a detailed understanding of the spatial relationships within the data. Both SWAV and MAE are ideal for handling the large, unlabeled geophysical datasets often used in mineral exploration. A powerful aspect of our MAE model is its ability to attend to a third dimension, such as the spectral dimension in hyperspectral data, allowing it to learn and extract features that represent both the spatial and spectral contents of imagery with hundreds of bands.

Why SSL Features Are So Powerful

The features generated by SSL models are incredibly versatile. Once a set of feature vectors has been extracted from the geophysical imagery, they can be used in a variety of ways to support exploration workflows.

For example, techniques like Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP) can be used to reduce the dimensionality of the feature space, making it easier to visualise and explore the data. This allows geoscientists to identify patterns and structures that may not have been visible in the raw imagery.

Clustering algorithms, such as the Leiden algorithm, can be applied to the feature vectors to identify regions with similar geological characteristics. This is particularly useful for tasks like unsupervised classification, where the goal is to group similar areas together based on their geophysical properties. When cluster labels are generated, the geoscientist can then visualise images from each cluster and interpret what they represent.

Figure 2: Overview of a clustering workflow on gravity image features extracted in this case from a ViT model. The model is trained on the original imagery and the extracted feature vectors are analysed through dimensionality reduction tools such as UMAP and Leiden clustering. Given the output cluster map, the original imagery captured by each cluster can be visualised.

Finally, similarity mapping allows geophysicists to search for regions that have similar textural features to a known reference point. For example, if a particular area is known to host a mineral deposit, similarity mapping can be used to locate other areas with similar geophysical signatures, providing valuable targets for further exploration.

Data Fusion at Different Scales

One of the most exciting possibilities offered by computer vision in geophysics is the ability to combine data from different sources and scales. For instance, magnetics and gravity datasets can be fused with satellite imagery to create a comprehensive, multi-scale representation of the subsurface. This data fusion enables geoscientists to analyse geological features at multiple levels of detail, providing a more complete picture of the subsurface. In this example, we have a regional magnetics image of the Hamersley Basin WA gridded with 100 m cells, and a Sentinel-2 multispectral image gridded at 10 m. For every cell of the magnetics, there are 100 pixels of Sentinel-2 data. We trained an SSL model to extract features from the Sentinel-2 data in patches that match the resolution of the magnetics grid, allowing us to match the contents of the Sentinel-2 data with the magnetics and directly analyse them together. A cosine similarity map was generated on the combined magnetics and Sentinel-2 dataset, mapping the similarity of these features to a reference location within a known banded iron formation. This shows how computer vision can be used to fuse incongruent datasets and interrogate the results in meaningful ways.

Figure 3: Left; simplified workflow to use SSL features to rescale and combine Sentinel 2 data with magnetics. Right top; Hamersley Basin, WA. 1:100k mapping polygons of banded iron formations (BIFs). Right bottom;  Cosine similarity map of ‘predicted’ BIFs in the Hamersley Basin, based on the fused Sentinel 2 and magnetics data. 

 

Conclusion

Computer vision is transforming the way geoscientists analyse and interpret geophysical imagery. From traditional methods like Haralick features to advanced SSL techniques like SWAV and MAE, CV offers a powerful toolkit for integrating disparate data together and unlocking insights from very large and complex datasets. By turning images into numbers and leveraging unsupervised learning, geoscientists can extract valuable information from their data, even in the absence of labelled training examples.

At Datarock, we’re continually experimenting and developing new approaches to generate insights from new and existing data. As CV technologies continue to evolve, we will continue to integrate them into geophysical exploration workflows, helping to drive smarter, more efficient decision-making.