You are Home   »   News   »   View Article

AI based subsurface image classification

Tuesday, October 3, 2023

Kadme has develop a tool to make images classifiable using AI, and then searchable, so that oil and gas companies can do a search of all the images in their electronic documents

We are used to using Google Images to search for photos. Many oil and gas companies would like to be able to search images in their own document store in the same way.

Oil and gas knowledge management company Kadme, based in Stavanger, Norway, has been building tools to do this, based on an algorithm to classify images. It can be a way to quickly search through unstructured data.

'We wanted to build a system that worked like Google Images, but instead of websites, it would have all our messy documents as the source,' said Jesse Lord, Lead, Product Strategy at Kadme, speaking at the Society of Professional Data Managers conference.

The results of the search shows images in an interface similar to Google Images. By clicking on them, you can retrieve the source document.

To do this, the system needs to know which page on the document that image has come from, so it can take the user to the right page. The system does this by first splitting up each document into separate documents, one for each page.

The system has been in development for 18 months, and released earlier in 2022, as part of its search system Lumin (previously known as WhereOil).

Since the trend is ever more unstructured data, systems like this should have increasing value, he says.

System structure

The development of the system starts by running existing detection and classification algorithms over the documents. For this, Lumen uses tools provided by Amazon Web Services.

The search system uses elastic search. You can type in a phrase, type in the type or 'class' of image you are looking for, and then search. A text search can look for the specific text, in the same document.

The system also needs to verify whether the user has permissions to see the specific document.

The documents are stored on a Lumen server, which is accessed via REST APIs.

Image classification

Technologies to classify images are quite mature, including for oil and gas images. For example, it can classify images as a seismic image, seismic images with wells, geosections, and seismic inversions, Mr Lord says.

From a technical perspective, the challenge is to set up an automated pipeline for classifying the images.

Kadme's system, Lumin, is designed for 'speed at scale', capable of searching through systems of over 12 petabytes.

Companies can also train their own model, rather than use Kadme's pre-trained model. 'As few as 100 pictures can produce a 'class' which is useful,' he says.

It does not use 'cloud AI' services from the cloud service providers, which can be quite expensive.

It would be less than a month for an operator to get up and running, allowing for things to go wrong, he says. If the information is all ready to index, it could be done in a week. It may take longer if the documents are on different systems, or if they want to deploy different systems in different regions.

The work involves connecting to the source data, downloading documents, and converting documents into pictures.

The AI model is 'almost the least impressive thing,' he says.

It is essential to have a system for users to provide feedback on whether they found what they were looking for, so the system can improve. 'Whatever model you build - if you don't connect it to feedback, eventually it will become irrelevant,' he said.


As an example, consider that you are trying to search for 'thin sections of ferroan dolomite' (dolomite which has an iron content) in the company's database.

Other examples could be someone looking for an image containing the term 'dunlin' but not 'missed pay'. Someone might want to look at geo-section type images, or a geosection of 'Namurian' (a stratigraphic layer found in Northwest Europe from 326 million years ago).

You might be looking for authigenic quartz (minerals formed, in place, within sediments and sedimentary rocks)

The system will look for documents containing the text you are looking for, and the sort of image you are looking for.

Associated Companies
comments powered by Disqus


To attend our free events, receive our newsletter, and receive the free colour Digital Energy Journal.


Latest Edition Jan-Feb 2024
Jan 2024

Download latest and back issues


Learn more about supporting Digital Energy Journal