Key take-aways

  • Direct engagement with DCMR Environmental Service Rijnmond

  • Created and trained custom machine learning models for large visual data sets 

  • Enabled rapid, reliable object detection in aerial photography data

Established challenges demand new solutions

Recently, I was part of a team tasked with finding an innovative, data-driven solution to this challenge. With manual detection proving extremely challenging, DCMR Environmental Service Rijnmond asked us to help them find a more efficient solution, powered by data science. In this blog I’ll walk you through our approach, and explain how novel machine learning models helped us solve a problem that previously appeared impossible to overcome without significant manual effort.

Defining our approach

Data science is a useful tool when you’ve got very specific questions to ask, and the answers are hidden in large data sets. In this case, our question was simple: “Where are the unregistered wet cooling towers in Rijnmond?”

We quickly realised that there was only one data set that would contain all of the answers to that question –overhead satellite photography of the region.

We obtained complete overhead visual data for all relevant areas from PDOK, knowing our answers –the locations of the cooling towers – would be in there somewhere. The next step was more challenging: training a machine learning model to spot wet cooling towers in an overhead photograph.

Testing and refining our model

With more visual data available than ever before, object detection and image analysis are both evolving rapidly today, with new methods, models and approaches constantly emerging. With no single package able to deliver everything we needed, we opted to use our own computer vision package, Envision, to develop our model.

Envision enabled us to train and use object detection to make accurate predictions of where unregistered wet cooling towers are. Plus, because it’s our package, we were free to tweak details and capabilities in the model to ensure results were as strong as possible.

During this project, that meant finding new ways to lay aerial images over one another and ‘tile’ them to ensure that no potential towers were lost on the edges of images.

Delivering the right outputs

With the model trained, it was now able to return predictions as coordinates on aerial photographs. These coordinates were then verified, and converted back into addresses for manual inspection and enquiry.

We monitored the quality and accuracy of outputs closely, tweaking the model as we went to help it learn and detect wet cooling towers more effectively. Once we were confident in the prediction quality of the model, we began using it to analyse new local aerial images, and find unregistered wet cooling towers in any defined location.

The end result –fully-automated tower detection

Today, empowered by the model we’ve put in place, DCMR can detect unregistered wet cooling towers without having to search for them manually. The model enables DCMR to input visual data for new focus areas, and quickly identify the cooling towers in that area.

The model will also help DCMR understand where to prioritise its efforts when inspecting all these new towers. By flagging factors like towers that are close to urban areas, or those exposed to other risk factors, it helps the team to make informed decisions about where the greatest risk of legionella may exist.

Equipped with that insight and information, the DCMR team is empowered to spend more of its time focusing on its most important goals –safeguarding public and environmental health.

Reflecting on the project

The reason I found this project so engaging was that in many ways, it’s data science at its best. We began with a clear, tangible problem that demanded a creative, data-driven solution, and we were able to combine our engineering and data expertise to deliver exactly that.

I’d encourage anyone considering a new data science project, or looking to transform their operations through data-driven innovation, to take a similar approach. Begin by identifying real problems that can’t be solved through traditional means. Those will always be your most compelling and valuable use cases, and the projects most likely to enable true transformation in your organisation. 

If you have any questions about this project, our data science approach, or how we can help you drive transformation through intelligent application of machine learning, please don’t hesitate to get in touch.