Leak detection using thermal imagery: Deep learning versus traditional computer vision state-of-the-art

As a cornerstone of climate-neutral heat supply in urban areas, district heating systems require monitoring to detect and mitigate leaks in their subterranean pipelines. Recent research has focused on an approach involving thermography, where leaks are detected as hot-spots in remote sensing imagery. To this end, various traditional computer vision algorithms have been implemented to automate anomaly detection.

This paper pursues a new approach that has so far received little attention in the context of leak detection in district heating pipelines: deep learning, specifically supervised semantic segmentation. By creating a generalisable, multi-stage training procedure to tackle the prevalent limited dataset problem, various architectures are tailored to this anomaly detection task, of which the SegFormer-B2 with Tversky loss is found to perform best. Via comprehensive quantitative, qualitative, explainable AI, and holistic evaluation, the model is assessed and compared to state-of-the-art traditional algorithmic alternatives. It is found to excel, outperforming previous intersection over union scores by almost 10 %pt and maintaining a high precision with little detriment to recall and detection rate.