Advanced science.  Applied technology.


Autonomous Aerial Detection of Hazardous Liquid Pipeline Leaks in Water Bodies, 10-R8862

Principal Investigators
Jake Janssen
Heath Spidle
Inclusive Dates 
07/01/19 - Current


The prevailing leak detection systems used today are unable to detect leaks that are smaller than 1% of the pipeline throughput. As such, many leaks go unnoticed until they reach a size large enough to be detected by Computational Pipeline Monitoring (CPM) systems or until a leak is observed. Because of commercially available solutions for detecting crude on water and the increasing pressure from regulating agencies for operators to improve their leak detection capabilities, there is an increased interest in developing technology to detect oil on water.

SwRI developed the Smart LEak Detection (SLED) system to autonomously monitor pipelines for hazardous chemical spills. SLED leverages visible and thermal cameras along with machine learning techniques to reliably detect the chemical fingerprint of small hazardous liquid leaks on surfaces typically found near pipelines. The foundations for SLED will be used for this project to see if we can adapt this approach to detect crude oil on the surface of water.


The goal of this project was to investigate visible and thermal cameras for detecting crude oil on water bodies. The first step of this project examined the sensors. The previous SLED project had already selected thermal and visible cameras, which were used as a starting point. The SLED thermal camera was used, but a higher resolution visible camera provided better detections.

The second step was to collect data on crude oil and non-crude oil materials that could fool the machine learning algorithm. This happened over three different collections. The first collection used four 4-foot by 4-foot pools of water with varying substrates to give different color contrasts. The second data collection used four 4-foot by 4-foot pools again, and a larger 10-foot diameter pool was used to provide more area for the oil to move. The last data collection used two 10-foot diameter pools and split the collection over two days. The first day consisted of negative training, consisting of leaves and black trash bags. The second day collected crude oil data with more variations in light, oil quantity, shadows, and types of oil than the two previous data collections.


The results of this project have shown favorable results that crude oil can be detected using visible and thermal cameras. The algorithm achieved 98% accuracy at detecting oil in pools used to collect data. This is promising and, with more diverse training data in different environments, we are confident that our algorithm can be further improved.