Applying Machine Learning Techniques to Improve Seabed Mapping Speed
Producing machine learning algorithms and software to improve existing filter-based software for more automatic seabed mapping.
Knowing what the sea floor looks like is very important for many activities, including exploration, transport, marine research and environment protection. This can be done in a number of ways but using sonar is still a common method as it is arguably the most cost-effective approach in industrial applications.
Among many sonar-based devices, GeoSwath is a popular interferometric instrument, designed and manufactured by Kongsberg GeoAcoustics Ltd, for seabed mapping. GeoSwath works by sending a single sound pulse, called a ping, down toward the seabed. The seabed, and any other objects within the beam (e.g. shipwrecks, bridge supports) will cause a reflection that is detected and used to provide a depth reading that is then subsequently used to produce a map of the seabed. However, a proportion of the returned data points are noise, and not related to the seabed or an object, so this then requires post-processing to remove them.
Currently the GeoSwath system process is performed using a series of filters that require configuration and monitoring by experienced human operators. Depending upon the nature of the seabed this can be a heavily human dependent process, so Kongsberg GeoAcoustics Ltd wanted to find out whether artificial intelligence could be applied to automate this process. They applied for an innovation voucher through EIRA to work with academics from the University of East Anglia on researching the feasibility of this idea, hoping to improve the efficiency of their current process.
With EIRA funding, two MSc students from the University of East Anglia (Holly Birch and Mike Sims) were selected to work on the project with UEA academics, Dr Wenjia Wang and Dr Ben Milner.
Dr Wenjia Wang is an associate professor of Artificial Intelligence at the School of Computing Sciences (CMP), UEA. He and his PhD students conduct research in the areas of data mining, machine learning, ensemble approach and artificial intelligence. He has been the Director of MSc in Computer Science since 2003 and the organiser for Masters Dissertation projects.
Dr Ben Milner is a senior lecturer in the School of Computing Sciences at UEA. His research interests are in machine learning applied to audio and speech applications These have included speech enhancement, noise reduction, whale detection and audio-visual processing.
The objectives of the project were:
- to review the relevant literature and methods
- to learn the existing GeoSwath software and its filtering algorithms and to investigate their performance
- to study the seabed survey data and write necessary programs to pre-process the data
- to apply some appropriate machine learning and artificial intelligence methods to classify the seabed survey sonar data with an aim of automating the process and
- to evaluate the effectiveness of the methods developed.
In completing these objectives, the team developed two automated classification approaches. ‘Classifiers’ are the means of categorising data into the pre-defined classes and generated from the data through some machine learning methods. Two different methods were trialled: supervised and unsupervised learning.
In the supervised approach, the classifiers need to be trained with labelled data and five types of classifiers were generated for comparison. Interestingly, the team found that the labelled data itself is not always reliable as “true” answers produced by the human experts are not always “true”, which can confuse the learning of the classifiers and affect their accuracy. In addition, the amount of the labelled data is limited.
The team also developed an unsupervised approach that does not require the labelled data. A hybrid-unsupervised algorithm was developed with a combination of two clustering methods and some expert rules.
These supervised and unsupervised methods were evaluated on different seabed terrains and their classification performance was measured with sensitivity (the accuracy of classifying the positive points) and specificity (the accuracy of classifying the negative points), and some other common metrics in machine learning.
The experimental results showed unsurprisingly that performance of these supervised classifiers varied quite significantly. Some produced very good sensitivity accuracies, as high as 98% but some classifiers were only around 75% accurate. A third approach, an ensemble method, was developed to address these big variances. An ensemble is a combination of multiple modules selected to work together to overcome each other’s weaknesses. Various types of ensembles were built and the testing showed that they were very accurate and also much more reliable.
The unsupervised approach also produced very encouraging results. Without using any labelled data in learning, the developed algorithm achieved sensitivity accuracy between 80% to 98% on the test data.
In summary, the results from the initial study have demonstrated the capability of AI/machine learning methods for classifying ping returns as genuine or noise and have a much-reduced dependency on experienced human intervention. In some cases, the AI methods were seen to be doing better than the human as they identified some targets correctly that the human missed due to the subjective nature of sonar data processing.
However, these methods were not developed fully, nor systematically tested with sufficiently validated datasets, so they require further developing and testing for making them technically ready for industrial applications. The methods also need further work on software implementation to make them suitably integrated into the GeoSwath system, so further research, development and evaluation is needed.
The academic team had the following to say about their experiences of working on the project:
“Thanks to EIRA for awarding GeoAcoustics a research voucher, which helped us to set up two MSc dissertation projects to carry out some feasibility studies in applying and developing AI and machine learning techniques to this real world problem. With the support from the company and my colleague Ben, we as a team completed the studies successfully and developed two different frameworks, which achieved better than expected results. These findings demonstrated the technical capability of AI methods for the problem and laid solid foundations for further development. The students gained a lot of real world experience, which helped them to find proper employment after their graduation. As academics, we were also benefited from this EIRA research voucher as we were provided with an opportunity to apply our expertise to a different real-world problem and also apply for new grants. The project as a whole could be further developed to be an impact case for the next REF (Research Excellence Framework).
So, as the leader of the project, I would also like to thank the people at the EIRA, particularly Ian and Beth at the Research and Innovation Office, UEA, and the team at GeoAcoustics, for their support from the grant application to the execution of this project.” – Dr Wenja Wang, UEA
“Working with GeoAcoustics on developing an AI method to improve seabed mapping was a truly interesting and worthwhile project for my MSc dissertation. I really appreciated the opportunity and enjoyed the challenges in researching unsupervised machine learning algorithms and expert rules, and exploring their best combination to form a hybrid method. Our method produced very promising results and could be used for future research.” – Holly Birch, MSc Computing Science (2018-19)
“My dissertation project was to develop AI methods to solve an interesting real-world problem. Thanks to the collaborative company who provided their proprietary software and their raw sonar data files. Although challenging, with the support from my supervisors and experts from the company, I quickly learned the necessary knowledge and skills and built a machine learning framework with some supervised models. I thoroughly enjoyed the project and was very satisfied with the reasonably good testing results.”– Michael Sims, MSc Advanced Computing Science (2018-19)
As a result of this feasibility work, Kongsberg GeoAcoustics successfully applied for an R&D grant through EIRA to continue the work enabling the company to have a fully operational, integrated AI process to enhance the accuracy of their existing GeoSwath system.
Kongsberg GeoAccoustics have been successful in applying for an EIRA R&D grant. The next phase of the project is due to complete in July 2020.
Capitalising on the findings in the innovation voucher project, the R&D aims to:
- Identify and establish some datasets on various seabed terrains with “ground” truth as industrial benchmarks for evaluating current and new mapping methods and tools
- To test the methods developed in the feasibility study thoroughly with the established benchmark datasets
- To improve these methods in terms of not only accuracy but also consistency
Finally, the next phase of the project will develop the design and implement the software modules in order to be integrated into the existing GeoSwath software GS4.
Photo by Joshua Fuller on Unsplash