This presentation was held at the 2020 NAFEMS UK Conference "Inspiring Innovation through Engineering Simulation". The conference covered topics ranging from traditional FEA and CFD, to new and emerging areas including artificial intelligence, machine learning and EDA.

Resource Abstract

Background: Invasive non-native species (INNS) pose a threat to asset owners. Their management and control, particularly Japanese Knotweed, have been largely ineffective. The development of this new tool is part of an ongoing effort to improve the methods used to understand the exact extent of invasive species, develop a historical record of their spread and simulate their future spread. The aim was to provide a resource that will inform the management efforts while providing time savings and keeping boots off ballast.

Methods: The development was undertaken by the Intelligent Data Analytics team at Mott MacDonald. We applied the best in class machine learning technologies to pre-collected and already available data sources. For the first task, an in-house object detection single-shot-style deep neural network, based on the YoloV3 research project (Redmon et al., 2018), was developed. This algorithm was trained using the data from a previous project that the ecology team at Mott MacDonald undertook using traditional methods. The previous project involved identifying four species of INNS from cab footage collected from the front of trains across rail tracks in South Wales . The algorithm was trained to perform the same task as the ecologists for Japanese Knotweed. The data from the previous project did not include sufficient instances of the other INNS along the CVL to effectively apply this methodology. For the second task, a second in-house algorithm was developed. It is an object detection and instance segmentation algorithm that is an adaptation of the Mask R-CNN research project (He et al., 2018). This algorithm highlights the pixels in an image that belong to an identified object. The data from the first task in addition to the data from the previous ecology project were used to train the algorithm to identify Japanese Knotweed in 4cm per pixel aerial imagery of all rail right of way in Wales.

Results: Results from testing of the first task yielded a very strong match to the manually-labelled data over the same patches of railway. Even with the limited labelled data used, the INNS detection tool correctly identified 93% of all instances of INNS within the test area. The second task is currently in the final stages of development and testing but has shown to accurately identify areas of INNS and provide quantified extents of INNS within a sample area.

Conclusion: The devised methods successfully identify instances of INNS along the rail track using cab footage. The development of task 2 provides quantified areas of instances of INNS. The application of this tool across the UK combining input data from multiple datasets (aerial imagery, satellite, linear asset imagery) could develop a holistic approach to identifying the presence and spread of INNS and feed directly into growth prediction models, on a much greater and more effective scale, to inform future asset management strategies. The successful deployment of the INNS tool forms part of the efforts of the Mott MacDonald Intelligent Data Analytics team to tackle issues through the combination of state-of-the-art machine learning algorithms with large datasets within the field of asset management. Other implementation, for example, wet beds, retaining walls, structural defects, ash dieback, and biodiversity net gain, are underway.