Workflow towards autonomous and semi-automized UXO Survey and Detection

Abstract

This paper presents a workflow for UXO detection based on multibeam data in combination with AUV-based ground truth. An artificial neuronal network (ANN) is trained on manually annotated multibeam data and aims for making UXO detection and the generation of target lists faster and more objective. Prior to annotation the data is checked according to several quality factors to ensure that it fits for the purpose of object detection. The quality and accuracy of annotations has an influence on the predicted probabilities of the ANN, as the probabilities of the annotations determined by the experts are considered during training. To make this whole workflow even more effective in terms of survey time, the quality check and ANN analysis can run automated on the survey vessel. While MBES mapping continues, autonomous underwater vehicles (AUVs) can be used to ground truth possible targets with additional sensors, such as geomagnetic and under water cameras. The more precise the training data, the more reliable the ANN outcome will be.