This paper presents a machine learning technique for using large unlabelled survey datasets to aid automatic classification. We have demonstrated the benefit of this technique on a simulated synthetic aperture sonar (SAS) dataset. We designed a machine learning model to encode a representation of SAS images from which new SAS views can be generated. This novel task requires the model to learn the physics and content of SAS images without the requirement for human labels. This is called self-supervised learning. The pre-trained model can then be fine-tuned to perform classification on a small amount of labelled examples. This is called semi-supervised learning. By using a simulated dataset we can step-by-step increase the realism to identify the sources of difficulty for applying this method to real SAS data, and have a performance bench mark from this more idealised dataset. We have quantified the improved accuracy for the re-view model (ours), against a traditional self-supervised approach (autoencoder), and no pre-training. We have also demonstrated generating novel views to qualitatively inspect the model's learned representation. These results demonstrate our re-view self-supervised task aids the downstream classification task and model interpretability on simulated data, with immediate potential for application to real-world UXO monitoring.

This content is only available via PDF.