Electronic Thesis and Dissertation Repository

Thesis Format

Monograph

Degree

Doctor of Philosophy

Program

Electrical and Computer Engineering

Collaborative Specialization

Planetary Science and Exploration

Supervisor

McIsaac, Kenneth A.

2nd Supervisor

Osinski, Gordon R.

Joint Supervisor

Abstract

Renewed interest in Solar System exploration, along with ongoing improvements in computing, robotics and instrumentation technologies, have reinforced the case for remote science acquisition systems development in space exploration. Testing systems and procedures that allow for autonomously collected science has been the focus of analogue field deployments and mission planning for some time, with such systems becoming more relevant as missions increase in complexity and ambition. The introduction of lidar and laser scanning-type instruments into the geological and planetary sciences has proven popular, and, just as with the established image and photogrammetric methods, has found widespread use in several research areas. However, the instrumentation and its data products have attributes, such as extreme accuracy, that can be leveraged in new ways. Success with algorithmic image processing systems such as TextureCam, AEGIS and others lead to the question: can an automated surface characterisation algorithm be created for this new type of data?

We developed a multi-stage, machine learning algorithmic pipeline for this purpose, utilising Adaptive Resonance as a generalised semi-automatic Feature of Interest (FoI) detector. Using data from the RIS4E analogue planetary exploration field deployments, we found that this system was highly effective at isolating FoIs at multiple scales, with some direction from an operator. In follow up work, we sought to examine the possibility of fully automating the process. We made additions to the pipeline to facilitate this, by creating a dimensional reduction (DR) based system, that used various metrics to detect a useful classification. We also increased the number and variety of datasets for testing. Finally, we considered the question of dimensionality and what effect the addition of exotic types of point cloud geometry descriptors, such as linearity and curvature, would have on the effectiveness of the classification process. We also added large, natural complex datasets to test the limits of the Adaptive Resonance network. By varying the DR configurations of the pipeline and point descriptors fed to Adaptive Resonance network, we found that tangent normal in combination with UMAP was the most effective, at 72%, for fully automatic FoI detection, from a variety of natural and synthetically generated point cloud datasets.

Summary for Lay Audience

With the potential for permanent human presence in the wider Solar System gathering momentum, this has reinforced an existing desire for more autonomous science collecting capability from remote exploration platforms. The development and rapid progression of high resolution sensing systems, such as lidar, presents the opportunity to increase the utilisation and effectiveness of planetary surface characterisation for exploration. However, leveraging this new technology for this purpose comes with a series of unique challenges.

In this work, we examined and quantified these challenges, with the aim of forming requirements to approach this exploration problem, resulting in a novel method of machine-learning based surface characterisation, centered around the Adaptive Resonance Theory machine learning technique. This structural extraction method transforms raw data commonly produced by high resolution laser and image sensor instruments, into a format that identifies geological surface features.

Initially, we demonstrated that with direction from the operator, the system was able to identify areas of significance in the point cloud data products, on both a general scale and at high resolution. The high resolution behaviour is specific enough that manual segmentation of such features would be both difficult and time consuming.

After this, we attempted to fulfil the requirement of high levels of automation by completely removing the requirement of user input. This extension was facilitated by the addition of two systems: a sequential processing functionality and a method for selecting the useful classifications using Dimensional Reduction (DR) techniques. We applied this on an expanded set of data, designed to better simulate a variety of terrain present on bodies in the Solar System.

Finally, a series of experiments were undertaken to more fully explore a variety of dimensions on the effectiveness of the system. Various combinations of commonly and used and more exotic dimensions were tried, to ascertain whether the system demonstrated preferential performance to any of them. Additional datasets consisting of more complex surface geometry, were used to examine system performance on data with less well-defined features.

By varying the DR methods of the pipeline and point dimensions fed to Adaptive Resonance network, we found that tangent normal dimension, in combination with the Uniform Manifold Approximation Projection (UMAP) DR technique, was the most effective, at 72%, for fully automatic surface geometry detection.

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS