Toward Global Localization of Unmanned Aircraft Systems using Overhead Image Registration with Deep Learning Convolutional Neural Networks

<p>Global localization, in which an unmanned aircraft system (UAS) estimates its unknown current location without access to its take-off location or other locational data from its flight path, is a challenging problem. This research brings together aspects from the remote sensing, geoinformatics, and machine learning disciplines by framing the global localization problem as a geospatial image registration problem in which overhead aerial and satellite imagery serve as a proxy for UAS imagery. A literature review is conducted covering the use of deep learning convolutional neural networks (DLCNN) with global localization and other related geospatial imagery applications. Differences between geospatial imagery taken from the overhead perspective and terrestrial imagery are discussed, as well as difficulties in using geospatial overhead imagery for image registration due to a lack of suitable machine learning datasets. Geospatial analysis is conducted to identify suitable areas for future UAS imagery collection. One of these areas, Jerusalem northeast (JNE) is selected as the area of interest (AOI) for this research. Multi-modal, multi-temporal, and multi-resolution geospatial overhead imagery is aggregated from a variety of publicly available sources and processed to create a controlled image dataset called Jerusalem northeast rural controlled imagery (JNE RCI). JNE RCI is tested with handcrafted feature-based methods SURF and SIFT and a non-handcrafted feature-based pre-trained fine-tuned VGG-16 DLCNN on coarse-grained image registration. Both handcrafted and non-handcrafted feature based methods had difficulty with the coarse-grained registration process. The format of JNE RCI is determined to be unsuitable for the coarse-grained registration process with DLCNNs and the process to create a new supervised machine learning dataset, Jerusalem northeast machine learning (JNE ML) is covered in detail. A multi-resolution grid based approach is used, where each grid cell ID is treated as the supervised training label for that respective resolution. Pre-trained fine-tuned VGG-16 DLCNNs, two custom architecture two-channel DLCNNs, and a custom chain DLCNN are trained on JNE ML for each spatial resolution of subimages in the dataset. All DLCNNs used could more accurately coarsely register the JNE ML subimages compared to the pre-trained fine-tuned VGG-16 DLCNN on JNE RCI. This shows the process for creating JNE ML is valid and is suitable for using machine learning with the coarse-grained registration problem. All custom architecture two-channel DLCNNs and the custom chain DLCNN were able to more accurately coarsely register the JNE ML subimages compared to the fine-tuned pre-trained VGG-16 approach. Both the two-channel custom DLCNNs and the chain DLCNN were able to generalize well to new imagery that these networks had not previously trained on. Through the contributions of this research, a foundation is laid for future work to be conducted on the UAS global localization problem within the rural forested JNE AOI. </p>

Name thesis:degree_name

Doctor of Philosophy in Geosciences (PhD)

Level thesis:degree_level

Dissertation

Year dc:date.available

2022

Author dc:creator

• Linck, Rachel

Advisor dc:contributor.advisor

• Cothren, Jackson D.

Contributors dc:contributor

• Rainwater, Chase E.
• Tullis, Jason A.

• Geospatial imagery
• terrestrial imagery
• geospatial data
• UAS imagery
• machine learning
• Geographic Information Sciences
• Remote Sensing
• Spatial Science

Repository record dc:identifier

https://scholarworks.uark.edu/etd/4545

OAI identifier oai:identifier

oai:scholarworks.uark.edu:etd-6095