Using machine learning to improve fish, insect, and UAV detection in pulsed lidar data

Abstract

The work in this dissertation focuses on detecting fish, insects, and unoccupied aerial vehicles (UAVs) in imbalanced lidar data. The highly imbalanced data produced by pulsed lidar systems makes detecting the objects of interest challenging and time-consuming; moreover, the detection process has traditionally been performed manually. By combining techniques from signal processing, machine learning, and imbalanced classification, I created automated data analysis pipelines that addressed the difficulties of manually analyzing imbalanced lidar data. Using my data analysis techniques, I was able to make contributions in all three application areas. In the fish detection application, I was able to automatically detect 86% of regions that contained fish. In the insect detection application, reliably identifying insects is one of the primary challenges in using lidar for long-term non-invasive insect monitoring. My methods identified 95% of the insect in the dataset, helping further lidar's readiness for insect monitoring. Lastly, my machine learning methods were able to detect 97% of the UAVs in the dataset, while we were only able to visually detect 43% of the UAVs. My contributions in each field will enable researchers that are not signal processing and machine learning experts to utilize automated techniques in their work, allowing researchers to spend more time applying their domain expertise to the problems they are actually interested in solving. By making contributions in three unique fields, I show that my data analysis pipeline is generally applicable for many detection applications.