UC San Diego

Improving oceanographic data collection involves two components: improving instrumentation, and improving the processing and analysis of resulting measurements. While advancing technology has improved and expanded data collection, processing these data has become a significant problem for under-resourced academic labs, rendering much of these data underutilized. The experiments described in this dissertation are contributions to the effort of improving data processing in oceanography. Chapter 1 examines parametrization of particle size distribution measurements in the Arctic, collected during decades of field expeditions by Dariusz Stramski, Rick Reynolds, and others. Experiments suggest that the commonly-used Junge-type power law model for parametrizing particle size distributions is insufficient, and cumulative distribution functions may offer a superior alternative. Particle size directly physically affects light, so the particle size distribution affects the signals of optical instruments. These results will increase the utility of satellite imagery, both by assisting the measurement of particle size from satellites, and by improving understanding of the impact of different seawater characteristics on optical signals. Chapter 2 discusses cutting-edge, technology-enabled survey image and 3D model techniques for studying coral reefs. The report focuses on lessons learned by the Sandin/Kuester labs from a decade of experience. Improving and standardizing data collection in the field allows research groups to pool datasets and compare results. Pooled databases are valuable for developing processing and analytical tools like neural networks, make those tools useful to more research groups, and enable ecological analysis at larger scales. Chapters 3 and 4 evaluate neural-network-assisted tools that can expedite taxonomic labeling of survey image products such as orthoprojections and pointclouds. Chapter 3 contains the first published investigation of 3D coral pointcloud segmentation with 3D neural networks. This new technology shows great promise, as evidenced by the time savings (36%) and high prediction accuracy in some contexts (~70-90%) recorded by our experiments, but requires more comprehensive, standardized datasets than what is currently available to fully develop or confidently evaluate. These tools are already capable of increasing the utility of survey imagery by expediting their annotation, which is necessary for many types of analysis, and further development will further improve their performance.