The Cooperative Institute for Climate, Ocean, and Ecosystem Studies

Hi all! My name is Katherine Low, and I’m a junior at UC Berkeley studying Integrative Biology and Marine Science. I worked with my mentors Sarah Chinn and Peter Mahoney over at the NOAA Alaska Fisheries Science Center, in the Marine Mammal Laboratory. Our goal was to further the development of a machine learning model for the purposes of identifying and counting pinnipeds from aerial imagery. We believe machine learning has the potential to greatly increase the efficiency at which aerial imagery is reviewed, allowing us to acquire population counts more quickly. These counts are important for informing conservation and management decisions.

Our initial version of the object identification model requires further fine tuning, for which accurate ground truth data is required. Much of my focus was on correcting the model’s outputs to produce this ground truth data. By sending those corrected images back into the training pipeline, the accuracy of the model will improve. Notably, I added multiple new classes to our model’s repertoire, allowing it to identify pups that are important markers of population health.

Additionally, I participated in the analysis of current model performance and bias, with the intention of establishing a correction factor that can be applied to population counts. We developed parameters for the Slicing Aided Hyper Interface (SAHI) step of the detection process, aiming to eliminate double counts and maximize both precision and recall. We discovered that the model has the potential to both under and overcount animals, complicating the process of establishing a uniform correction factor.

Because the SAHI parameters were based off of a species-varied dataset, we then investigated harbor seals specifically. We found that the model had a severe tendency to overcount harbor seals. To correct this, we raised the confidence threshold to eliminate false positives, but a significant number of “phantom annotations” still remained. While this higher confidence threshold increased performance in the harbor seals alone, it was not stronger for the entire species-variable dataset. This implies specific SAHI parameters and correction factors must be established for different pinniped species. Looking forward, this project will continue through further model fine-tuning and correction factor development.

Through this program, I gained valuable experience in my field. Additionally, I enjoyed connecting with others in the Marine Mammal Laboratory. I have learned an exceptional amount about the day-to-day workings of marine mammal science, the varied career paths of those I worked alongside, and what I should expect as I look towards graduate school. I would like to thank my mentors and everybody at CICOES who made this experience possible, and I am incredibly grateful for the opportunity.

Project Introduction

Research Poster