Automating the Boring Stuff: A Deep Learning and Computer Vision Workflow for Coral Reef Habitat Mapping
Title | Automating the Boring Stuff: A Deep Learning and Computer Vision Workflow for Coral Reef Habitat Mapping |
Publication Type | Thesis |
Year | 2020 |
Authors | Pierce, J |
Degree and Program | Master of Science |
Degree | Oceanography |
Number of Pages | 57 |
Date Published | December |
University | University of New Hampshire |
Location | Durham, NH |
High-resolution underwater imagery provides a detailed view of coral reefs and facilitates insight into important ecological metrics concerning their health. In recent years, anthropogenic stressors, including those related to climate change, have altered the community composition of coral reef habitats around the world. Currently the most common method of quantifying the composition of these communities is through benthic quadrat surveys and image analysis. This requires manual annotation of images that is a time-consuming task that does not scale well for large studies. Patch-based image classification using Convolutional Neural Networks (CNNs) can automate this task and provide sparse labels, but they remain computationally inefficient. This work extended the idea of automatic image annotation by using Fully Convolutional Networks (FCNs) to provide dense labels through semantic segmentation. Presented here is an improved version of Multilevel Superpixel Segmentation (MSS), an existing algorithm that repurposes the sparse labels provided to an image by automatically converting them into the dense labels necessary for training a FCN. This improved implementation—Fast-MSS—is demonstrated to perform considerably faster than the original without sacrificing accuracy. To showcase the applicability to benthic ecologists, this algorithm was independently validated by converting the sparse labels provided with the Moorea Labeled Coral (MLC) dataset into dense labels using Fast-MSS. FCNs were then trained and evaluated by comparing their predictions on the test images with the corresponding ground-truth sparse labels, setting the baseline scores for the task of semantic segmentation. Lastly, this study outlined a workflow using the methods previously described in combination with Structure-from-Motion (SfM) photogrammetry to classify the individual elements that make up a 3-D reconstructed model to their respective semantic groups. The contributions of this thesis help move the field of benthic ecology towards more efficient monitoring of coral reefs through entirely automated processes by making it easier to compute the changes in community composition using 2-D benthic habitat images and 3-D models. | |
https://scholars.unh.edu/thesis/1436 |