Background/Question/Methods Oyster reefs are a key component of estuarine ecosystems. They provide services such as water filtration, carbon sequestration, and shoreline protection as well as food and habitat for other organisms. However, they are increasingly at risk due to over-harvesting, disease, and habitat degradation. Mapping oyster reefs accurately is crucial for much-needed preservation and restoration, but spatial data is currently lacking. New approaches to semi-automated mapping are now possible due to recent advances in artificial intelligence (AI) and the widespread availability of open data, authoritative and accessible to the public. The AI model in this study “learns” by comparing aerial imagery to predefined training samples. This study used color-infrared imagery from 2019 with a spatial resolution of 0.6 meters, provided by the USDA's National Agriculture Imagery Program (NAIP). The training data came from the most comprehensive dataset publicly available, provided by the Florida Fish and Wildlife Conservation Commission-Fish and Wildlife Research Institute (FWC-FWRI). FWC-FWRI's dataset pulls from the best available sources, but it is still incomplete and imperfect. Using a subset of this data as a starting point, with 20% set aside for validation, an AI model was trained to accurately map oyster reefs in Oyster Bay, Wakulla County, Florida. The model was trained in a Jupyter notebook using the Python programming language and esri's implementation of the U-Net convolutional neural network architecture.
Results/Conclusions Results were quite successful; after only 10 epochs of training, the model was able to produce oyster reef maps that were considerably more accurate than the data provided by FWC-FWRI. Training took about one hour on the laptop used in this study. With respect to training data, the model reported a 95% accuracy and an F1 score or dice coefficient of .665. This is good because the training data itself was imperfect. Using this imperfect training data, the AI model successfully produced a more complete, accurate, and up-to-date map of oyster reefs. There was no evidence of overfitting, and the model generalized well enough to work outside of the immediate training area. The model could be improved by further refining the training data and by giving it more training time. This study demonstrates the usefulness of AI and computer vision in the semi-automated mapping of ecologically significant features. All data used in the development of the model is publicly available and free to use.
