Open Collections

Kwok, Jasmine; Zhou, Shirley; Xiong, Vanessa; Wang, Wenwen

Two problematic phytoplankton on the coast of British Columbia (BC) are Alexandrium spp., which produces saxitoxin that can cause Paralytic Shellfish Poisoning, and Pseudo-nitzschia spp., which produces domoic acid that can cause Amnesic Shellfish Poisoning. Phytoplankton such as Alexandrium spp. can release biotoxins in Harmful Algal Blooms (HABs). Climate change is exacerbating the development of HABs with shifting climatic variables such as temperature and nutrient availability that affect phytoplankton dynamics. Currently, biotoxin levels are monitored by the Canadian Shellfish Sanitation Program, but the system fails to consistently and inclusively inform First Nations’ communities about safe locations and times to harvest seafood. The risk of these poisonings and insufficient testing of shellfish harvesting sites along the coast of BC harms the food accessibility and sovereignty for First Nations along the Pacific coast. Using the variables of biotoxin concentration, sea-surface temperature, air temperature, salinity, precipitation, and NO₃ and PO₄ concentration, Multi-layer Perceptron (MLP) and Random Forest machine-learning models were trained to conduct a 10-day prediction of biotoxin concentration locations exceeding regulation level. To develop the model, the area under the receiver operating characteristic (ROC) curve and feature importance plots were analysed for model accuracy and data applicability. We found that predicting biotoxin levels from a Random Forest model compared to MLP revealed similar levels of accuracy. The model largely uses data from global ocean and biogeochemical climate reanalysis models, rather than in-situ measurements. This is due to reanalysis data being open source; more complete; and having better spatial and temporal resolution. Further, we found that the reanalysis data had high correlation with observational data in our study area and would serve as a sufficient proxy. Overall, the Random Forest model using all predictor variables had the highest accuracy of 91%. In predicting AST exceedances with that model, air temperature, sea surface temperature and PO₄ concentration at 0.5m depth had the highest importance. In predicting PST exceeding regulatory threshold with the same model, results showed that salinity and PO₄ concentration at 0.5 m and 20 m depth had the highest importance. This model and its findings can help support First Nations food security and sovereignty by informing shellfish safety programs of high-risk areas and providing Indigenous harvesters an independent way of assessing harvest area status.

Text

University of British Columbia. ENVR 400

Vancouver : University of British Columbia Library

10.14288/1.0442341

Science, Faculty of; Earth, Ocean and Atmospheric Sciences, Department of

Undergraduate

DSpace