(this is a technical report that is a work in progress and hopefully a future published paper)
Introduction:
Red Tide is a general term for events that are caused by the rapid reproduction of algae that change the color of the ocean surface. Sometimes these events are referred to as harmful algal blooms (HABs) when they coincide with the degradation of the environment such as the release of toxins. One impact that HABs have on the environment is the accumulation of toxins in the tissues of mollusks. These toxins have little effect on the mollusks themselves, but when accumulated at a high enough concentration and ingested by humans, they can result in paralytic shellfish poisoning (PSP) and even death (Grattan et al).
The Philippines is one country that suffers significant health and economic consequences from HABs. The Filipino Bureau of Fisheries and Aquatic Resources (BFAR) reports that between 1983 and 2002 there were 2122 cases of PSP and 117 deaths (Bajarias et al), ranking the Philippines above all other Asian countries (Azanza and Taylor). During particularly bad outbreak years, the economic effects can result in millions of pesos in losses, disruption of international trade as import bans are imposed, and the displacement of thousands of artisan fishers (Bajarias et al).
The first recorded case of PSP in the Philippines coincided with visual reports of red tides in the year of 1983 (Estudillo). Since then, PSP and red tides have proliferated throughout the archipelago (Bajarias et al) and the Filipino government has taken measures to limit losses caused by red tides. A national red tide task force was created and implemented a monitoring program that utilizes site sampling and bioassays to determine toxicity. Due to budgeting constraints, the frequency and duration of the monitoring surveys vary between sites (Bajarias et al). Reports from this red tide monitoring program were used in this study as the response variable for this forecast.
Example of report:
An artificial neural network (ANN) is used in this study to predict the toxicity status of different sites in the Philippines. ANNs have shown promising results in forecasting HABs and shellfish contamination and were chosen for their ability to extract nonlinear relationships in data.
Methods:
Response Variable
The labels for this model, and therefore its predictive focus, are the Red Tide bulletins published by BFAR. These reports are released at an irregular interval, typically about two weeks apart, and state the status of green mussel toxicity for many sites across the Philippines. Toxicity is reported as a binary toxic or not-toxic status, so the model’s output is binary as well.
These reports are archived on Facebook up until 2012. An information request was also submitted to the Filipino Bureau of Fisheries, which provided the status of red tides back to the program’s inception in 2002. Together, these reports provide a coarse but 20-year timeline of toxicity status for a limited number of sites in the Philippines. About 8 sites have been tested regularly since the program’s inception.
Data collection:
A seemingly novel aspect of this project was the use of Google Earth Engine (GEE) for the entirety of data collection. GEE is a free online database with thousands of terabytes of spatial and temporal data that is hosted and accessed through Google’s own computing power. This data can be queried and downloaded using JavaScript commands through GEE’s API.
A number of different datasets were used, including temperature, salinity, water velocity, and water height from the Hybrid Coordinate Ocean Model; percent cloud coverage from the NCEP-DOE Reanalysis 2 project; and precipitation from the Global Precipitation Measurement mission.
Every one of these datasets is extrapolated through space and time and is subject to considerable error. Data was organized into weekly blocks, and the mean value for all variables was collected at various locations and associated with the site of reported toxins. This data was paired with the toxin labels and organized into time series for training.
Training:
A long short-term memory (LSTM) sequential neural network was used because of the importance that sequence has in the input data, since it is a time series. The LSTM structure is a variant of the recurrent neural network (RNN) structure and allows states to be carried from one time step to the next, in this case from one week to the following week. The data was loaded forward-facing, meaning the oldest weeks were fed into the model first, while the most recent weeks were fed in last. This was done under the assumption that the most recent environmental conditions carry the most weight on the status of toxicity.
RESULTS
Surprisingly, out of the three models below, the one with the least lag before its prediction interval (two weeks) performs the worst. The best performing model is the one that predicts four weeks into the future.
The three plots below show the performance of each model on the entire dataset, distinguishing between the training and validation sets. The model has never seen the validation set, so its performance should carry more weight when evaluating the model.
-
-
Bibliography
Andres, John Kristoffer, Aletta T. Yñiguez, Jennifer Mary Maister, Andrew D. Turner, Dave Eldon B. Olano, Jenelyn Mendoza, Lilibeth Salvador-Reyes, and Rhodora V. Azanza. 2019. “Paralytic Shellfish Toxin Uptake, Assimilation, Depuration, and Transformation in the Southeast Asian Green-Lipped Mussel (Perna Viridis).” Toxins 11 (8): 468. https://doi.org/10.3390/toxins11080468.
Azanza, R. V., and F. J. Taylor. 2001. “Are Pyrodinium Blooms in the Southeast Asian Region Recurring and Spreading? A View at the End of the Millennium.” Ambio 30 (6): 356–64.
Bajarias, Fe Farida A, Masaaki Kodama, and Setsuko Sakamoto. n.d. “In the Philippines: Option.”
Cho, Hyungmin, U-Jin Choi, and Heekyung Park. 2018. “DEEP LEARNING APPLICATION TO TIME-SERIES PREDICTION OF DAILY CHLOROPHYLL-A CONCENTRATION.” In , 157–63. Naples, Italy. https://doi.org/10.2495/EID180141.
Chollet, François, and Joseph J. Allaire. 2018. Deep Learning with R. Shelter Island, NY: Manning Publications Co.
Cruz, Rafaela C., Pedro Reis Costa, Susana Vinga, Ludwig Krippahl, and Marta B. Lopes. 2021. “A Review of Recent Machine Learning Advances for Forecasting Harmful Algal Blooms and Shellfish Contamination.” Journal of Marine Science and Engineering 9 (3): 283. https://doi.org/10.3390/jmse9030283.
Gacutan, R. Q., M. Y. Tabbu, E. J. Aujero, and F. Icatlo. 1985. “Paralytic Shellfish Poisoning Due to Pyrodinium Bahamense Var. Compressa in Mati, Davao Oriental, Philippines.” Marine Biology 87 (3): 223–27. https://doi.org/10.1007/BF00397798.
Grattan, Lynn M., Sailor Holobaugh, and J. Glenn Morris. 2016. “Harmful Algal Blooms and Public Health.” Harmful Algae, Harmful Algal Blooms and Public Health, 57 (July): 2–8. https://doi.org/10.1016/j.hal.2016.05.003.
Montojo, Ulysses M., Setsuko Sakamoto, Mirriam F. Cayme, Norvida C. Gatdula, Elsa F. Furio, Juan R. Relox, Shigeru Sato, Yasuwo Fukuyo, and Masaaki Kodama. 2006. “Remarkable Difference in Accumulation of Paralytic Shellfish Poisoning Toxins among Bivalve Species Exposed to Pyrodinium Bahamense Var. Compressum Bloom in Masinloc Bay, Philippines.” Toxicon 48 (1): 85–92. https://doi.org/10.1016/j.toxicon.2006.04.014.