What if we could somehow forecast when toxic algal blooms or fish kills happen? This would help our coastal communities especially those dependent on shellfish or fish farming to have more options on what to do with their tahong (mussels) or talaba (oysters) or bangus (milkfish) if they had an inkling that their shellfish might become toxic and unharvestable, or if a fish kill could happen sometime soon. This paper (https://www.sciencedirect.com/…/artic…/pii/S0048969719361698) is our attempt at building forecasting models for toxic blooms and fish kills at Bolinao-Anda in Pangasinan. This site unfortunately experiences both events frequently and this is where we have more detailed data.
We used a machine learning method called the random forest classification to try to predict if particular combinations of conditions led to a toxic bloom or a fish kill event. But a constraint we have to work with is that the data we use for forecasting are only the ones we can get in real-time from the water using sensors: temperature, salinity, dissolved oxygen, pH and chlorophyll. Otherwise, if we included other likely important factors like the amount of nutrients in the water that cannot be measured in real-time, then the forecasting capability also suffers.
How did our random forest models do? They actually performed quite well: 96.1% accuracy for fish kills, and 97.8% accuracy for shellfish bans. For both of the models, the most important contributing factor was dissolved oxygen. Looks promising, but now the real test is to actually deploy the models in the field together with the sensors we're currently developing and see how they truly perform in actual conditions!
