Blue-Cloud Hackathon: EMODnet Chemistry for ‘Early Warning System for Severe Eutrophication’
In nearshore sea waters, algal overgrowth and decreased bottom oxygen concentration below deadly thresholds endanger benthic and demersal organisms. This environmental condition is mainly triggered by human-made eutrophication: enrichment of nutrients in estuaries and coastal areas for example by industry and agriculture.
During the Blue-Cloud Hackathon, from 7 to 9 February 2022, the EutroWarn team decided to use EMODnet Chemistry data to propose the development of early warning systems for possible severe eutrophication events.
While sophisticated dynamical models, e.g. MedEAF, are used for short term predictions (72h), forecasts on a monthly data basis are rare and challenging. Nevertheless, knowledge of eutrophication development in the order of months would be very helpful to initiate investigations and warnings. The strength of the artificial neural networks is that they are able to extract complex patterns and interactions between the variables and time series. Therefore we think that they might contribute to monthly predictions based on time series.
The competition was mainly open to marine scientists and researchers, data scientists, ICT experts, innovators, and students. A total of 32 teams took part in the hackathon, featuring 149 participants from 73 countries. Participants were challenged to discover and use Blue-Cloud, the online platform dedicated to marine sciences of the European Open Science Cloud, to face specific challenges within three topics and one wildcard:
- Understanding the Ocean
- Feeding the World
- Predicting environmental risks
- Wildcard: Hack the Blue-Cloud.
The EutroWarn team was led by Information and Communication Technology (ICT) & Artificial Intelligence (AI) experts of the Alfred Wegener Institute (AWI) and included marine scientists and data stewards of the Italian National Institute of Oceanography and Applied Geophysics (OGS). To develop its idea, called ‘Early Warning System for Severe Eutrophication’, the team retrieved from the BlueCloud Data discovery & access service a small subset of the EMODnet eutrophication and acidity time series data for the Northern Adriatic region, in the Mediterranean Sea: an area highly influenced by river input. EutroWarn used the ODV software to filter for time series with a sufficient number of data points and the presence of the following measurements: temperature, salinity, oxygen, phosphate, phosphorus, silicate, nitrate, nitrite and chlorophyll-a.
Then, it used two machine learning approaches to predict one month ahead and classify the state of eutrophication. The Team used artificial neural networks. The first approach was based only on the multivariate data, i.e. EutroWarn trained a neural network model with a recurrent and a convolutional layer with moving data chunks of 6 months, where the AI experts fed the first 5 months into the model and predicted the 6th month. To evaluate the prediction skill of the model, AWI and OGS used a test dataset with data yet ‘unknown’ to the model and additionally compared the outcome with a simple autoregressive model.
The second approach was a supervised learning setup, where the marine scientists identified severe eutrophication states and labeled every month as ‘severe’, ‘nothing’, ‘unclear’. The AI and ICT experts trained a multilayer perceptron model with the expert knowledge to create a skillful ‘human-like’ classification / prediction system for severe eutrophication. The skill of the model is evaluated again on yet ‘unknown’ data.
Decision and policy makers should be informed about the negative impacts of severe eutrophication events and should take measures to reduce such events, e.g., protection of remaining wetlands, limitations regarding fertilizers in agriculture.