Venice lagoon chlorophyll-a evaluation under climate change conditions: A hybrid water quality machine learning and biogeochemical-based framework

Abstract

Climate change presents a significant challenge to lagoon ecosystems, which are highly valued coastal environments known for their provision of unique ecosystem services. As important as fragile, lagoons are vulnerable to both natural processes and anthropogenic activities, and this vulnerability is exacerbated by the impacts of climate change, which are likely to result in severe ecological consequences. The complexity of water quality (WQ) processes, characterized by compounding and interconnected pressures, highlights the importance of adequate sophisticated methods to estimate future ecological impacts on lagoon environments. In this setting, a hybrid framework is introduced where Machine Learning (ML) and biogeochemical (BGC) models are integrated in a sequential modelling approach. This integration exploits the unique strengths offered by both models. The ML model allows capturing and learning linear and nonlinear correlations from historical data; the BGC interprets and simulates complex environmental systems subject to compounded pressures, building on identified causal relationships. Multi-Layer Perceptron (MLP) and Random Forest (RF) ML algorithms are trained, validated and tested within the Venice lagoon case study to assimilate historical WQ data (i.e., water temperature, salinity, and dissolved oxygen) and spatio-temporal information (i.e., monitoring station location and month), and to predict changes in chlorophyll-a (Chl-a) conditions. Then, projections from the BGC model SHYFEM-BFM for 2019, 2050, and 2100 timeframes under RCP 8.5 are integrated into the ML model (composing the hybrid ML-BGC model) to evaluate Chl-a variations under future biogeochemical conditions forced by climate change projections. Moreover, the SHYFEM-BFM standalone Chl-a projections are also used to compare the hybrid and the BGC scenarios. Annual and seasonal Chl-a predictions are developed by classes based on two classification modes (median and quartiles) established on the descriptive statistics computed on historical data. Results from the case study showed as the RF successfully classifies Chl-a with an overall model accuracy of about 80% for the median and 61% for the quartiles modes. Concerning future climate change scenarios, results revealed a decreasing trend for the lowest Chl-a values (below the first quartile, i.e. 0.85 µg/l) moving to the far future (2100), with an opposite rising trend for the highest Chl-a values (above the fourth quartile, i.e. 2.78 µg/l). On the seasonal level, summer remains the season with the highest Chl-a values in all scenarios, although in 2100 a strong increase in higher Chl-a values is also expected during the springtime one. The proposed hybrid framework represents a valuable approach to strengthen both multivariate Chl-a modelling and scenarios analysis, by placing artificial intelligence-based models alongside biogeochemical models.