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

• DOI: 10.1016/j.ecolind.2023.111245
• Publication Date: 2023-11-10T19:45:51Z
• Authors: F. Zennaro; E. Furlan; D. Canu; L. Aveytua Alcazar; G. Rosati; C. Solidoro; S. Aslan; A. Critto
• Keywords: Biogeochemical Cycle; Hindcast; Vulnerability

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