Multi-risk assessment in the Mediterranean Sea using Random Forest

Marine coastal ecosystems (MCEs) are of vital importance for human health and well-being. They are well known carbon sinks playing a key role in climate regulation mechanisms on Earth. They also provide support in climate adaptation, providing ecosystem services like coastal protection, flood control, etc. However, MCEs result to be globally threatened by increased environmental pressures, either related to climate change or to more direct human impacts. Focusing on the Mediterranean Sea, the growing interplay among human-made and climate-related pressures could affect the survival of seagrasses meadows, the most representative ecosystem in this ecoregion. In this regard, it is necessary to understand how cumulative impacts induced by the complex interplay among multiple stressors could affect them, also considering the effect of potential climate changes.Based on these considerations, the state of art of this research field was reviewed, showing the machine learning potential for the understanding and unraveling of environmental dynamics underpinning ecosystems functioning and health. Generally tree based models resulted as the most suitable for this task, and above all the mostly applied methods in the analyzed studies. Building on these applications, the key objective of this Thesis is to develop a Machine Learning model (i.e., Random Forest) supporting a sound evaluation and scenario analysis of multiple risks resulting from the interaction among human-made and climate-related pressures affecting seagrass meadows. To this aim, GIS-based monitoring data representing stress factors (e.g. change in the temperature and salinity regime, nutrients concentrations, coastal development) are integrated into the assessment flow as model predictors. On the other hand, data on ecological conditions (i.e. distribution, pattern of connectivity, shannon index) and ecosystem services capacity (i.e. carbon sequestration and denitrification) are used as response variables. Particularly, they are classified into three classes respectively low, moderate and high MCEs state/services capacity. The model performance is estimated by applying spatial cross-validation and it achieves an overall accuracy about 82%. Particularly, the Random Forest is able to perfectly classify the connectivity pattern; on the other hand, distribution, carbon sequestration and denitrification variables are classified rather well (showing lower performance in predicting the second class- moderate MCEs state/services capacity, as it is not perfectly separable from the others); the model shows uncertainty for the shannon index. Overall, results of the analysis showed that the ecological conditions of seagrass meadows are mainly determined by human-related pressures (identified with the distance from river mouths and main urban centers variables), as well as to changes in nutrient concentration and sea surface temperature. The model is then used to predict response variables against different future scenarios. Specifically, future scenarios from climate numerical models, for the 2050 and 2100 timeframes, were considered for variables related to temperature and salinity. Particularly, they were used to predict future conditions of the ecosystem, showing a decrease in seagrass coverage, and related benefits, in both future timeframes. The developed model provides useful predictive insight on possible future conditions and dynamics underpinning seagrass ecosystem response to multiple pressures (including CC), supporting marine managers towards a more effective ecosystem-based adaptation and management measures, needed to protect and restore MCEs, as well as to achieve the “good environmental status” as required under the EU Marine Strategy Framework Directive. This thesis was carried out in the frame of the H2020 MaCoBioS project (https://macobios.eu) in collaboration with the Foundation Centro Euro-Mediterraneo sui Cambiamenti Climatici (www.cmcc.it)