Relevance of Landscape Metrics in Predicting Water-Related Ecosystem Services demonstrated on the Arno River Basin Italy

Quantification of the impact of Land Use and Land Cover Change on ecosystem services is crucial for identifying critical areas requiring conservation efforts and sustainable land practices. Landscape metrics, which quantify the spatial arrangement of land cover types within a landscape, have emerged as valuable tools for systematically understanding such impacts and operationalize land use changes. In this study, we explored the contribution of landscape metrics to predicting water-related ecosystem services, mainly water provisioning at the watershed scale represented by runoff. We employed a random forest model to approximate   distributed  maps of runoff  for the Arno River Basin in Italy, obtained from a nationally used gridded hydrological water balance model .   Open access earth observation data were used as environmental predictors, including hydroclimatic variables, land use classes, and of landscape metrics related to runoff. The out of bag error is used to assess model performance along with variance and bias, and a leave one sub-watershed out a time is used for validation.  Our results demonstrated that despite a relatively low feature importance compared to other direct hydrological predictors like precipitation and temperature, landscape metrics, especially the core area index of forest and agricultural land use, captured significant interactions between forest and agricultural land use and their influence on water provision. In wet and normal conditions where precipitation is the predominant factor, the significance of these metrics intensifies, whereas in dry conditions characterized by dominant groundwater recharge processes, their relevance diminishes.  Leveraging land use data and earth observations, this approach clarifies the complex LULCC-ecosystem service relationships, informing strategies that balance ecosystem multifunctionality with environmental sustainability despite limitations in comparison to process based models.