Factors Predicting the Occurrence of Ecologically Valuable Springs: Springs: A Random Forest Modelling Approach

Spring habitats are among the most threatened aquatic ecosystems in Switzerland. Effective protection requires assessing their ecological value (e.g., structural and faunistic condition), yet the existing FOEN method is resource-intensive and expensive. A modelling approach could simplify this assessment, thereby facilitating protection efforts. Previous research has largely focused on modelling spring occurrence and its predictors, rather than ecological value. This highlights the need for a new approach to modelling the ecological quality of springs. The aim was to determine whether structural parameters from the FOEN method can predict faunistic quality, and whether landscape and topographic variables can predict the structural and faunis tic quality of springs. An exploratory data analysis was performed on more than 750 Swiss springs structurally and faunistically assessed using the FOEN method. Additional potential predictor variables were generated in QGIS from publicly available landscape and topographic datasets. Various Ran dom Forest regression and classification models were implemented in R. Faunistic results were modelled using respective structural data, while both structural and faunistic results were modelled using landscape and topographic variables. Key predictors were identified through permutation-based variable importance measures. When modelling faunistic results using structure data, regression models showed similar per formance, with R² values never exceeding 0.25, indicating only weak predictive tendencies. Classification models reached a maximum accuracy of 0.43 with lower balanced accuracy, reflecting poor prediction of the lowest fauna classes. Therefore, accurate prediction of the fau nistic condition from the structural condition was not possible. Modelling structural and faunistic results with landscape and topographic data yielded better results for structural conditions, with the best regression model achieving an R² of 0.25, while the faunistic modelling did not exceed 0.15. Classification accuracy was substantially higher for structural modelling than for faunistic modelling, though balanced accuracy was similar. Overall, accurate prediction of ecologically valuable spring occurrence from landscape and topographic variables was not feasible. Across datasets, key predictor variables included Elevation, Number of Structures, Land Use, and Lithology, among others. This study demonstrates the application of Random Forest models with field-derived data to assess spring habitats. The findings indicate that successful modelling of the faunistic quality requires additional structural parameters not currently included in the FOEN method, highlight ing the need for a methodological expansion. Limitations include imbalanced class distributions dominated by highly rated springs. Future research should identify critical structural predictors that define the ecological value of springs and explore alternative modelling approaches, such as hybrid models or neural networks, to improve predictive performance.