Effect of landscape composition on catchment flow components across Germany using machine learning

The study was conducted on the basis of spatial information from 30 river catchments in Germany. The tank model, as one of the conceptual rainfall-runoff models, has been widely used to understand the dynamics of rainfall-runoff interactions for hydrological modelling. Determining the influence of spatial catchment variables on the variability of model parameters is a significant challenge in hydrological modelling. This study compares three machine learning algorithms, including Boosted Regression Trees, Support Vector Machines, and Random Forests, to analyze the effect of catchment variables on the parameters of the tank models. The study was conducted based on spatial information using land use/land cover, soil related characteristics, and catchment characteristics. The results showed that soil related-variables, especially texture (0.01–14.7 %) and moisture content (0.05–19.44 %) at different soil depths, mainly controls the hydrological responses and largely improve the accuracy of predictions. Soil organic carbon (0.05–8.86 %) was important in defining the retention capacity of water and soil permeability and, hence, infiltration and runoff behavior. Land use/Land cover types, especially mixed forests (0.22–27.35 %) and transitional woodland-shrub (0.84–33.92 %), considerably affected subsurface flow and water storage levels, showing the importance of vegetation in controlling water flow and storage. Moreover, results indicated that the RF model proved the most reliable and robust method, providing higher accuracy (0.86 ≥ r 2 ≤ 0.95), generalization, and the ability to capture complicated nonlinear relationships in data comparing to Boosted Regression Trees (0.59 ≥ r 2 ≤0.99), Support Vector Machines (0.33 ≥ r 2 ≤ 0.91). • The study emphasizes the high potential of machine learning algorithms in enhancing hydrological predictions and water resource management. • RF algorithm can capture complex nonlinear relationships governing hydrological process-response. • Silt, mixed forests, transitional woodland-shrub, and soil moisture content were key in predicting water flow and storage within catchments.