Aalipour et al. (2026) Effect of landscape composition on catchment flow components across Germany using machine learning

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-01-06
• Authors: Mehdi Aalipour, Mirhossein Mousavinezhad, Naicheng Wu, Ali Torabi Haghighi, Nicola Fohrer, Bahman Jabbarian Amiri
• DOI: 10.1016/j.ejrh.2025.103058

Research Groups

• Department of Geography and Spatial Information Techniques, Ningbo University, Ningbo, China
• Department of Civil Engineering, Faculty of Engineering, University of Mohaghegh Ardabili, Ardabil, Iran
• Water, Energy, and Environmental Engineering Research Unit, University of Oulu, Oulu, Finland
• Department of Hydrology and Water Resources Management, Institute of Natural Resource Conservation, Kiel University, Kiel, Germany
• Department of Regional Economics and Environment, Faculty of Economics and Sociology, University of Lodz, Lodz, Poland

Short Summary

This study utilized machine learning algorithms to analyze how landscape composition (land use/land cover and soil characteristics) influences the parameters of the conceptual Tank model across 30 German catchments. The findings reveal that soil-related variables, particularly texture and moisture content, along with specific land use types like mixed forests and transitional woodland-shrub, are the primary controls on hydrological responses, with the Random Forest model demonstrating the most robust predictive performance.

Objective

• To determine how and to what extent the landscape composition, including land use/land cover (LULC) and soil-related variables, explains the variation in different catchment flow components, thereby improving the understanding of rainfall-runoff processes at the catchment scale.

Study Configuration

• Spatial Scale: 30 river catchments in Germany, with areas ranging from 53 km² to 737 km².
• Temporal Scale: Tank model parameters were derived from long-term daily hydrological and meteorological data (from Amiri et al., 2016). Spatial input data for machine learning were from 2000, 2020, and 2023.

Methodology and Data

• Models used:
  • Hydrological Model: Tank model (four-tank version)
  • Machine Learning Models: Boosted Regression Trees (BRT), Support Vector Machines (SVM), Random Forests (RF)
• Data sources:
  • Tank model outputs: Calibrated parameters from Amiri et al. (2016).
  • Catchment characteristics: Catchment boundary, area, hydrological density slope, and main river length from HydroATLAS (2023).
  • Land Use/Land Cover (LULC): Corine Land Cover map (2000) at 100 m resolution.
  • Soil-related variables: Soil texture (clay, silt, sand content), soil organic carbon (SOC), and water content from Poggio et al. (2021) and Turek et al. (2023) at 250 m resolution (data from 2020).

Main Results

• The Random Forest (RF) model demonstrated the most reliable and robust predictive performance for regionalizing tank model parameters, with r² values ranging from 0.86 to 0.95 in the testing phase, outperforming Boosted Regression Trees (0.59 ≤ r² ≤ 0.99) and Support Vector Machines (0.33 ≤ r² ≤ 0.91).
• Soil-related variables were identified as major controls on hydrological responses:
  • Soil texture (0.01–14.7 %) and soil moisture content (0.05–19.44 %) at different depths significantly improved prediction accuracy. Silt content at 15–30 cm (14.7 % importance for infiltration coefficient a0) and 5–15 cm (12.4 % for a0) were key predictors.
  • Soil organic carbon (SOC) (0.05–8.86 %) was crucial for water retention capacity and soil permeability, influencing infiltration and runoff behavior, particularly in intermediate layers (SOC 0–15 cm with 8.86 % importance).
  • Soil moisture content at 100–200 cm depth (WC10P 100–200) was highly influential (~8.5 % for a0, ~19 % for water storage level Ha2), highlighting the role of subsurface water availability.
• Land Use/Land Cover (LULC) types considerably affected flow components:
  • Mixed forests (0.22–27.35 %) and transitional woodland-shrub (0.84–33.92 %) significantly influenced subsurface flow and water storage levels. Transitional woodland-shrub showed very high importance (>30 % for sub-surface flow a1, base flow d1, and sub-base flow water storage level Hc1).
  • Agricultural land (0.082–11.28 %) affected base flow (d1) and water storage levels (Ha1, Hb1).
  • Pastures (0.34–8.05 %) contributed to reducing surface runoff (a2).
• Catchment characteristics also played a role:
  • Catchment area (0.2–16.98 %), main river length (~22.9 % for percolation coefficient b0), and slope (~7.5 % for intermediate flow b1) influenced tank model parameters.

Contributions

• This study bridges the gap between conceptual rainfall-runoff models (specifically the Tank model) and spatial analysis by integrating machine learning techniques to understand spatial-hydrological interactions.
• It quantifies the influence of landscape composition (LULC and soil properties) on the variability of different catchment flow components, enhancing the physical understanding of hydrological model parameters.
• It provides new hydrological insights for Germany by identifying specific soil texture, soil moisture, soil organic carbon, and LULC types (e.g., mixed forests, transitional woodland-shrub) as primary drivers of hydrological responses.
• The research demonstrates the effectiveness and robustness of the Random Forest model for regionalizing hydrological parameters, offering a valuable tool for more precise and spatially informed hydrological forecasts, particularly in data-limited regions.

Funding

• National Natural Science Foundation of China (No. 52279068)
• Internal research competitions of the University of Lodz (grant increased by 2 % for universities in the Excellence Initiative - Research University competition in Poland)
• Alexander von Humboldt Stiftung in Germany

Citation

@article{Aalipour2026Effect,
  author = {Aalipour, Mehdi and Mousavinezhad, Mirhossein and Wu, Naicheng and Haghighi, Ali Torabi and Fohrer, Nicola and Amiri, Bahman Jabbarian},
  title = {Effect of landscape composition on catchment flow components across Germany using machine learning},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2025.103058},
  url = {https://doi.org/10.1016/j.ejrh.2025.103058}
}