Enhanced streamflow prediction with SWAT using support vector regression for spatial calibration: A case study in the Illinois River watershed, U.S.

📝 Summary

This study presents a hybrid modeling approach combining the Soil and Water Assessment Tool (SWAT) with Support Vector Regression (SVR) to improve monthly streamflow prediction in the Illinois River Watershed (IRW), USA. Traditional SWAT calibration using SWAT-CUP is time-consuming and often inaccurate in dry seasons or ungauged watersheds. The new SWAT-SVR model uses SWAT outputs and drainage area as SVR inputs, bypassing heavy calibration and enabling more accurate spatial predictions. The model showed better performance in the wet season and for medium streamflows (5–30 m³/s), with applicability for watersheds ranging from 500 to 3000 km². This article can be accessed from https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0248489. 

⭐ Highlights

1. Problem Statement

• Rainfall-runoff relationships are complex and nonlinear.

• Existing models (conceptual, physically-based, or data-driven) each have limitations.

• SWAT is powerful but calibration is data-intensive and time-consuming.

• SVM (SVR) can handle nonlinear patterns but needs sufficient data.

2. Solution

• Developed a hybrid SWAT-SVR model to combine the strengths of physical modeling (SWAT) and machine learning (SVR).

• Used streamflow from SWAT and upstream drainage area as SVR inputs.

• Applied spatial calibration and leave-one-out validation for 13 USGS stations in the IRW.

3. Model Performance

• SWAT-SVR outperformed SWAT-CUP, especially in the dry season.

• SWAT-CUP failed in dry-season prediction (all simulations were rated "Unsatisfactory").

• SWAT-SVR had:

  1. Good prediction of medium flows (5–30 m³/s),

  2. Poor prediction of high flows (>30 m³/s) due to insufficient data for training,

  3. Satisfactory to Very Good performance for 67% of stations in annual simulations.

4. Spatial Suitability

• Best performance observed for basins with drainage areas between 500–3000 km².

• Flow duration curves showed good match in medium and low flows.

5. Efficiency

• SWAT-SVR reduces calibration effort: only three SVR parameters (C, γ, ε) to tune.

• Useful in ungauged or limited-data watersheds by borrowing spatial information from proxy sites.

6. Conclusion

• The SWAT-SVR hybrid model is promising for monthly streamflow prediction in watersheds with:

  • Medium size (500–3000 km²),

  • Medium flows (5–30 m³/s),

  • Limited gauging data.

• Offers a data-efficient alternative to conventional SWAT calibration methods.

• Particularly beneficial for regional applications and ungauged basin studies.