Investigating the Sensitivity of Hydrological Model Calibration to Noise in Forcing and Response Variables

Abstract

In this study, the sensitivity of hydrological model calibration to noise in both input and output variables was investigated using the Soil and Water Assessment Tool (SWAT). Singular Spectrum Analysis (SSA) was utilised to remove noise fully as well as partially from the input data before using it for setting up the model for the lower Columbia River basin. The Wavelet Transform (WT) method was used for partially denoising the input data. It was observed that partially denoised datasets using the WT method yielded the highest statistical performance with performance metrics improving from 5% to 16% than the base model and the fully denoised model. A comparative analysis indicated that full denoising did not provide superior statistical model performance, whereas partial denoising achieved a more effective balance between reducing the extracted high-frequency components and preserving variability that contributed to model calibration. This research provides critical insights into input data preprocessing methodologies for hydrological modelling.