Integration of machine learning and particle filter approaches for forecasting soil moisture

A Physics‐Aware Machine Learning‐Based Framework for Minimizing Prediction Uncertainty of Hydrological Models

“…For brevity, the histogram of model parameters resulting from residual resampling technique only has been illustrated in Figure 4. The distribution of the model parameters appears to be Gaussian, exemplifying a reduction in the prediction uncertainty arising from the model parameters (Figure 4) (Sen et al., 2020; Tandon et al., 2022). Besides the Gaussian distribution of model parameters, a narrow variation of the model parameter range across the mean (i.e., a narrow spread) also advocates reduced parametric uncertainty (Sen et al., 2020; Tandon et al., 2022).…”

“…Apart from the advantage of estimating the PI, the PF technique also simultaneously quantifies the model parameter uncertainty resulting from the PI. This provides valuable information like the percentage of observed streamflow values encompassed within the PI, the deviation of the ensemble mean from the observation, the over or under‐estimation of the streamflow, and the symmetry around the observation (Sen et al., 2020; Tandon et al., 2022). The PF technique, however, suffers from the drawback of particle degeneracy (please refer to the Appendix for details on particle degeneracy), which was reduced by selecting the appropriate resampling technique (Boli'cboli´c et al., 2004; Elfring et al., 2021; Kuptametee & Aunsri, 2022; T. Li et al., 2015).…”

“…Please note that these Bayesian‐based data assimilation techniques are methods that utilize the knowledge of the prior distribution of the data (or observations) to estimate their posterior distribution (Wikle & Berliner, 2007). While the data assimilation techniques can be configured for quantifying any specific or combination of different sources of uncertainty (Abbaszadeh et al., 2019; Pathiraja et al., 2018), several studies only considered quantifying the parametric uncertainty ignoring other sources (Y. Liu et al., 2012; Sen et al., 2020; Tandon et al., 2022; Weerts & El Serafy, 2006). This might possibly be due to the multitude of degrees of freedom arising from different sources of uncertainty which makes the computation complex.…”

A Novel Physics‐Aware Machine Learning‐Based Dynamic Error Correction Model for Improving Streamflow Forecast Accuracy

“…Graphical assessment was made by fitting the parameters in the form of a histogram and the distribution curve using the Kernel density estimation. The distribution appeared to be Gaussian, and the lesser variance (or spread) indicated the significant reduction in uncertainty arising from the model parameters (Sen et al, 2020;Tandon et al, 2022).…”

“…Despite several attempts being made to improve the forecast accuracy of streamflow at longer lead times, several practical challenges still remain. These challenges may vary from understanding and modeling the complex underlying physical processes to quantifying the uncertainty induced by the input, model structure, and parameter (Kasiviswanathan et al., 2013, 2016; Kurian et al., 2020; Liu et al., 2015; Pathiraja et al., 2016; Singh et al., 2020; Tajiki et al., 2020; Zhao et al., 2011, 2013).…”