Summary of  work 

Value of process understanding in the era of machine learning: A case for recession flow prediction

The use of machine learning (ML) models in the field of hydrology is not something new. Since the early 1990s, ML models such as Artificial Neural Networks (ANN) and Support Vector Machines (SVM) have been employed in rainfall-runoff modeling. However, recently, advanced ML models (like LSTM) have shown superior performance compared to well-established hydrological models in daily streamflow prediction. This has led researchers to question whether we truly need an understanding of hydrological processes in this era of machine learning. Nonetheless, predictions made by ML models often lack physical consistency. In this study, we attempt to answer the question of whether a physical process understanding in hydrology is necessary. We have chosen a conceptual model, Power Law Regression (PLR), and two machine learning models, ANN and Long Short-Term Memory (LSTM), to predict recession flow based solely on past streamflow information. Additionally, we have proposed process-informed versions of the ML models, PI-ANN and PI-LSTM. Our results suggest that a simple conceptual model developed with a clear understanding of physical processes can outperform advanced ML models like LSTM. Furthermore, we demonstrate that a basic ML model like ANN can compete effectively when incorporating process understanding. This study provides a timely contribution to the field of hydrological science, highlighting the importance of process understanding."

A Canberra distance-based complex network classification framework using lumped catchment characteristics

Modelling hydrological processes is challenging due to process complexity and limited availability of observed data. This makes it more complex in case of ungauged basins where observed data is not available. The traditional solution is to transfer the information from gauged to ungauged basins called ‘regionalization’. However, the success of a regionalization method depends mainly the degree of similarity between the gauged catchments and the ungauged catchments. Recent studies use the concept of complex network to classify hydrologically similar catchments for gauged catchments. We introduce Canberra distance metric to construct complex networks. This enables the use of complex networks for catchment classification for ungauged basins. We compare the results with commonly used K-means clustering algorithm. This study is carried out for 494 USA catchments.