Streamflow Forecasting at Ungauged Locations Using Deep Learning Networks, Driven by Soil Moisture States and Meteorological Data

Traditionally streamflow forecasting is accomplished using process-based hydrological models. These models could range from simple lumped type to more detailed distributed models. Lumped type models are easy to setup while distributed models require considerable skill and experience for setup. Due to the growing availability of large amounts of spatial and temporal data from various sources, such as remote sensing and re-analyses, and recent advances in the computational power, machine learning models are gaining momentum for solving applied engineering problems, triggering conceptual shifts perhaps led by rapid progress in data science and availability of ready-to-be-deployed software tools. These models have the ability to extract complex dynamical nonlinearities without explicitly defining the involved physical processes and their governing mathematical formulations, as is followed in the case of hydrological models. It is believed that new trends and conceptual shifts are essential for generating new knowledge, challenging or validating prevailing assumptions, and enhancing operational applications, which may include several water management-related functions, hydropower generation operations, and flood risk management across a range of temporal and spatial scales.

In this study, two deep learning variants of machine learning models, i.e., (1) the attention-based encoder-decoder bidirectional long short-term memory (AB-ED-BiLSTM) network and (2) the attention-based encoder-decoder bidirectional gated recurrent units (AB-ED-BiGRU) network, were tested on multiple watersheds selected from the Ottawa River Basin, Canada. After developing and successfully evaluating watershed-specific models, regional versions of both models were developed and tested based on the leave-one-watershed-out strategy to emulate an ungauged scenario. Both models were driven mainly by soil moisture states of watersheds and meteorological data in order to evaluate their usefulness for streamflow forecasting at ungauged locations. Although not as ideal as one would desire, these models demonstrated reasonable skill in forecasting streamflow with one to seven days lead time when assessed in terms of coefficient of determination, Nash-Sutcliff Efficiency, and Kling-Gupta Efficiency performance metrics. However, considerable discrepancies were noticed in simulating peak flow values for certain watersheds. Overall results of the study suggest that soil moisture driven machine learning models can potentially be used to develop streamflow forecasting tools for ungauged locations, with AB-ED-BiGRU being computationally an inexpensive option compared to the AB-ED-BiLSTM model. Additional investigations will be required to improve their performance further, e.g., by employing multiple soil moisture products, available through remote sensing and re-analyses sources, and ensemble modelling techniques. Based on continuous scientific progress, emerging machine learning frameworks and architectures, and better understanding of the origins and limitations of existing models, improved hydrological forecasting at ungagged locations can be made possible. In essence, this study contributes towards enhancing our understanding of the role of soil moisture in developing machine learning based streamflow modelling and forecasting tools to support operational applications at ungauged locations, which are often neglected when developing real-time streamflow forecasting systems.