Machine Learning Methods for Modeling Streamflow in Intermittent Rivers and Ephemeral Streams

The Intermittent rivers and ephemeral streams (IRES) are the most widespread elements of the global surface water network and there is a growing reliance on them for a variety of hydro-ecological services. Including IRES in short- and long-term water planning and management endeavors requires appropriate tools that are capable of modeling the unique characteristics of these streams and providing reliable information about their aquatic states (flow or no-flow) and flow dynamics. Considering the limitations of physics-based and water balance methods in modeling flow discontinuities, machine learning techniques are viewed as promising, flexible, and powerful tools to simulate the nonlinear and highly variable dynamics of IRES flow. In this dissertation, innovative data-driven methodologies were developed to enhance the predictive ability of machine learning streamflow models for capturing the entire IRES flow spectrum. In the first study, a sequence of Bayesian-inspired classifiers that progressively relax the assumptions made during the application of Bayes’ theorem are developed and compared against the classic Naïve Bayes Classifier with respect to their predictive performance as hydro-climatic models in capturing IRES aquatic state series. This study answered the following questions: Are Bayesian classifiers appropriate for modeling IRES aquatic states? What are the impacts of the potential violations of the naïve assumptions (e.g., conditional independence and the blanket use of Gaussian marginals) on the performance of Bayesian classifiers? How much complexity is warranted for modeling aquatic states in IRES systems using Bayesian-Inspired approaches? Two headwater IRES in Texas were chosen as testbeds: one with majority no-flow states and another with majority flow states. The testbeds were also used to develop two sets of block bootstrapping synthetic streamflow sets for further assessment over a range of intermittencies. According to the results, models that explicitly capture at least a portion of the correlation among the input features, provide better estimations of the likelihoods associated with the aquatic states (as measured by PBIAS, NSE, and RSR metrics) which in turn lead to better estimation of node state dynamics (evaluated using the theory of runs and Markov Chains). Copula theory provided a convenient approach to capture conditional dependence among inputs and its integration with Bayes’ theorem proved beneficial. In the second study, an innovative IRES flow modeling framework, referred to as the Classification Regression Intermittent Streamflow Prediction (CRISP) is developed. The CRISP framework builds from the fact that different hydrological processes control the flow regimes in IRES under flow and no-flow conditions. Therefore, the streamflow is best conceptualized as arising from a mixture distribution. A generative Bayesian Classifier based on multivariate Gaussian distribution (CGJD-BC) was used to model the discrete (flow and no flow regimes) part and a state-of-the-art Self-Attention Long-Short Term Memory (SA-LSTM) network was used to predict non-zero flow rates. The proposed algorithm was used for predicting monthly IRES flow using meteorological fluxes (precipitation and evaporation) and compared to a SA-LSTM model that was founded on the assumption that streamflow data came from a single underlying distribution (the common practice). The proposed CRISP framework exhibited superior predictions with respect to a suite of existing and new performance evaluation metrics for IRES with various intermittency ratios. Furthermore, the SA-LSTM model based on the continuum assumption was unable to truly predict any zero flow in either case and sometimes led to unrealistic non-zero flow predictions. The CRISP model provides a better alternative to model flows in temporary rivers.

Embargo status: Restricted until 01/2024. To request the author grant access, click on the PDF link to the left.