New Mexico Geological Society Annual Spring Meeting — Abstracts

Joint low flow events across multiple watersheds can have far reaching consequences for agriculture, environment, and human communities. However, predicting the spatial joint occurrence of low streamflow remains a challenging task due to the number of factors affecting streamflow, ranging from meteorology to watershed characteristics.

This study investigates the spatial correlation patterns of hydrological drought across a large number of instrumented watersheds across the Conterminous United States, with a focus on developing predictive models for joint drought probabilities. We base our analysis on a large dataset (Gages-2) which includes historical streamflow observations, climatic variables, and basin physical characteristics. At each site, we stochastically simulate long streamflow time series data by randomizing the phase of observed streamflow in the wavelet domain. By means of this approach, we preserve spatial and temporal correlation of observed streamflow and at the same time obtain long simulations from which we infer the joint probability of crossing low-flow thresholds across station pairs

To further predict spatial low flow occurrence across watersheds, we propose a predictive Machine learning model by comparing the predictive capability of tree-based models (XGBoost and Random Forest) with a base Multiple Linear Regression model. This model can be used to predict the joint probability of drought events, learning from both climatic variables and basin characteristics as predictors. Our preliminary results suggest that a technique based on XGBoost is the most robust approach to solve this spatial prediction problem.