Expecting surprise : towards information-based streamflow monitoring network evaluation through flow duration curve estimation in ungauged basins

Prediction in ungauged basins is fundamentally a problem of information transfer under uncertainty. This thesis applies information theory to estimating river discharge magnitude and frequency across the Pacific Northwest of North America, while testing foundational assumptions in broader global regions. It is motivated by two needs: reliable prediction of long-term streamflow distributions and information-based evaluation of monitoring networks. The thesis first establishes data and methodological foundations. A consistent dataset of catchment descriptors supports prediction in gauged and ungauged basins. A conceptual framework for automated, interactive quality documentation, Camel Farrier, supports assessment of environmental data products for their intended use. Streamflow distribution estimation is then examined for sensitivity to prior assumptions and procedural artifacts. Kernel density estimators are assessed as references for model comparison, and a fixed discrete quantization based on rating curve measurement uncertainty provides a consistent, pre-specified support. The thesis then evaluates out-of-sample flow duration curve prediction using parametric models based on catchment attributes, nonparametric donor transfer using k-nearest neighbours (kNN), and rainfall–runoff modelling using long short-term memory (LSTM) neural networks. Comparisons assess the benefit of richer inputs and greater model complexity, with upper and lower benchmarks providing consistent context for predictive skill. kNN and LSTM methods perform comparably across much of the sample, while simpler parametric models remain competitive in some settings. Low rank correlations among model errors suggest complementary information, and inter-model ensembles yield small performance gains. Finally, the thesis tests whether catchment attributes can predict pairwise divergence between daily discharge distributions. This experiment advances an approach to quantifying the marginal information value of monitoring locations by reframing network evaluation as minimizing uncertainty across the ungauged region served. The model outperforms a naive benchmark in just over half of evaluated cases, indicating modest predictive skill, while wide prediction intervals reflect substantial uncertainty. Future work must improve predictive skill and establish where extrapolation is reliable. Sparse sampling of location pairs may leave substantial portions of the ungauged region’s high-dimensional attribute space unrepresented, complicating assessment of prediction-interval calibration.