Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Master → Etats Unis → 2020 → Statistical precipitation downscaling model for Southern California developed using random forest : Evaluation of performance and impacts on hydrologic metrics

San Diego State University (2020)

Statistical precipitation downscaling model for Southern California developed using random forest : Evaluation of performance and impacts on hydrologic metrics

Flint, Kelly M

Titre : Statistical precipitation downscaling model for Southern California developed using random forest : Evaluation of performance and impacts on hydrologic metrics

Auteur : Flint, Kelly M

Université de soutenance : San Diego State University

Grade : Master of Science (MS) 2020

Résumé
Projected climate changes in southern California are likely to induce feedback from natural systems, including alterations to existing flow regimes, which can disrupt aquatic and riparian ecosystems and negatively impact native species. Hydrologic metrics used to assess such alterations are often estimated from streamflow simulated with rainfall-runoff models. Resolutions of existing precipitation datasets used to initialize runoff generation in rainfall-runoff models may be too coarse to capture sub-kilometer and sub-daily variations in precipitation occurrence and intensity typical of precipitation regimes in southern California. This research uses a Random Forest algorithm to downscale 3-hourly, 2x2.5-degree atmospheric outputs from NASA GISS Model E2-R to 3-hourly, 250-meter precipitation estimates for southern California. Hourly observations for years 1985 – 2012 (28 years) obtained from more than 700 precipitation gauges were used to train the algorithm. Daily and 3-hourly versions of the spatially downscaled precipitation dataset and PRISM daily precipitation totals were used to initialize HEC-HMS models for three case studies to evaluate performance in complex terrain. Simulated streamflow and subsequently derived hydrologic metrics were compared to observed streamflow values ; streamflow simulated with PRISM precipitation totals served as a benchmark for relative performance evaluation. The resulting statistical downscaling model tended to overestimate low- to moderate intensity precipitation and underestimate high intensities ; however, it showed improved precipitation estimates when compared to PRISM at selected gauges. Accuracy of models calibrated with the downscaled dataset exhibited slight, but significant improvement on those calibrated with PRISM. While both datasets underestimated streamflow magnitudes, the downscaled precipitation produced inaccurate dry conditions less often. Temporal resolution had limited impact on simulated streamflow, except in the flashy and ephemeral system, for which 3-hourly precipitation improved model performance. Continued efforts to improve precipitation estimates will provide more accurate depictions of baseline conditions and projected change for well-informed mitigation and adaptation strategies within the realm of water resources.

Présentation et version intégrale

Page publiée le 22 décembre 2021