Geospatial patterns in runoff projections using random forest based forecasting of time-series data for the mid-Atlantic region of the United States

Brandi Gaertner

doi:10.1016/j.scitotenv.2023.169211

Geospatial patterns in runoff projections using random forest based forecasting of time-series data for the mid-Atlantic region of the United States

Sci Total Environ. 2024 Feb 20:912:169211. doi: 10.1016/j.scitotenv.2023.169211. Epub 2023 Dec 12.

Author

Brandi Gaertner¹

Affiliation

¹ The Pennsylvania State University, 2217 Earth and Engineering Sciences Building, University Park, PA 16802, United States. Electronic address: BAH6009@psu.edu.

PMID: 38097071
DOI: 10.1016/j.scitotenv.2023.169211

Abstract

This research explores the geospatial patterns of historical runoff for the period 1958-2021 in the Mid-Atlantic region and uses these time-series data plus nine external climatic and hydrologic variables to predict future runoff for the period 2022-2031. Gridded, average monthly climatic water balance data were obtained from the TerraClimate dataset. A cluster analysis of the long term (1958-2021) historical runoff found 13 significant temporal trends, which tend to form large contiguous regions associated with climate gradients and topographic patterns. The runoff time-series clusters, and the associated time-series of 9 TerraClimate variables, were used to generate random forest based forecast models to predict future (2022-2031) runoff. The random forest-based forecast with the greatest accuracy included inputs of actual evapotranspiration, climate water deficit, minimum, average, and maximum temperature, and vapor pressure deficit. The final model predicted significantly increasing runoff in nine of the 13 clusters.

Keywords: Climate change; Climate water deficit; Evapotranspiration; Forest-based regression; Hydrology; Machine learning; Temperature; Time series cluster analysis; Vapor pressure deficit.