GSA Connects 2021 in Portland, Oregon

Paper No. 209-4
Presentation Time: 8:50 AM


BURNS, Erick1, MORDENSKY, Stanly1, LIPOR, John2, CURTIS, Jennifer A.3 and SANDO, Roy4, (1)U.S. Geological Survey, Geology, Minerals, Energy, and Geophysics Science Center, 2130 SW 5th Ave, Portland, OR 97201, (2)Electrical & Computer Engineering, Portland State University, Portland, OR 97201, (3)U.S. Geological Survey, 716 UNIT E W Cedar Street, Eureka, CA 95501, (4)U.S. Geological Survey, Wyoming-Montana Water Science Center, Bozeman, MT 59717

We propose two measures of hydrologic response and apply data-driven machine-learning strategies to understand the effects of geology on hydrology for approximately 192,000 km2 of the Northwest Volcanic Province, covering parts of California, Idaho, Nevada, and Oregon. Specifically, we investigate how: 1) precipitation-dependent spring density varies as recharge exceeds the capacity of aquifers to transmit groundwater; and 2) fast, intermediate, and slow components of streamflow provide information about geologically moderated flow paths. We use the following inputs: A) climate forcing provided as daily precipitation and minimum and maximum air temperature grids; B) soil-water holding capacity from soils databases; and C) a hydrology-themed interpretation of existing State-level geologic maps. From these inputs, we develop a simplified climate-soil relation model to predict annual available water (i.e., precipitation that becomes either groundwater or surface-water), providing the hydrologic forcing that is partitioned by the geology into spring and direct-runoff stream flow. In this way, the machine learning problem can be framed as a forcing-response problem with two filters: the air-temperature/soil filter translates daily precipitation into annual available water, and the geology filter translates available water into either spring density or annual stream-hydrograph components. Results broadly demonstrate that hypothesized patterns in hydrologic response are consistent with findings from a limited number of past focus studies within the region, allowing inference of regional permeability and hydrologic response in ungaged basins. Younger volcanogenic terranes tend to have higher primary permeability than older terranes and are associated with higher spring density and lower direct-runoff to streams. Fault-derived secondary permeability is variably important across the tectonically heterogeneous landscape, and in older geologic terranes, secondary permeability controls much of the basin-scale groundwater flow. Because recent faulting is common in young volcanic terranes (i.e., faulting and rock age are correlated), linear machine learning methods may not reliably fully separate the influence of each mechanism independently. For these cases, non-linear branching methods of regression (e.g., XGBoost) can help identify subtleties within the features (i.e., input data).