Scott Steinschneider’s research while affiliated with Cornell University and other places

California faces cycles of drought and flooding that are projected to intensify, but these extremes may impact water users across the state differently due to the region's natural hydroclimate variability and complex institutional framework governing water deliveries. To assess these risks, this study introduces a novel exploratory modeling framework informed by paleo and climate‐change based scenarios to better understand how impacts propagate through the Central Valley's complex water system. A stochastic weather generator, conditioned on tree‐ring data, produces a large ensemble of daily weather sequences conditioned on drought and flood conditions under the Late Renaissance Megadrought period (1550–1580 CE). Regional climate changes are applied to this weather data and drive hydrologic projections for the Sacramento, San Joaquin, and Tulare Basins. The resulting streamflow ensembles are used in an exploratory stress test using the California Food‐Energy‐Water System model, a highly resolved, daily model of water storage and conveyance throughout California's Central Valley. Results show that megadrought conditions lead to unprecedented reductions in inflows and storage at major California reservoirs. Both junior and senior water rights holders experience multi‐year periods of curtailed water deliveries and complete drawdowns of groundwater assets. When megadrought dynamics are combined with climate change, risks for unprecedented depletion of reservoir storage and sustained curtailment of water deliveries across multiple years increase. Asymmetries in risk emerge depending on water source, rights, and access to groundwater banks.

This study proposes a novel hybrid method that substantially accelerates and improves deep learning (DL) model development for streamflow prediction. The method leverages a combination of a long short-term memory (LSTM) network and random forests. A hybrid encoder-decoder model is designed, where a pre-trained LSTM is utilized as an encoder to extract temporal features from the input data. Subsequently, the random forest decoder processes the encoded information to make streamflow predictions. Our method was tested on 421 catchments in the continental United States and 324 in Germany, both selected from two CAMELS datasets. The hybrid method has several benefits. First, it is much more efficient and robust than training LSTMs on each catchment individually (~14x faster). Second, it is much less computationally expensive than LSTM fine-tuning (i.e., feasible on a CPU-based workstation). Third, it achieves superior accuracy compared to a catchment-wise training strategy (e.g., 9.2 % improvement in the median in Nash-Sutcliffe Efficiency (NSE)), shows competitive performance compared to regional LSTM models when trained with fewer data, and through fine-tuning, improves regional LSTM performance in out-of-training samples by 13.13 % (median NSE). To our knowledge, this is the first decision-tree model integrated within a DL workflow to enhance fine-tuning efficiency of pre-trained models in new locations. This hybrid approach holds significant promise for future applications in hydrological modeling, particularly considering the imminent rise of geospatial foundation models in hydrology that will rely on transfer learning techniques for effective deployment.

This study proposes a novel hybrid method that substantially accelerates and improves deep learning (DL) model development for streamflow prediction. The method leverages a combination of a long short-term memory (LSTM) network and random forests. A hybrid encoder-decoder model is designed, where a pre-trained LSTM is utilized as an encoder to extract temporal features from the input data. Subsequently, the random forest decoder processes the encoded information to make streamflow predictions. Our method was tested on 421 catchments in the continental United States and 324 in Germany, both selected from two CAMELS datasets. The hybrid method has several benefits. First, it is much more efficient and robust than training LSTMs on each catchment individually (~14x faster). Second, it is much less computationally expensive than LSTM fine-tuning (i.e., feasible on a CPU-based workstation). Third, it achieves superior accuracy compared to a catchment-wise training strategy (e.g., 9.2 % improvement in the median in Nash-Sutcliffe Efficiency (NSE)), shows competitive performance compared to regional LSTM models when trained with fewer data, and through fine-tuning, improves regional LSTM performance in out-of-training samples by 13.13 % (median NSE). To our knowledge, this is the first decision-tree model integrated within a DL workflow to enhance fine-tuning efficiency of pre-trained models in new locations. This hybrid approach holds significant promise for future applications in hydrological modeling, particularly considering the imminent rise of geospatial foundation models in hydrology that will rely on transfer learning techniques for effective deployment.

This study proposes a novel hybrid method that substantially accelerates and improves deep learning (DL) model development for streamflow prediction. The method leverages a combination of a long short-term memory (LSTM) network and random forests. A hybrid encoder-decoder model is designed, where a pre-trained LSTM is utilized as an encoder to extract temporal features from the input data. Subsequently, the random forest decoder processes the encoded information to make streamflow predictions. Our method was tested on 421 catchments in the continental United States and 324 in Germany, both selected from two CAMELS datasets. The hybrid method has several benefits. First, it is much more efficient and robust than training LSTMs on each catchment individually (~14x faster). Second, it is much less computationally expensive than LSTM fine-tuning (i.e., feasible on a CPU-based workstation). Third, it achieves superior accuracy compared to a catchment-wise training strategy (e.g., 9.2 % improvement in the median in Nash-Sutcliffe Efficiency (NSE)), shows competitive performance compared to regional LSTM models when trained with fewer data, and through fine-tuning, improves regional LSTM performance in out-of-training samples by 13.13 % (median NSE). To our knowledge, this is the first decision-tree model integrated within a DL workflow to enhance fine-tuning efficiency of pre-trained models in new locations. This hybrid approach holds significant promise for future applications in hydrological modeling, particularly considering the imminent rise of geospatial foundation models in hydrology that will rely on transfer learning techniques for effective deployment.

The risk of compound coastal flooding in the San Francisco Bay Area is increasing due to climate change yet remains relatively underexplored. Using a novel hybrid statistical-dynamical downscaling approach, this study investigates the impacts of climate change induced sea-level rise and higher river discharge on the magnitude and frequency of flooding events as well as the relative importance of various forcing drivers to compound flooding within the Bay. Results reveal that rare occurrences of flooding under the present-day climate are projected to occur once every few hundred years under climate change with relatively low sea-level rise (0.5 m) but would become annual events under climate change with high sea-level rise (1.0 to 1.5 m). Results also show that extreme water levels that are presently dominated by tides will be dominated by sea-level rise in most locations of the Bay in the future. The dominance of river discharge to the non-tidal and non-sea-level rise driven water level signal in the North Bay is expected to extend ~15 km further seaward under extreme climate change. These findings are critical for informing climate adaptation and coastal resilience planning in San Francisco Bay.

Compound coastal flooding due to astronomic, atmospheric, oceanographic, and hydrologic drivers poses severe threats to coastal communities. While physics-driven approaches are able to dynamically simulate temporally and spatially varying compound flooding generated by multiple drivers with correlations between some of them, computational burdens limit their capability to explore the full range of conditions that contribute to compound coastal hazards. Data-driven statistical approaches address some of these computational challenges; however, they are also unable to explore all possible forcing combinations due to short observational records, and projections are typically limited to a few locations. This study proposes a hybrid statistical–dynamical framework for compound coastal flooding analysis that integrates a stochastic generator of compound flooding drivers, a hydrodynamic model, and machine learning-based surrogate models. The framework was demonstrated in San Francisco Bay (SF) over the past 500 years with accuracy similar to the physics-driven approach but with much higher computational efficiency. The stochastic generator of compound flooding drivers is developed by coupling a sea surface temperature (SST) reconstruction model with a climate emulator, weather generator, and model of the hydrological and reservoir system. Using reconstructed SSTs as input, the generator of compound flooding drivers is employed to simulate time series of the forcing factors contributing to compound flooding (e.g. surge, waves, river discharge, etc) in SF Bay. A process-based hydrodynamic model is built to predict total water levels varying in time and space throughout SF Bay based on stochastically generated drivers. The machine learning-based surrogate models are then developed from a relatively small library (several hundred) of hydrodynamic model simulations to efficiently predict water levels for compound flooding analysis under the full range of stochastic drivers. This study contributes a hybrid statistical–dynamical framework to better understand the spatial distribution and temporal evolution of compound coastal-fluvial flooding, along with the relative contributions of drivers in complex nearshore, estuarine, and river environments for centennial timescales under past, present, and future climates.

The threat that climate change poses to water resource systems has led to a substantial and growing number of impact studies. These studies follow two approaches: (a) top‐down studies are driven by projections of future climate change provided by downscaled general circulation models (GCMs); and (b) bottom‐up studies are driven by the systematic evaluation of exploratory scenarios. Top‐down approaches produce realistic scenarios rooted in the simulation of thermodynamic and dynamic processes represented in GCMs, but the internal resolution of these processes make it difficult to link vulnerabilities to discrete components of change. Bottom‐up approaches link vulnerabilities to discrete components of change through the structured evaluation of exploratory scenarios, but the lack of insight rooted in climate change processes can lead to the development of implausible scenarios. This paper evaluates exploratory scenarios developed through thermodynamic and dynamical guided perturbations motivated by GCM‐bound insights. The result is a hybrid approach that bridges a gap between top‐down and bottom‐up approaches. This yields several advantages. First, emerging vulnerabilities are linked to distinct thermodynamic and dynamic processes that are modeled in GCMs with differential likelihoods and plausible ranges of change. Second, the structured evaluation of process‐informed exploratory scenarios link system vulnerabilities to distinct components of climate change. An illustrative case study demonstrates perturbations linked to thermodynamic and dynamical processes have a large impact on stakeholder‐defined flood and drought performance, and the structured evaluation of process‐informed exploratory scenarios find nuanced infrastructure‐specific vulnerabilities that would be difficult to identify using an alternative approach.