[1] Previous observations and modeling studies of Lake Superior have only partly elucidated its large-scale circulation, in terms of both the climatological state and interannual variability. We use an eddy resolving, three-dimensional hydrodynamic model to bridge this gap. We simulate Lake Superior circulation and thermal structure from 1979 to 2006 and consider the mechanisms responsible for the flow. Model results are compared to available direct observations of temperature and currents. Circulation in the lake is primarily cyclonic during all seasons, and a two-gyre structure is sometimes present. Surface circulation patterns in winter mimic wind directions but become organized in summer by the presence of thermal gradients. On the annual mean, nearshore currents are controlled by thermal gradients, while offshore flow is primarily determined by the wind. From a uniform bathymetry simulation, we determine that topographic variations cause small-scale structures in the open lake flow and are critical to the development of nearshore-offshore temperature gradients. The lake exhibits significant variability in current speed and direction on synoptic time scales, but coherent patterns of interannual variability are not found. Long-term trends due to changing meteorological forcing are found. Model results suggest the increase in lake surface temperature (0.37°C/decade) is significantly correlated to increases in wind speed above the lake (0.18 m/s/decade), increased current speeds (0.37 cm/s/decade), and declining ice coverage (−886 km 2 /yr).

Free-water measurements of dissolved oxygen (DO) in lakes are becoming common and provide opportunities for estimating ecosystem processes, such as gross primary production (GPP) and ecosystem respiration (R). The models used to estimate metabolism often subsume biological processes into one parameter each for GPP and R. However, high-frequency DO observations made over days show diverse patterns at multiple time scales, suggesting a complex suite of processes controls DO dynamics. Can we improve metabolism estimates and predictive ability for DO at diel scales by adding complexity to the models? In this study, we use data from two north temperate lakes to test a variety of metabolism models representing a suite of non-linear metabolic processes. To test whether alternative models can be discriminated, we simulated DO with assumed parameter values and auto-regressive noise, and fit the models to the simulated DO. The most complex model could be discriminated from simpler models and provided the most accurate and precise predictions. However, when models were fit to observed DO data from the sensor network, the simplest model predicted DO as well as the most complex one. The added complexity did not improve model performance. An analysis of the model residuals indicates that physics may explain some of the DO pattern not predicted, especially high-frequency oscillations and anomalies that appear to coincide with weather patterns. Under reasonably stable weather conditions and at scales of a few days, simple metabolism models explain the bulk of diel DO variability.

East Asian regions in the North Pacific have recently experienced severe riverine flood disasters. State-of-the-art neural networks are currently utilized as a quick-response flood model. Neural networks typically require ample time in the training process because of the use of numerous datasets. To reduce the computational costs, we introduced a transfer-learning approach to a neural-network-based flood model. For a concept of transfer leaning, once the model is pretrained in a source domain with large datasets, it can be reused in other target domains. After retraining parts of the model with the target domain datasets, the training time can be reduced due to reuse. A convolutional neural network (CNN) was employed because the CNN with transfer learning has numerous successful applications in two-dimensional image classification. However, our flood model predicts time-series variables (e.g., water level). The CNN with transfer learning requires a conversion tool from time-series datasets to image datasets in preprocessing. First, the CNN time-series classification was verified in the source domain with less than 10% errors for the variation in water level. Second, the CNN with transfer learning in the target domain efficiently reduced the training time by 1/5 of and a mean error difference by 15% of those obtained by the CNN without transfer learning, respectively. Our method can provide another novel flood model in addition to physical-based models.

No abstract