Strategic deep‐water observations enhance probabilistic parameter estimation of lake hydrodynamic models

Calibration of physics-based hydrodynamic models of lake water temperature often faces challenges including limited observational data and parameter equifinality, that is, many parameterizations yield similar goodness-of-fit to observations. This study presents a framework to investigate how the number and location of observation depths used in calibration influence temperature prediction accuracy and parameter equifinality. We leverage a computationally efficient one-dimensional physics-based model (GLM; General Lake Model), and apply probabilistic parameter estimation within a Generalized Likelihood Uncertainty Estimation framework. Using 7 years of hourly water temperature data from Lake Mendota (WI, USA), we quantify parameter equifinality and assess predictive performance using multiple metrics across single-depth, dual-depth, and full-profile observation. Results show that single-depth deep-water observations consistently outperform shallower depths and sometimes even full-profile observations, significantly reducing equifinality and achieving accurate match throughout the water column. Contrary to expectations, full-profile data do not consistently yield superior predictive performance across the water column. Instead, they dilute the influence of deep-water observations on parameter selection. Dual-depth scenarios show that single-depth surface mixed layer observations can be vastly improved by adding a second observation depth. These findings provide practical examples for optimizing observational strategies and advancing lake hydrodynamic modeling. In particular, this work supports the management of drinking water reservoirs by informing the strategic selection of fewer, well-placed deep-water observations. In doing so, it focuses model development toward critical depths—for example, for drinking water reservoirs that commonly rely on deep-layer intakes.