Lake: Characteristics, Modeling, and Management

• Lake is a naturally occurring or man-made inland water body notable for its variable geometry, hydrological interactions, and critical roles in ecological and socioeconomic systems.
• Advanced remote sensing and machine learning methods, such as Swin-Unet with MIoU ≈ 0.905, enable precise mapping and temporal monitoring of lake boundaries and morphometry.
• Operational management of lakes uses predictive control and optimization strategies, achieving effective flood mitigation, water supply reliability, and equitable human-lake interaction.

A lake is a naturally occurring or anthropogenic inland water body, typically characterized by a closed or nearly closed basin with spatially and temporally varying water levels, morphometry, and physicochemical properties. Lakes play critical roles in hydrological, ecological, climatic, and socioeconomic systems. They are foci of physical modeling in fluid mechanics, objects of large-scale remote sensing, drivers and recorders of biogeochemical cycles, and serve as crucial ecosystem and recreational resources. Contemporary research employs advanced statistical, computational, and remote-sensing methodologies to interrogate both their physical characteristics and their human-environment interfaces.

1. Geometric and Morphometric Characterization

Quantitative morphological analysis of lake geometry is fundamental for hydrological and climatic modeling. The area $A$ and perimeter $P$ of a lake exhibit scale-dependent relationships; for lakes with fractal boundaries, perimeter scales with area according to $P \sim A^{1/D}$, where $D$ denotes the fractal dimension. Studies of large tundra lakes in the Russian High Arctic identified bifurcation in $D$ for lakes exceeding 100 km², with observed values ranging from approximately 1.43 to 1.95, indicating increasing shoreline complexity and potential self-similarity with scale (Sudakov et al., 2017).

Lake-area distributions empirically follow power-law tails, $prob(A) \sim A^{-\tau}$, with $\tau \sim 2$ for large lakes ($A > 70$–100 km²), signifying a preponderance of small lakes and progressive rarity of large lakes. These metrics are vital for upscaling methane flux estimates and modeling permafrost-lake system evolution (Sudakov et al., 2017).

2. Remote Sensing and Extraction Methodologies

High-resolution mapping and temporal monitoring of lake boundaries and areas are achieved via machine learning–based semantic segmentation of satellite imagery. Pure-transformer models, such as Swin-Unet, have demonstrated superior performance over standard convolutional neural networks in pixel-wise classification of “lake/non-lake” regions. The hierarchical, window-based attention mechanism (W-MSA, SW-MSA) in Swin-Unet enables efficient encoding of local and cross-window contextual cues. Training employs Tversky loss for class imbalance mitigation, and evaluation is based on mean Intersection-over-Union (MIoU), with top models attaining MIoU ≈ 0.905 and pixel accuracy of 0.98 on test sets (Han et al., 2024).

These models underpin the creation of high-temporal-resolution global lake datasets, such as GLAKES-Additional, which provides biennial lake area delineations from 1990 to 2021 for 152,567 lakes (>0.5 km²) worldwide. Preprocessing steps include aggregation of multi-decadal Landsat data, vector masking for shoreline and river exclusion, contour extraction, assignment of persistent identifiers, and temporal gap-filling (Han et al., 2024).

3. Spatiotemporal Dynamics and Climate Attribution

Longitudinal satellite-derived datasets enable the analysis of lake-area dynamics in response to climatic drivers. Stacked LSTM (Long Short-Term Memory) neural networks ingest sequences consisting of contemporaneous area, precipitation, vapor pressure, and temperature, achieving predictive RMSE as low as 0.317 km². Attribution analysis demonstrates that biennial lake area variance is primarily explained by precipitation anomalies (~60%), with temperature (~25%) and vapor pressure contributing the remainder. Regional heterogeneity is pronounced: while most lakes globally display slow area increases under moderate warming/precipitation scenarios, mid-latitude wetlands may shrink if evaporative demand surpasses hydrological input (Han et al., 2024).

Temporal comparisons using multi-decadal imagery additionally reveal that individual lakes may experience substantial areal changes (up to ±50% over four decades), with no overall bias toward universal growth or contraction, suggesting interacting climatic, geomorphological, and drainage processes (Sudakov et al., 2017).

4. Hydrodynamics and Mathematical Modeling

The hydrodynamics of lakes are governed by the so-called "lake equations," derived from shallow-water theory for incompressible, inviscid flows in domains with variable bathymetry. The system

$$\begin{aligned} \nabla\cdot \left(b\mathbf{u}\right) &= 0 \ \partial_t \mathbf{u} + (\mathbf{u}\cdot\nabla)\mathbf{u} &= -\nabla h \end{aligned}$$

where $b(x)$ is the spatially-varying depth, $P$0 is the vertically-averaged horizontal velocity, and $P$1 is the surface deviation, encapsulates mass and momentum conservation with impermeable boundary ($P$2 on $P$3). The potential vorticity $P$4 (with scalar vorticity $P$5) satisfies a transport equation $P$6, conserving integrated circulation.

In the singular limit, initial vorticity concentration evolves as a point vortex with trajectory governed by

$P$7

or, equivalently, $P$8, where $P$9 is the canonical rotation operator. This geometric law determines the leading-order drift of localized vortices and, thus, informs both analytical studies of geophysical flows and subgrid-scale parameterizations in large-scale numerical models (Dekeyser et al., 2019).

5. Statistical and Machine-Learning–Based Modeling

Lake surface water temperature (LSWT) prediction is operationalized via probabilistic spatiotemporal deep learning models. Bayesian spatiotemporal graph convolutional neural networks (BSTNN) fuse meteorological covariates, depth-specific simulations, and in-situ measurements at multiple spatial nodes, with Bayesian LSTM layers capturing temporal autocorrelation and Bayesian GCN layers encoding spatial adjacency via diffusion kernels. The full posterior predictive distribution is approximated via variational inference, minimizing the evidence lower bound (ELBO). For Lake Geneva, BSTNN achieves $P \sim A^{1/D}$0, RMSE = 1.82 °C, and 90% coverage with interval width $P \sim A^{1/D}$1, outperforming competing deep Bayesian methods, especially under sparse observation regimes (Stalder et al., 2021).

Transferability of these models is enabled due to modular decoupling of temporal and spatial structure, requiring only adaptation of the adjacency matrix and prior tuning for new lakes. However, sensitivity arises at shoreline nodes, which have limited spatial connectivity and simulation fidelity (Stalder et al., 2021).

6. Lake Operation and Management via Optimization

Reservoir and regulated lake management involves optimal scheduling of releases to balance competing objectives: flood avoidance, maintenance of minimum water levels, and water-supply reliability. Model Predictive Control (MPC) is formulated as a quadratic program over prediction horizons, incorporating physical constraints from mass-balance, state-dependent nonlinear release bounds, and soft penalties for unmet demand and threshold violations. In operational tests at Lake Como, hourly MPC achieves flood-RMSE of 0.5215 m and zero dry-violation hours, within ~5% of deterministic dynamic programming (DDP) offline benchmarks, but with a significantly higher rate of acceptable demand-deficit events. Solver runtimes ($P \sim A^{1/D}$20.3 s per hour) permit real-time deployment (Cestari et al., 2022).

Daily MPC approximations, justified under Gaussian fluctuation assumptions for sub-daily flows, retain comparable flood and demand performance, facilitating use where only daily data are available (Cestari et al., 2022).

Control Strategy	Flood-RMSE (m)	Demand-RMSE (m³/s)	Dry-Violation Hours
Hourly MPC	0.513	40.34	0
Daily MPC	0.527	40.88	0

7. Human-Lake Interaction and Socioecological Analytics

Emerging conceptual frameworks, such as “lakeplaces,” integrate lakes with their immediate human-use catchments. A lakeplace consists of a lake and its surrounding census-administrative blocks capturing the intersection of water and human activity. High-resolution mobile positioning data enables estimation of visitation intensities, temporal demand clusters, and segregation of on-lake (boating, swimming) versus around-lake (park, entertainment) activities (Xiong et al., 4 May 2025). The scaled intensity metric

$P \sim A^{1/D}$3

and on-lake contribution index

$P \sim A^{1/D}$4

quantify lakeplace popularity and functional typology.

Spatial autocorrelation in visitation (Moran’s I = 0.24, $P \sim A^{1/D}$5) underscores clustering mechanisms in urban settings, while demographic association analyses reveal that both local and non-local users of lakeplaces tend to be disproportionately high-income and majority-white compared to the general population, suggesting equity challenges in lake access (Xiong et al., 4 May 2025). The integration of spatiotemporal big data facilitates dynamic monitoring and planning for equitable, sustainable urban water-resource management.