Contributions of Water Surface Area and Evaporation Intensity to Long- Term Reservoir Evaporation Loss Changes Across China

Technical Summary: Contributions of Water Surface Area and Evaporation Intensity to Long-Term Reservoir Evaporation Loss Changes Across China

Problem Statement
Reservoir evaporation represents a critical yet often under-accounted water loss in China's water-resource system. While existing studies have improved estimates of evaporation magnitude, they frequently emphasize either evaporation depth (intensity) or total volume without explicitly partitioning the observed changes between the two driving factors: the exposed water surface area and the evaporation intensity. This distinction is vital because the mechanisms driving loss changes dictate different management responses. If losses are driven primarily by surface expansion, management should focus on reservoir development and surface regulation. If driven by intensity changes, management must address changing atmospheric demand on existing surfaces. This study addresses the gap in national-scale mechanism separation, aiming to quantify how much of the long-term change in reservoir evaporation loss is attributable to changes in water surface area versus changes in evaporation intensity.

Methodology
The study analyzed monthly records from the Res-CN dataset for 3,168 high-quality reservoirs across mainland China from 1984 to 2020. The analytical workflow proceeded as follows:

1. Variable Construction: Annual evaporation loss ($V$), mean water surface area ($A$), and evaporation intensity ($I$) were derived from monthly data. Intensity was calculated as an area-normalized daily depth ($mm \cdot day^{-1}$) using the relationship $V = A \cdot I \cdot D / 1000$.
2. Trend Analysis: Long-term trends for loss, area, and intensity were assessed at the reservoir level using the non-parametric Mann-Kendall test and Sen's slope estimator.
3. Decomposition: Long-term changes in evaporation loss were decomposed into area and intensity effects using logarithmic differences between the first and last valid years. The contribution of each factor was calculated as the percentage of the absolute log-change it represented.
4. Machine Learning: Elastic Net and LightGBM models were employed to identify dominant predictors of spatial variation in long-term mean evaporation loss and intensity. Features included hydroclimatic variables (e.g., reference evapotranspiration, temperature), topography, reservoir morphology, and land cover.

Key Results

• Spatial and Seasonal Patterns: Reservoir evaporation exhibited strong spatial heterogeneity and a summer-dominant seasonal cycle. July was the peak evaporation month for 64.1% of reservoirs, and June for 22.3%.
• Trend Analysis: Annual evaporation loss increased significantly in 55.7% (1,764) of the reservoirs. In contrast, trends in evaporation intensity and water surface area were more heterogeneous, with a majority of reservoirs showing no significant trend in either variable individually.
• Decomposition Findings: The decomposition revealed that evaporation intensity was the dominant driver of long-term loss changes. On average, intensity changes contributed 57.0% to the total evaporation-loss change, while area changes contributed 43.0%. Among reservoirs with increasing losses, the intensity contribution was even higher at 57.8%.
• Predictive Modeling: Machine-learning models successfully reproduced spatial variations in evaporation loss with high skill ($R^2 = 0.954$ for log-transformed loss). Key predictors identified included reference evapotranspiration, reservoir area, topography, temperature, and sunshine duration. The model for evaporation intensity showed lower predictive skill ($R^2 = 0.649$), suggesting that current predictors do not fully capture the complex physical processes governing area-normalized evaporation depth.

Significance and Claims
The authors claim that their primary contribution is the structural separation of evaporation-loss changes into area and intensity components at a national scale. The results indicate that limiting water surface expansion alone is insufficient for managing reservoir evaporation losses. Because intensity changes account for the majority of the long-term loss increase, effective reservoir management must also account for changing atmospheric demand on existing reservoir surfaces.

The study positions evaporation loss not merely as a secondary water-balance term but as a volumetric accounting problem with two separable components. By demonstrating that the "intensity" term (driven by atmospheric demand and local conditions) often outweighs the "area" term, the paper argues for a more nuanced approach to water-resource planning that integrates evaporation-aware management into effective storage and supply-reliability assessments.

Limitations and Uncertainty
The paper modestly acknowledges several sources of uncertainty:

• Data Quality: Area estimates for small reservoirs are susceptible to mixed pixels and shoreline complexity, which can amplify errors in intensity calculations. Consequently, intensity statistics were treated as relative comparisons rather than absolute benchmarks for open-water depth.
• Statistical Assumptions: The Mann-Kendall test was applied without pre-whitening or autocorrelation adjustments, which could affect significance classifications if serial correlation exists.
• Endpoint Sensitivity: The decomposition relied on the first and last valid years, making the results potentially sensitive to anomalous endpoint years.
• Physical Attribution: While machine learning identified predictive signals, the study notes that factors such as reservoir operation, heat storage, fetch, and bathymetry are not fully characterized in the available data, limiting the ability to provide a complete physical attribution for intensity changes.