GLake-DICDOC Dataset

These cover key regions including Asia (Chinese lakes and reservoirs; the Sayan-Altai region of Central Asia), Africa (24 lakes and long-term records for Lake Victoria), South America (high-altitude lakes in Ecuador), and North America/the Arctic (Western Canada; 30-year time series from Quebec-Newfoundland; prairie lakes; Arctic-alpine lakes; and lake clusters in the Sierra Nevada). Additionally, specialized experimental data focusing on biological metabolic and photochemical processes were integrated, resulting in a representative sample system that spans global geography, ecological gradients (from fresh to saline waters), and climatic zones (from tropical to polar).

The spatial distribution of data points shown in Figure 2.1 reveals higher sampling densities in North America, Europe, and East Asia, contrasted with relative sparsity in the interior of South America, Central Africa, and remote high-latitude regions—reflecting the uneven nature of historical monitoring networks.Regarding the data processing workflow, a unified standardization of measurement units was first implemented: based on IUPAC atomic weights, concentrations of DOC, DIC, and total nitrogen (TN) were uniformly converted to mg·L⁻¹, while total phosphorus (TP)—being a trace element—was retained in μg·L⁻¹ to balance ecological accounting conventions with numerical precision.

Subsequently, a multi-level logical data cleaning process was implemented: ① Thresholds based on physicochemical principles (e.g., pH ∈ [3.0, 11.0]) were established to filter out values ​​clearly resulting from instrument malfunctions; ② For skewed variables such as carbon, nitrogen, and phosphorus, an improved interquartile range (IQR) method (IQR × 1.5) was used to identify outliers; however, stratified quality control based on water type (freshwater vs. saline water) was applied to retain genuine high values ​​found in extreme habitats—such as salt lakes in arid regions and high-altitude alpine lakes—thereby preventing a loss of ecological representativeness; ③ For lakes with repeated sampling across multiple sites and time points (e.g., hotspots like Lake Taihu and Lake Erie), data were aggregated seasonally using the median value to represent the lake for that specific period, thereby eliminating biases caused by varying monitoring intensities and achieving spatiotemporal dimensionality reduction.

Following this cleaning and aggregation, the final dataset comprises 1,827 valid lake entities (i.e., 1,827 records), each identified by a unique Hylak_ID. The columns encompass three categories of variables: (1) Morphological covariates (surface area [km²], shoreline length [km], volume [km³], mean residence time [days], and catchment area [km²]); (2) Hydrochemical variables (DOC [mg·L⁻¹], DIC [mg·L⁻¹], pH [dimensionless], electrical conductivity [μS·cm⁻¹], TN [mg·L⁻¹], TP [μg·L⁻¹], dissolved oxygen [mg·L⁻¹], chlorophyll-a [μg·L⁻¹], etc.); and (3) Environmental drivers (mean annual temperature [°C], annual precipitation [mm], slope [%], forest cover [%], cropland cover [%], soil organic carbon [g·kg⁻¹], soil erosion modulus [t·km⁻²·a⁻¹], etc.).

Notably, there are 1,642 paired observations for DIC and DOC, with a missing data rate of 10.2%; this is primarily attributed to early monitoring programs failing to simultaneously measure carbon fractions or to certain regions (such as Africa and the interior of South America) providing only a single carbon metric. To address the issue of Missing Not At Random (MNAR) data, this study avoided simple deletion or mean imputation; instead, it employed the iterative random forest-based multiple imputation algorithm proposed by Stekhoven & Bühlmann (implemented via the `missMDA` package in R 4.2.1).

This algorithm reconstructs missing values ​​by leveraging nonlinear covariation among variables without requiring distributional assumptions: it begins with mean imputation, followed by the iterative fitting of random forest models—ordered by ascending missingness rates—using the column with missing values ​​as the response variable and all other variables as predictors, continuing until the prediction Mean Squared Error (MSE) converges.

This strategy preserves sample integrity for high-latitude and less-developed regions while effectively maintaining the biogeochemical structure linking lake morphology, climate, and hydrochemistry. Cross-validation (5-fold CV) of the imputed data yielded an average relative error of < 8.3%; notably, reconstruction errors for DIC (RMSE = 12.7 mg·L⁻¹) and DOC (RMSE = 4.1 mg·L⁻¹) fell within reasonable limits relative to measurement precision.

All data are stored in standard CSV format (UTF-8 encoding), ensuring compatibility with major analysis platforms such as R, Python (pandas), ArcGIS, and SPSS. Citations for the original literature and datasets are provided in references [34, 50–64], and the complete dataset and metadata are available via the supplementary materials or designated data repositories (e.g., Zenodo).