CO₂ Emissions Associated with Groundwater Storage Depletion in South Korea: Estimation and Vulnerability Assessment Using Satellite Data and Data-Driven Models

Abstract

Groundwater is crucial in mediating the interactions between the carbon and water cycles. Recently, groundwater storage depletion has been identified as a significant source of carbon dioxide (CO₂) emissions. Here, we developed two data-driven models—XGBoost and convolutional neural network–long short-term memory (CNN-LSTM)—based on multi-satellite and reanalysis data to monitor CO₂ emissions resulting from groundwater storage depletion in South Korea. The data-driven models developed in this study provided reasonably accurate predictions compared with in situ groundwater storage anomaly (GWSA) observations, identifying relatively high groundwater storage depletion levels in several regions over the past decade. For each administrative region exhibiting a decreasing groundwater storage trend, the corresponding CO₂ emissions were quantified based on the predicted GWSA and respective bicarbonate concentrations. For 2008–2019, XGBoost and CNN-LSTM estimated CO₂ emissions to be 0.216 and 0.202 MMTCO₂/year, respectively. Furthermore, groundwater storage depletion vulnerability was assessed using the entropy weight method and technique for order of preference by similarity to ideal solution (TOPSIS) to identify hotspots with a heightened potential risk of CO₂ emissions. Western South Korean regions were particularly classified as high or very high regions and susceptible to groundwater storage depletion-associated CO₂ emissions. This study provides a foundation for developing countermeasures to mitigate accelerating groundwater storage depletion and the consequent rise in CO₂ emissions.

1. Introduction

The amount of CO₂ in the atmosphere continues to increase due to the combustion of fossil fuels and deforestation [1]. In an attempt to mitigate climate change on a global scale, concerted efforts are being made to reduce CO₂ emissions worldwide [2]. To understand climate change clearly and enact effective policies for mitigating this phenomenon, a holistic understanding of the water and carbon cycle processes is required [3,4]. Globally, on average, water bodies emit 7700 million metric tons of CO₂ (MMTCO₂) annually, with streams and rivers contributing 6600 MMTCO₂/year and lakes and reservoirs accounting for 1100 MMTCO₂/year [5]. Therefore, monitoring changes in water use due to accelerated climate change and the attendant changes in CO₂ emissions from water bodies is paramount.

Groundwater also contributes to atmospheric carbon dynamics through the interaction between the carbon and water cycles; it releases CO₂ into surface waters as well as the atmosphere. The atmospheric CO₂ contribution is affected by changes in the carbon flux between groundwater and the atmosphere. In recent years, increased water use, climate change-induced events such as droughts, and declining river flows have led to a greater dependence on groundwater, causing changes in carbon flux. Consequently, groundwater storage depletion has been widely observed globally [6,7,8,9,10,11]. Jasechko et al. [6] examined 1693 aquifers worldwide and found that groundwater storage declined in 36% of aquifers from 2000 to 2022. They also confirmed the acceleration of groundwater storage depletion by observing that in 30% of the 1693 aquifers, groundwater decline in the early 21st century exceeded the rate of decline in the late 20th century.

Groundwater depletion, a previously unrecognized factor, has now been reported to contribute to CO₂ fluxes. There is no net change in CO₂ emissions in a steady state of the groundwater aquifer system, where recharge and discharge are balanced [12]. However, if groundwater storage is depleted at a rate exceeding its recharge, CO₂ is released into the atmosphere [13,14]. Consequently, the increasing trend of groundwater depletion due to groundwater extraction, driven by factors such as climate change and drought, is likely to result in higher CO₂ emissions. Wood and Hyndman [12] reported that groundwater depletion was responsible for the release of approximately 1.7 MMTCO₂/year into the atmosphere in the United States during 2000–2008, and it has been ranked among the top 20 sources of carbon emissions as documented by the US Environmental Protection Agency [15]. Mishra et al. [16] investigated CO₂ emissions due to groundwater depletion and pumping in India using groundwater observation wells and the Gravity Recovery and Climate Change Experiment (GRACE) satellite. They estimated that during 2002–2016, the total groundwater depletion was 122–199 billion m³, and the corresponding annual CO₂ emission was approximately 0.72 MMTCO₂/year in India. Based on these findings, they emphasized the importance of understanding the relationship between groundwater and carbon emissions in preventing adverse environmental impacts.

However, analyzing the spatiotemporal relationships between carbon flux and groundwater storage is challenging due to the sparsity of observational data, discontinuities in data provision periods, and the low spatial resolution of available satellite data [11]. Groundwater storage observations, which often rely on point-based measurements, may not be sufficient for capturing groundwater storage depletion associated with spatiotemporal variations in CO₂ emissions. As an alternative, combining satellite data with advanced data-driven models, such as machine and deep learning models, has recently garnered interest. This approach can provide a clearer idea of groundwater storage at different temporal and spatial scales, thereby overcoming the limitations of existing satellite data. For example, data-driven models have been used to predict temporal changes in groundwater storage [17,18,19], groundwater potential [20,21,22], and groundwater drought [23,24,25], as well as for the downscaling of groundwater storage data [26,27,28]. The existing studies predominantly employed data-driven models that relied on single meteorological variables, such as precipitation or temperature, or a restricted set of input parameters. However, groundwater dynamics are influenced by a complex interplay of multiple factors, necessitating a more comprehensive and diverse set of input variables. In this study, we developed a data-driven model to predict groundwater storage anomalies (GWSAs) and depletion by integrating an extensive range of input variables, encompassing satellite-based meteorological, hydrological, and vegetation data. CO₂ emissions resulting from groundwater storage depletion can be estimated more accurately by utilizing a combination of data-driven models and multi-satellite data to predict spatiotemporal variations in groundwater storage.

With projections indicating a significant future decline in groundwater storage, recognizing the vulnerability to such depletion and identifying high-risk areas become imperative, especially considering the potential for associated CO₂ emissions. In this regard, vulnerability assessment has emerged as a crucial tool in water resource management [29]. While the definition of vulnerability varies, it fundamentally refers to the degree of susceptibility or inability to cope with damages and negative impacts from climate change or human activities [30]. Adger [31] further defines vulnerability as the state of being susceptible to harm due to exposure to stresses associated with environmental and social change, coupled with a lack of adaptive capacity. Therefore, this study assessed vulnerability to groundwater storage depletion using the entropy weight and technique for order of preference by similarity to ideal solution (TOPSIS) methods, which considers spatial and temporal factors, including both natural and anthropogenic impacts. In addition, potential hotspots were identified by analyzing regions susceptible to groundwater storage depletion using vulnerability assessment and associated CO₂ emissions based on data-driven models. Vulnerability assessment using this technique enables the identification of areas at risk of increased CO₂ emissions due to groundwater storage depletion.

In this study, we investigated the impacts and vulnerabilities of groundwater storage depletion on CO₂ emissions in South Korea using machine and deep learning models based on multi-satellite data. The specific objectives were as follows: (1) to predict the GWSA based on hydrological factors using multi-satellite and reanalysis data and to validate the spatiotemporal GWSA prediction and distribution by comparing them with in situ groundwater observations; (2) to quantify the contribution of groundwater depletion to CO₂ emissions using the predicted GWSA; and (3) to investigate the extent of vulnerability to groundwater storage depletion that may result in CO₂ emission changes using the entropy weight method and TOPSIS.

2. Study Area and Data

2.1. Study Area

South Korea is located in East Asia (33°N to 39°N and 125°E to 130°E; Figure 1a) and is characterized by a temperate climate with a mean annual temperature of 12–14 °C and four distinct seasons. It is predominantly influenced by East Asian monsoons and typhoons. The mean annual precipitation is 1000–1800 mm, and rainfall occurs mainly during the Changma season (from July to August). South Korea consists of 160 administrative regions in 17 provinces (A01, Seoul; A02, Gyeonggi-do; A03, Incheon; A04, Gangwon-do; A05, Chungcheongnam-do; A06, Chungcheongbuk-do; A07, Daejeon; A08, Sejong; A09, Jeollabuk-do; A10, Jeollanam-do; A11, Gwangju; A12, Gyeongsangbuk-do; A13, Gyeongsangnam-do; A14, Daegu; A15, Ulsan; A16, Busan; A17, Jeju; Figure 1b). The Jeju Island region (A17) was excluded from the study area. Monthly groundwater-level data (2003–2019) from 166 stations (blue dots in Figure 1c) of the National Groundwater Monitoring Network (NGMN) were used. In addition, soil texture classification data (Figure 1c), based on a detailed soil map (1:25,000) of the National Institute of Agricultural Sciences and Korean Rural Development Administration, were used to calculate the specific yield (S_y) and convert groundwater-level data to groundwater storage. The proportions of clay (0.23%), silty clay loam (4.08%), clay loam (2.85%), silt (0.21%), silt loam (21.15%), sandy clay loam (0.59%), loam (27.61%), sandy loam (39.92%), loamy sand (3.06%), and sand (0.29%) in the soil were noted. Sandy loam, loam, and silt loam dominate the soil texture in South Korea (~89%). South Korea has a diverse geological landscape, which contributes to various types of aquifers, including alluvial, fractured rock, and volcanic aquifers. Alluvial aquifers are found in river valleys and coastal regions and are composed of unconsolidated sediments like sand, gravel, and silt, which have high porosity and permeability, making them significant sources of groundwater. In mountainous and hilly regions, fractured rock aquifers are prevalent. Jeju Island is known for its volcanic aquifers formed in basaltic rock.

2.2. Data

In this study, the input datasets of data-driven models consisted of the precipitation (P), mean temperature (T), soil moisture content (SM), normalized difference vegetation index (NDVI), modified normalized difference water index (MNDWI), and terrestrial water storage anomaly (TWSA) data, which were used to predict GWSA (

Table 1. Information on the satellite and reanalysis data.

Variables	Data	Properties	Reference
P	TMPA3B43V7	0.25° resolution, January 2003–December 2019	https://disc.gsfc.nasa.gov/datasets/TRMM_3B43_7/summary, accessed on 20 November 2023
T and SM	GLDAS NOAH025_M	0.25° resolution, January 2003–December 2019	https://ldas.gsfc.nasa.gov/gldas, accessed on 20 November 2023
NDVI and MNDWI	Landsat 5 and 8	30m resolution, January 2003–February 2013 (Landsat 5) March 2013–December 2019 (Landsat 8)	https://earthexplorer.usgs.gov, accessed on 20 November 2023
TWSA	CSR RL06M	0.25° resolution, January 2003–April 2017 (GRACE) June 2018–December 2019 (GRACE-FO)	https://grace.jpl.nasa.gov, accessed on 20 November 2023

).

The P data were obtained from the TMPA3B43V7 dataset produced by the Tropical Rainfall Measuring Mission (TRMM) satellite, which has been widely applied and validated in this study area [32,33]. Monthly P datasets from 2003 to 2019 with a spatial resolution of 0.25° were used. The T and SM values used in this study were provided by the global land data assimilation system (GLDAS, [34]). Monthly T and SM datasets (Noah025_M) from 2003 to 2019 with a spatial resolution of 0.25° were used.

The NDVI and MNDWI values were sourced from datasets provided by Landsat 5 (TM) and 8 (OLI) satellites. A fast line-of-sight atmospheric analysis using the spectral hypercube algorithm was performed for radiometric and atmospheric correction [35]. A total of 2652 processed images were mosaicked via nearest-neighbor sampling. The 30 m NDVI and MNDWI values over the entire period (2003–2019) were resampled to 0.25° grids.

TWSA data were determined using the equivalent water thickness converted from mass change observations provided by the GRACE satellite [36]. The GRACE data were processed at the Center for Space Research (CSR), Jet Propulsion Laboratory, and German Research Centre for Geosciences. In this study, we used the monthly TWSA data from the CSR mascon solution (release 6) at a resolution of 0.25° [37,38]. The CSR mascon data were improved upon through the measurement noise of traditional spherical harmonic coefficients through coefficient (C20, C30) replacement and filtering, thereby reducing leakage errors between coastal and land masks. Wang et al. [39] assessed the CSR mascon data, finding it to exhibit the lowest uncertainty among the available GRACE mascon solutions. GRACE data have been extensively applied and compared with in situ measurements in various studies in South Korea, demonstrating its reliability and validity [18,24,33,40]. Cubic spline interpolation was employed to fill in the monthly data gaps caused by battery issues and the 11-month hiatus between the GRACE and GRACE-FO products from July 2017 to May 2018. Cubic splines have been widely used in numerous studies due to their relative simplicity and effectiveness in gap-filling between GRACE and GRACE-FO [18,41,42,43,44].

The in situ GWSA represents the deviation for a specific month relative to the baseline average from January 2004 to December 2009. The monthly changes in groundwater-level data from 166 stations (blue dots in Figure 1c) were converted to GWSA values in terms of equivalent height by multiplying S_y by groundwater-level changes. The median S_y value for each 0.25° grid was calculated using soil texture information (Figure 1c), as reported by [45,46].

3. Methods

3.1. Prediction of GWSA

3.1.1. Machine and Deep Learning Models

Groundwater in situ data are limited in its spatial representation as it consists of point data. Although South Korea has 166 groundwater observation stations, the data are insufficient to calculate GWSA by administrative region. Consequently, satellite data present challenges for nowcasting and forecasting due to latency in data provision by institutions such as NASA. This study aims to develop data-driven models that consider the lag time of satellite and reanalysis data to predict accurate spatiotemporal GWSA.

In this study, machine (XGBoost) and deep learning (convolutional neural network–long short-term memory; CNN-LSTM) models were developed to predict GWSA (Step 1 in Figure 2).

XGBoost is a machine learning model that makes up for the shortcomings of the gradient boost model using residuals, which reduce errors by weighting data. It adds an overfitting prevention technique to gradient tree boosting, which is a supervised learning algorithm that sequentially combines gradient-boosting models. XGBoost is a highly accurate algorithm for regression problems (i.e., time series prediction) because it uses parallel computation to learn faster and prune unnecessary subtrees to reduce overfitting and achieve greater generalizability [47]. The CNN-LSTM model is powerful for time series prediction due to its ability to extract and learn both spatial and temporal features, leading to more accurate and reliable predictions [48].

The input variable sets for the data-driven models included satellite and reanalysis data from m months prior (t-m), specifically P(t-m), T(t-m), SM(t-m), NDVI(t-m), MNDWI(t-m), and TWSA(t-m) (Table 1 and Figure 3). The GWSA(t) is the output variable (prediction values) and was validated with in situ data (actual values) from 166 groundwater observation wells. The total period (2003–2019) of the dataset was divided into training (80%) and test (20%) periods.

The models were implemented using the XGBoost and Keras libraries in Python (version 3.6). The input variables were standardized using the min–max scaling method, which scales the data to have values between 0 and 1 to prevent overfitting in this study. The XGBoost structure consists of weak decision trees and each tree focuses on reducing the residuals (errors between the actual and predicted values) from the previous iteration to create strong learners (Figure 3a). The loss function was selected as the root mean square error (RMSE) to optimize errors in the training process in this study.

The CNN-LSTM structure consists of CNN, LSTM, and dense layers (Figure 3b). The convolutional layer extracts spatial features through an activation function (the ReLU function was used) after performing a convolutional product operation on the input data. Subsequently, the max pooling layer reduces the size of the detected feature data. Thereafter, the LSTM layer performs a time series prediction by extracting temporal features [18]. The RMSE was used as a loss function for updating the network as with the Adam optimizer in this study.

3.1.2. Hyperparameter Tuning

Hyperparameters are values that are often set by the user of an algorithm based on experience. The optimal value is not predefined and must be determined through multiple iterations. Determining optimal hyperparameter values is essential for the model to achieve high performance. Three main methods are used for tuning hyperparameters: grid search, random search, and Bayesian optimization. In this study, we used Bayesian optimization, which can handle high hyperparameter counts and large search spaces. Such methods are more efficient than grid or random search methods because they balance the exploitation and utilization of the search space [49].

Table 2. Hyperparameter tuning ranges for XGBoost and CNN-LSTM models.

Model	Hyperparameter	Setting Value Range
XGBoost	Number of estimators	{1, …, 1000}
	Maximum depth	{3, 5, …, 15}
	Minimum child weight	{1, 2, …, 9}
	Subsample	{0.5, 0.6, …, 0.9}
	Colsample_bytree	{0.5, 0.6, …, 0.9}
	Gamma	{0, 0.1, …, 0.7}
	Lambda	{0, …, 100}
	Alpha	{0, …, 100}
	Learning rate	{0.001, …, 0.01}
	Time lags (month)	{1, 2, …, 12}
CNN-LSTM	Number of filters	{16, …, 128}
	Number of hidden layers	{1, 2, 3}
	Hidden unit	{0, …, 100}
	Dropout	{0.1, 0.2, …, 0.5}
	Learning rate	{0.001, …, 0.01}
	Epochs	{1, 2, …, 100}
	Iteration	{1, 2, …, 100}
	Time lags (month)	{1, 2, …, 12}

lists the hyperparameter settings used to tune the XGBoost and CNN-LSTM models and the ranges for each hyperparameter. The parameters to be optimized for XGBoost comprise tree-specific hyperparameters (number of estimators, maximum depth, minimum child depth, subsample, and colsample_bytree), which control the overall behavior and learning process of the model, and learning task-specific hyperparameters (gamma, lambda, alpha, and learning rate). In the case of CNN-LSTM, hyperparameters related to the network structure (number of filters and hidden layers, hidden units, and dropout) and those used for learning (learning rate, epochs, and iteration) were optimized. The loss function was mean square error. For both models, the time lags of 1–12 months were set for predicting the GWSA time series.

3.2. Quantification of CO₂ Emission Due to Groundwater Storage Depletion

3.2.1. Trend Analysis of Model-Predicted GWSA

In this study, the Mann–Kendall test and Sen’s slope were used to identify regions suffering groundwater storage depletion (Step 2 in Figure 2). The Mann–Kendall test is a nonparametric statistical technique that determines a trend within a certain confidence level [50], and Sen’s slope represents the slope of a linear trend [51]. As these methods are largely unaffected by missing data or outliers, they are useful for analyzing time series trends [52,53]. For each administrative region, areas with groundwater storage depletion were identified as those for which the Mann–Kendall test and Sen’s slope (p < 0.05) highlighted a decreasing trend in the GWSA predicted by the XGBoost and CNN-LSTM models.

3.2.2. CO₂ Emission from Groundwater Storage Depletion

Wood and Hyndman [12] suggested that groundwater contributes to CO₂ emissions through the following processes: a groundwater recharge from precipitation, reaction of recharged groundwater in aquifers, and re-exposure of groundwater to the atmosphere. Thus, CO₂ emissions can be estimated by groundwater storage depletion according to groundwater use or flow to rivers and oceans and the equivalent CO₂ concentration in groundwater. During the process of groundwater recharge from precipitation, water and carbon dioxide react when water enters the unsaturated zone. As water passes this zone, ${\mathrm{H}}_2\mathrm{O}$ reacts with ${\mathrm{CO}}_2$ according to Equation (1), producing ${\mathrm{H}}^{+}$ and bicarbonate (${\mathrm{HCO}}_3^{-}$) ions. Most aquifers contain gravel, sand, clay, and calcite (${\mathrm{CaCO}}_3$), which react with the produced ${\mathrm{H}}^{+}$ ions to generate ${\mathrm{HCO}}_3^{-}$ and ${\mathrm{Ca}}^{2+}$ ions (Equation (2)). When groundwater is extracted by pumping or discharged into rivers or the ocean, it becomes exposed to the atmosphere, causing CO₂ to be released and ${\mathrm{CaCO}}_3$ to precipitate, as described in Equation (3). Based on a conservative assumption, half of the ${\mathrm{HCO}}_3^{-}$ concentration in groundwater is converted to CO₂ through the hydrological cycle [12,16]. Therefore, the ${\mathrm{HCO}}_3^{-}$ concentration in groundwater plays an important role in CO₂ emissions due to groundwater storage depletion. Equation (4) was used in this study to convert the ${\mathrm{HCO}}_3^{-}$ (mg/L) concentration in groundwater to the equivalent CO₂ concentration (${\mathrm{CO}}_2\mathrm{e}$; mg/L). The ${\mathrm{HCO}}_3^{-}$ concentration was obtained from the NGMN and converted to the equivalent CO₂ concentration for each administrative region. Subsequently, atmospheric ${\mathrm{CO}}_2\mathrm{Emission}$ MMTCO₂/year) due to groundwater storage depletion ($\mathrm{GWS}\mathrm{depletion}$) was estimated by Equation (5) based on groundwater storage depletion (km³/year) and equivalent CO₂ concentration (kg/m³) [12,16].

$${\mathrm{CO}}_2+{\mathrm{H}}_2\mathrm{O}\rightleftharpoons {\mathrm{H}}_2{\mathrm{CO}}_3\rightleftharpoons {\mathrm{H}}^{+}+{\mathrm{HCO}}_3^{-} \left(\mathrm{subsurface}\right)$$

(1)

$${\mathrm{H}}^{+}+{\mathrm{HCO}}_3^{-}+{\mathrm{CaCO}}_3={\mathrm{Ca}}^{2+}+2{\mathrm{HCO}}_3^{-} \left(\mathrm{subsurface}\right)$$

(2)

$${\mathrm{Ca}}^{2+}+2{\mathrm{HCO}}_3^{-}={\mathrm{H}}^{+}+{\mathrm{HCO}}_3^{-}+{\mathrm{CaCO}}_3={\mathrm{CO}}_2\uparrow +{\mathrm{H}}_2\mathrm{O}+{\mathrm{CaCO}}_3\downarrow \left(\mathrm{surface}\right)$$

(3)

$${\mathrm{CO}}_2\mathrm{e}=\left({\mathrm{HCO}}_3^{-}\times 44\right)/61$$

(4)

$${\mathrm{CO}}_2\mathrm{Emission}=\mathrm{GWS}\mathrm{depletion}\times {\mathrm{CO}}_2\mathrm{e}$$

(5)

3.3. Determining Groundwater Storage Depletion and CO₂ Emission Vulnerability Using Entropy Weight and TOPSIS Method

Quantitative vulnerability assessment methods are often based on multi-criteria decision-making (MCDM), which involves selecting the most important or best alternative among several potential criteria. In particular, TOPSIS [54], an MCDM technique integrated with a geographic information system framework, has been widely utilized in vulnerability assessment as a rational and relatively simple computational method for deriving rankings by calculating the distances between positive and negative ideal solutions. TOPSIS is used for assessing vulnerability to various environmental phenomena, such as droughts [55,56], floods [57,58], and heat waves [59,60]. Determining the weights for the indicators is the most important factor in TOPSIS. Research is being conducted on the entropy–TOPSIS method, which combines TOPSIS with the entropy weight method [61,62] and can calculate weights objectively.

3.3.1. Selection of Indicators

The proposed methodology for quantifying CO₂ emissions resulting from groundwater storage depletion is based on the premise that groundwater storage exhibits a decreasing trend. Thus, an examination of vulnerability to groundwater storage depletion can reveal areas with a potentially high risk of CO₂ emissions. In this study, we employed the entropy–TOPSIS method to assess vulnerability to groundwater storage depletion. This method provides a comprehensive assessment by considering multiple factors simultaneously. Additionally, it enables an easy comparison of vulnerability across regions of interest, facilitating the identification of areas at a heightened risk.

We selected three indicators that have a positive effect on vulnerability to groundwater storage depletion: the groundwater use per unit area, number of annual non-rainy days, and cultivation land per unit area. We also selected three indicators that have a negative effect: annual GWSA (predicted by XGBoost and CNN-LSTM), precipitation, and the water supply ratio (Step 3 in Figure 2;

Table 3. Vulnerability assessment indicators for groundwater storage depletion.

Effect	Indicators	Source	Period
Positive	Groundwater use per unit area (m3/year/km2)	Korea Groundwater Annual Report	2004–2019
	Annual non-rainy days (day)	Korea Annual Hydrological Report
	Cultivation land per unit area (%)	Korean Statistical Information Service
Negative	Annual GWSA (mm/year)	XGBoost and CNN-LSTM models
	Annual precipitation (mm/year)	Korea Annual Hydrological Report
	Water supply ratio (%)	Waterworks Statistics

). An increase in groundwater use per unit area has a direct impact on the depletion in groundwater storage. Consequently, as groundwater usage intensifies, the vulnerability of the area to groundwater storage depletion correspondingly increases.

Annual non-rainy days were selected based on the assumption that an increase in the number of non-rainy days reduces the extent of recharge, thereby reducing groundwater storage. Cultivation land per unit area was selected considering that an increase in cultivation land is expected to concurrently increase the amount of groundwater used for irrigation.

A negative GWSA indicates lower groundwater storage compared with the average (from 2004 to 2009), which is considered to increase vulnerability. A reduction in annual precipitation can lead to drought conditions, resulting in increased groundwater use due to reduced surface water availability; moreover, annual precipitation directly affects groundwater recharge rates. Hence, a lower annual precipitation is considered to increase the risk of groundwater storage depletion. Finally, if the water supply ratio in a region decreases, more people in that region will use groundwater, which is considered to increase vulnerability to groundwater storage depletion.

3.3.2. Determining Weights Using Entropy Weight Method

Calculating appropriate weights is necessary for determining the importance of indicators for vulnerability assessment. Representative methods include the equal weight, Delphi, and analytical hierarchy process methods; however, these methods rely on the subjective judgment of the researcher. Hence, we used the entropy weight method in this study, which is a simple and objective process for calculating the weights of the selected indicators. This method determines the weight according to the variance of each indicator; the more agglomerated the indicator values, the larger the weight ([61,62]; Figure 4).

To calculate the entropy weight, the indicators must be standardized because they might have different units. In this study, we used the min–max method to scale the indicators to values between 0 and 1. Equation (6) represents the conversion processes for the indicators with positive or negative effects on vulnerability:

$$r=\left\{\begin{array}{c}{\displaystyle \frac{x-\min \left(x\right)}{\max \left(x\right)-\min \left(x\right)}}, when x is a positive indicator\\ {\displaystyle \frac{\max \left(x\right)-x}{\max \left(x\right)-\min \left(x\right)}}, when x is a negative indicator\end{array}\right.$$

(6)

where $r$ and $x$ denote the scaled and original values of indicators, respectively.

A representative matrix $R$ can be constructed based on the standardized indicator values, as expressed in Equation (7), where $n$ is the number of indicators and $m$ is the number of regions to be analyzed.

$$R=\left[\begin{array}{ccc}{r}_{11}& \cdots & r_{1n}\\ ⋮& r_{ij}& ⋮\\ r_{m1}& \cdots & r_{mn}\end{array}\right]$$

(7)

The indicator values in the constructed matrix are used to establish a standardized index ($p_{ij}$).

$$p_{ij}={\displaystyle \frac{r_{ij}}{{\sum }_{i=1}^m{r}_{ij}}}$$

(8)

Subsequently, the entropy ($E_j$) is calculated for each indicator, as shown below:

$$E_j=-k{\displaystyle \sum }_{i=1}^m{p}_{ij}\ln p_{ij}$$

(9a)

$$k={\displaystyle \frac{1}{\ln \left(m\right)}}$$

(9b)

where $k$ denotes a constant representing the number of regions.

A scale $d_j$ representing the diversity of the indicators is calculated using Equation (10), and the weight $w_j$ can be obtained using Equation (11).

$$d_j=1-E_j$$

(10)

$$w_j={\displaystyle \frac{d_j}{{\sum }_{j=1}^n{d}_j}}$$

(11)

3.3.3. TOPSIS

TOPSIS considers multiple criteria simultaneously and selects the best alternative for a given purpose, based on the distance between the positive and negative ideal solutions [63].

The TOPSIS decision-making process is as follows. First, given a representative matrix $R$ (Equation (6)), each scaled indicator ($r_{ij}$) is assigned a weight ($w_j$) determined using the entropy weight method (Equation (12)).

$$v_{ij}=w_j{r}_{ij} \left(i=1,2,\dots ,m, j=1,2,\dots ,n\right)$$

(12)

Second, the positive ($A^{+}$) and negative ideal solutions ($A^{-}$) for each indicator (Equations (13a,b)) for the weighted scaled values ($v_{ij}$) can be calculated from the maximum and minimum values of each indicator:

$$A^{+}=\left\{v_1^{+},v_2^{+},\cdots ,v_n^{+}\right\}, v_j^{+}=\left\{\begin{array}{c}\max \left(v_{ij}|i=1, 2, \dots , m)|j\in J^{+}\right)\\ \min \left(v_{ij}|i=1, 2,\dots ,m)|j\in J^{-}\right)\end{array}\right.$$

(13a)

$$A^{-}=\left\{v_1^{-},v_2^{-},\cdots ,v_n^{-}\right\}, v_j^{-}=\left\{\begin{array}{c}\min \left(v_{ij}|i=1, 2,\dots ,m)|j\in J^{+}\right)\\ \max \left(v_{ij}|i=1, 2,\dots , m)|j\in J^{-}\right)\end{array}\right.$$

(13b)

where $J^{+}$ and $J^{-}$ are associated with an indicator, with $j$ having a positive and negative effect, respectively.

Third, the distance of the weighted scaled indicator value from the corresponding positive and negative ideal solution ($S_i^{+}$ and $S_i^{-}$) is calculated (Equations (14a,b)). After the aforementioned steps were completed in this study, the relative closeness coefficient ($C_i^{+}$) was calculated for each administrative region (Equation (15)). The vulnerability of each administrative region to groundwater depletion was ranked based on the order of magnitude of the relative closeness coefficient.

$$S_i^{+}=\sqrt{{\displaystyle \sum }_{j=1}^n{\left(v_j^{+}-v_{ij}\right)}^2}$$

(14a)

$$S_i^{-}=\sqrt{{\displaystyle \sum }_{j=1}^n{\left(v_j^{-}-v_{ij}\right)}^2}$$

(14b)

$$C_i^{+}={\displaystyle \frac{S_i^{-}}{S_i^{+}+S_i^{-}}}$$

(15)

4. Results

4.1. Model Performance in GWSA Prediction

Table 4. Optimal hyperparameter values for XGBoost and CNN-LSTM models.

Model	Hyperparameter	Optimal Value
XGBoost	Number of estimators	88
	Maximum depth	5
	Minimum child weight	5
	Subsample	0.7
	Colsample_bytree	0.8
	Gamma	0.0
	Lambda	1.0
	Alpha	0.0
	Learning rate	0.1
	Time lags (month)	8
CNN-LSTM	Number of filters	72
	Number of hidden layers	2
	Hidden unit	70 (layer 1), 60 (layer 2)
	Dropout	0.01
	Learning rate	0.01
	Epochs	60
	Iteration	50
	Time lags (month)	7

summarizes the optimal hyperparameter values to minimize error through Bayesian optimization. Figure 5 shows the GWSA time series predicted by the XGBoost and CNN-LSTM models using optimal hyperparameters overlaid on the in situ GWSA observed in South Korea. Due to the 1–12-month lag time, the GWSA predictions span from January 2004 to December 2019. As observed in the figure, the predictions for both models match all the in situ observations, indicating that the models accurately captured the seasonality of the in situ data.

Figure 6 presents scatterplots depicting the correlation between the model-based predictions and the in situ GWSA for the test period (August 2016–December 2019). The correlation coefficients (r) were 0.84 and 0.76 for XGBoost vs. in situ and CNN-LSTM vs. in situ, respectively. The RMSE values with respect to the in situ observations were 22.45 and 26.88 mm/month for XGBoost and CNN-LSTM, respectively. Thus, the XGBoost model outperformed the CNN-LSTM model (in terms of RMSE and r) for the test period.

Figure 7 shows the spatial distribution maps of the RMSE and r values between the model results and in situ GWSA distributions using inverse distance weighting during the test period. The red grids indicate large errors and non-significant correlations, whereas the blue grids represent significant correlations and small errors.

4.2. Estimation of CO₂ Emission Due to Groundwater Storage Depletion

Trend analyses were conducted to identify regions with significant groundwater storage depletion and estimate the resulting CO₂ emissions. Figure 8a,b show the spatial distributions of the trends determined using the Mann–Kendall test and Sen’s slope. As seen in Figure 8a, the XGBoost-predicted GWSA exhibited a significant decreasing trend (p < 0.05) in the northern (A01, A02, A03, and A04 provinces in Figure 1b) and western (A05, A09, and A10 provinces in Figure 1b) regions, accounting for 32.0% of the total area; the GWSA values were also correspondingly low in these regions during 2008–2019, as observed in Figure 8c. Additionally, the CNN-LSTM-predicted GWSA exhibited a significant decreasing trend (p < 0.05) in 16.2% of the total area, similar to that of the XGBoost-predicted GWSA, except for a few A04 provinces. Separately, significant upward trends (p < 0.05) were observed for both the XGBoost- and CNN-LSTM-predicted GWSA in the southern and central regions (Figure 8a,b), which similarly concurred with the corresponding increase in GWSA in these regions during 2008–2019 (Figure 8c,d). The overall spatial distributions of the GWSA predicted by XGBoost and CNN-LSTM were similar. However, the magnitudes of the GWSA predictions from CNN-LSTM were lower than those from XGBoost.

The bicarbonate concentration was obtained from major anion analysis data of NGMN (Figure 8e; https://gims.go.kr, accessed on 20 November 2023). The median bicarbonate concentration in South Korea (~120 mg/L) is lower than that observed in the Indo-Gangetic Plain (~300 mg/L, [16]) and the United States (~190 mg/L, [12]). Figure 8e shows that notably high concentrations of ${\mathrm{HCO}}_3^{-}$ (>250 mg/L) were recorded in the southeastern (A12 and A14 provinces) and northwestern regions (A01 and A02 provinces).

Atmospheric CO₂ emissions resulting from groundwater storage depletion were estimated using Equation (5). For these calculations, only the model-predicted groundwater storage depletions and the ${\mathrm{HCO}}_3^{-}$-derived equivalent CO₂ concentrations for 2008–2019 were used, because bicarbonate concentration data from the NGMN in Korea have been available only since 2008. A comparison of the CO₂ emission estimates between XGBoost and CNN-LSTM revealed consistent emission patterns, as shown in Figure 9a,b. Both models estimated GWSA values below approximately –200 mm/year in the northwestern A01, A02, and A03 provinces and the eastern A05 province (see Figure 8c,d). Furthermore, a significant decrease was noted in groundwater storage in these provinces, contributing to CO₂ emissions. Notably, Seoul, the capital of Korea, releases more than 0.015 MMT CO₂/year, the highest level identified by both models.

The CO₂ emissions estimated by XGBoost and CNN-LSTM were approximately 0.216 and 0.202 MMTCO₂/year during 2008–2019, respectively. The yearly total CO₂ emissions (2008–2019) are compared between XGBoost and CNN-LSTM in Figure 9c. Emissions for 2011 were excluded because NGMN did not provide the corresponding bicarbonate concentration data. XGBoost and CNN-LSTM estimated the total cumulative CO₂ emissions for 2008–2019 to be 2.596 and 2.426 MMTCO₂, respectively. From 2012 to 2015, a gradual increase in CO₂ emissions was observed, which can be attributed to a corresponding decrease in precipitation during the same period (Figure 9c,d).

4.3. Evaluation of Groundwater Storage Depletion and CO₂ Emission Vulnerability

Figure 10 shows the weights of the six indicators for vulnerability estimation using the entropy weight method. The weights were calculated separately for GWSA predictions using XGBoost and CNN-LSTM and were similar in both cases. The water supply ratio had the highest weight (0.351–0.358), followed by groundwater use per unit area (0.246–0.251), cultivation land per unit area (0.204–0.209), GWSA (0.094–0.113), non-rainy days (0.057–0.058), and finally precipitation (0.029–0.030). The use of the entropy weight method resulted in the positive and negative indicators being weighted at 50% each, indicating that indicator selection was appropriate. The water supply ratio had the highest weight due to its relatively low variance, indicating considerable agglomeration. In contrast, non-rainy days and precipitation had lower weights because their values vary widely across administrative regions in Korea, resulting in low agglomeration. With regard to GWSA, CNN-LSTM had a higher weight than XGBoost did, which is considered to be due to the difference in magnitudes between the XGBoost-predicted and CNN-LSTM-predicted GWSA (Figure 8c,d).

Vulnerability to groundwater storage depletion was evaluated in each administrative region of all provinces using TOPSIS based on the entropy weight method (entropy–TOPSIS). Vulnerability was classified into five levels (very low, low, moderate, high, and very high) based on the closeness coefficient estimates obtained from TOPSIS using the geometric interval function of the ArcGIS tool. The vulnerability maps prepared using the XGBoost and CNN-LSTM predictions are shown in Figure 11.

Both vulnerability maps highlight a relatively large area in the western region as having very high vulnerability. By comparison, the vulnerability of a large proportion of the eastern regions was observed to be low or very low. The higher vulnerability to groundwater storage depletion in the western regions is visually consistent with the spatial distribution of GWSA (Figure 8c,d) and its trends (Figure 8a,b), which exhibited a decreasing pattern in these regions.

Figure 12 presents maps depicting the co-occurrence of CO₂ emissions resulting from groundwater storage depletion (Figure 9a,b) and vulnerability (Figure 11) to groundwater storage depletion; the legend shows the counts of the corresponding regions. These maps reveal that over 80% of the administrative regions that experienced CO₂ emissions due to groundwater storage depletion also exhibited moderate, high, or very high vulnerability to groundwater storage depletion. Based on the XGBoost and CNN-LSTM predictions, we respectively identified 52 and 40 administrative regions as hotspots, representing potential areas that are highly vulnerable to groundwater storage depletion and consequent CO₂ emissions (legends in Figure 12).

5. Discussion

5.1. Groundwater Storage Depletion Using Data-Driven Models Based on Multi-Satellite and Reanalysis Data

Recent research has shown that groundwater storage depletion impacts atmospheric CO₂ emissions, underscoring the importance of identifying spatiotemporal groundwater storage depletion and monitoring the associated CO₂ emissions. Estimating CO₂ emissions resulting from groundwater storage depletion necessitates a precise assessment of changes in groundwater storage, along with the trends and magnitudes of depletion. The GRACE satellite, in conjunction with GLDAS and other satellite-based hydrologic components, has enabled the monitoring of global groundwater changes [64,65]. Various studies have conducted comprehensive assessments of groundwater storage depletion using GRACE data. For instance, Ramjeawon et al. [8] reported a net groundwater storage depletion of 9.25 × 10⁸ m³/year (141.87 mm/year) for the Usutu-Mhlatuze region (6520 km²) in South Africa from 2002 to 2021 based on GRACE data. Similarly, Mohammed et al. [9] documented groundwater storage depletion in the Arabian Peninsula, estimating a depletion of 87.6–96.61 mm/year from 2002 to 2020. Previous methods primarily utilized the water balance equation to estimate groundwater storage depletion [11,64,65]. In this study, we developed a data-driven model to predict GWSA and groundwater storage depletion by innovatively integrating GRACE data, reanalysis data (GLDAS), and satellite-derived (TRMM and Landsat) hydrological data.

We predicted spatiotemporal GWSA in South Korea and groundwater storage depletion through the trend test. The average groundwater storage depletion for the period between 2008 and 2019 of the XGBoost and CNN-LSTM models is 148.78 mm/year and 254.11 mm/year, respectively. Overall, the statistical results indicate that the model GWSA predictions and in situ observations were in agreement. Both XGBoost and CNN-LSTM provided relatively accurate predictions, not only temporally but also spatially. Consistent with our findings, other researchers have demonstrated the effectiveness of machine and deep learning models in accurately predicting groundwater storage [17,18,19]. The XGBoost model outperformed the CNN-LSTM model overall, particularly in the southeast region, where the CNN-LSTM model had a higher RMSE. The accuracy of the models in this region was relatively low because of the complexity of the coastal area, where groundwater levels fluctuate more than in the other regions. To address the relatively large errors in this region, it will be necessary to enhance the model by incorporating various input data or further tuning hyperparameters.

5.2. Relationships between Groundwater Storage Depletion-Associated CO₂ Emissions and Vulnerability

Using the spatiotemporal GWSA predicted by data-driven models, CO₂ emissions due to groundwater storage depletion in South Korea from 2008 to 2019 were estimated. Although the emissions (0.202–0.216 MMTCO₂/year) are lower than those of the United States (~1.7 MMTCO₂/year; [12]) and India (~0.72 MMTCO₂/year; [16]), they are considered relatively high given South Korea’s significantly smaller area compared to these countries. Notably, the highest annual CO₂ emissions were recorded in 2015, with estimated values of 0.537 and 0.450 MMTCO₂ for XGBoost and CNN-LSTM, respectively, concurring with precipitation being lowest that year (Figure 9d). This follows the assumption that a decrease in precipitation reduces groundwater storage, which consequently increases CO₂ emissions. Further research is required to investigate the relationship among droughts resulting from reduced precipitation, water scarcity, and CO₂ emissions.

Additionally, we analyzed regions vulnerable to groundwater storage depletion to identify potential hotspots of the associated CO₂ emissions. As observed in Figure 11, the western regions (A02, A03, A05, A08, and A09 provinces) constitute a large proportion (~ 50%) of the areas with high and very high vulnerability. In contrast, the eastern regions (A04, A15, and A16 provinces) cover more areas with low and very low vulnerability. The vulnerability of province A05 was classified as very high (Figure 11c,d); this province contains the largest proportion of vulnerable areas (Figure 11a,b). These areas predominantly consist of agricultural cultivation zones, which have a low water supply ratio compared with that of other regions, indicating high groundwater usage. Province A05 has been severely affected by recent droughts, which have exacerbated its vulnerability by reducing groundwater recharge and forcing increased groundwater extraction.

In addition, numerous co-occurrences of areas with high CO₂ emissions and vulnerabilities are observed in the western region, particularly in the A02, A03, and A05 provinces. Notably, 70% of the total area in the A05 province showcased co-occurrences of higher CO₂ emissions and high or very high vulnerability (Figure 12). Our findings indicate that the western regions (A02, A03, and A05 provinces) of South Korea are particularly susceptible to groundwater storage depletion and the associated potential risk of CO₂ emissions.

5.3. Significance and Limitations

Point measurements of groundwater are insufficient to identify spatiotemporal changes in GWSA, groundwater storage depletion, and their associated CO₂ emissions. The data-driven model developed in this study can analyze and predict GWSA changes and groundwater storage depletion across different temporal and spatial scales. This approach is expected to play a crucial role in understanding future trends, variations, and patterns of groundwater availability. However, the GWSA prediction results from this model have a spatial resolution of 0.25°, which limits the analysis of groundwater dynamics in detail. Therefore, future research should consider applying methods such as downscaling to improve resolution. It is necessary to employ various data-driven models utilizing a range of precise satellite data to enhance the accuracy of groundwater-related issue assessments. Furthermore, integrating advanced data-driven models and incorporating more comprehensive datasets will be essential to refine predictions and better capture the complexities of groundwater systems.

This study is significant as it represents the first effort in Korea to estimate CO₂ emissions resulting from groundwater storage depletion.

Table 5. Average CO₂ Emissions in Inventories of South Korea (2008–2019).

Sector		CO2 Emission (MMTCO2/Year)	Percentage (%)
Energy		585.408	86.064
Industry		53.690	7.893
Agriculture	Rice paddies	7.070	1.039
	Farmland soil	5.199	0.764
	Livestock manure treatment	4.662	0.685
	Livestock enteric fermentation	4.305	0.633
	Crop residue burning	0.017	0.002
Land use Land use change	Farmland	2.944	0.433
	Wetland	0.312	0.046
Waste	Landfill	7.641	1.123
	Burning	6.499	0.956
	Wastewater treatment	1.716	0.252
	Etc.	0.516	0.076
Groundwater storage depletion		0.202–0.216	0.030–0.032

shows the average CO₂ emissions in Inventories of South Korea (2008–2019; https://kosis.kr/, accessed on 1 July 2024). CO₂ emissions due to groundwater storage may account for a small portion of total CO₂ emissions compared to other sectors, but it is imperative to include groundwater-related activities in considering carbon fluxes. The CO₂ emissions from groundwater storage depletion are also higher than those from crop residue burning and are comparable to emissions from wetlands and waste (etc.). Therefore, the CO₂ emissions resulting from groundwater storage depletion are significant and should not be overlooked. Additionally, recognizing the substantial impact of groundwater depletion on carbon emissions underscores the importance of groundwater management practices to mitigate climate change. These insights can inform groundwater management decisions and future CO₂ emission inventories in South Korea.

Additionally, this study analyzed the relationship between groundwater storage depletion-associated CO₂ emissions and vulnerability. It reveals that western areas experienced high vulnerability to groundwater storage depletion and concurrent CO₂ emissions. This finding demonstrates a correlation between groundwater storage depletion vulnerability and associated CO₂ emissions. Furthermore, the insights gained from this study can assist the decision-making process and resource management strategies aimed at mitigating both groundwater storage depletion and its impact on CO₂ emissions.

In this study, only CO₂ emissions resulting from the depletion of GWSA were estimated. The impact of groundwater storage depletion on CO₂ emissions is driven by a variety of mechanisms. As groundwater levels decline, water must be pumped from greater depths, requiring more energy. This increase in energy consumption often relies on fossil fuels, which leads to higher CO₂ emissions. For instance, Kazakis et al. [66] quantified groundwater depletion in aquifers in the Eastern Thermaikos Gulf, Mouriki, and Marathonas basins in Greece, and explored the application of managed aquifer recharge (MAR) to address groundwater depletion and reduce associated CO₂ emissions. They also proposed approaches economically to reverse groundwater storage depletion by MAR operation scenarios and strategies to mitigate the impacts of groundwater storage depletion. Consequently, various analyses and strategies are needed to mitigate groundwater storage depletion and the resulting CO₂ emissions.

Groundwater storage depletion can also reduce irrigation water availability, impacting crop yields, and stressing natural vegetation, especially in arid and semi-arid regions, leading to additional CO₂ emissions. Furthermore, groundwater storage depletion can reduce the extent of wetlands, which are significant carbon sinks, causing stored carbon to be released as CO₂ or methane (CH₄). Therefore, understanding the future impact of groundwater storage depletion on CO₂ emissions will require a detailed analysis of the relationship between carbon and the hydrological cycle.

6. Conclusions

Groundwater storage depletion has recently been identified as a source of atmospheric CO₂ emission. The fact that it is considered one of the top 20 sources in the USA, along with major sources such as fossil fuel combustion and non-energy use of fuels, implies that CO₂ emissions due to groundwater depletion must be accounted for in future CO₂ budgets.

Key conclusions drawn from this study include

• Data-driven models based on multi-satellite and reanalysis data can predict spatiotemporal GWSA with relatively high accuracy.
• CO₂ emissions from groundwater storage depletion in South Korea were estimated to be 0.216 and 0.202 MMTCO₂/year using XGBoost and CNN-LSTM models, respectively.
• Western regions of South Korea are highly or very highly vulnerable to groundwater storage depletion and prone to CO₂ emissions.
• A correlation relationship was identified between the co-occurrence of CO₂ emissions from groundwater storage depletion and vulnerability to groundwater storage depletion.