Forecasting of groundwater level variation under changing climate in Chhattisgarh state using deep learning technique

Climate change is the long-term change of climatic factors. In the 21st century, humans are the most responsible for climate change (Nations 2022). Its impact on natural resources is quite notable. The urgency of the global climate change problem has grown. So, investigating climatic conditions and variables has become increasingly popular. Groundwater is one of the most precious natural resources. In India, human activities rely more on groundwater than surface water (Singh & Dey 2024). The agricultural sectors in India depend on groundwater, and the ratio of groundwater and surface water used increases yearly. The vast use of groundwater results in fluctuations in groundwater levels (GWLs) in pre-monsoon and post-monsoon. These abrupt fluctuations in groundwater affect the farmers' ability to estimate their area sowing crops. The greatest method for communities to adjust to changing climatic circumstances is the precise climate data that can be used to create efficient strategies for both mitigation and adaptation (Reddy & Saravanan 2023). Therefore, exact and trustworthy GWL predictions offer crucial data on the future availability of groundwater and may be used as the foundation for management choices and plans. Climate models are constantly being updated. Models have been updated regarding higher spatial resolution, new physical processes, and biogeochemical cycles (Bader et al. 2008; Chakraborty et al. 2017). In 2021, the IPCC released its sixth assessment report, which included CMIP6 climate models. In contrast to its predecessors, the most recent CMIP6 GCMs produce a considerably more accurate picture of Earth's physical processes (Calvin et al. 2023). The CMIP6 models, in particular, use Shared Socioeconomic Pathways (SSPs) to forecast future occurrences (Schlund et al. 2020). Climatic factors such as precipitation and temperature vary in future climatic scenarios. SSPs scenarios are created by an international team of climate scientists and economists to guide the climate change research community for an integrated, multidisciplinary examination (Riahi et al. 2017). According to the IPCC study, the SSPs have five storylines: sustainable development, regional competition, inequality, fossil-fueled development, and middle-of-the-road development. The CMIP6 model comes with a certain region-wise bias (Yeboah et al. 2022). This study uses biased, corrected, and freely available data (Mishra et al. 2020). A total of 13 CMIP6 models are available in India. Each model has historical data and four SSP scenarios (SSP126, SSP245, SSP370, and SSP585). This data comes with 0.25⁰ × 0.25⁰ spatial resolution and three parameters, i.e. precipitation, t_max and t_min.

Recent advancements in artificial intelligence (AI) and machine learning (ML) have revolutionized hydrological modeling. Traditional statistical models, such as autoregressive integrated moving averages and seasonal autoregressive integrated moving averages, have been widely used for hydrological time series forecasting (Perera et al. 2024). However, their linear assumptions limit their capability to capture complex, non-linear groundwater dynamics. The artificial neural network (ANN) approach has recently proved its value in forecasting GWLs by effectively handling non-linear relationships and spatial-temporal dependencies (Joshi et al. 2020; Sivakumar et al. 2024). It has also been successfully employed in various hydrological forecasting applications, including rainfall–runoff modeling (Rathnayake et al. 2023). Also, some studies have concentrated on the viability of using ML methods, particularly the branch of ML known as deep learning (DL), in different subfields of hydrology. Some ultramodern engineering uses of ML, such as reinforcement learning (RL) and other innovative methods for applying ML to conventional prediction problems, have not yet been explored. The best approaches to use this geospatial, temporal dataset and enhance integrated hydrological system modeling are provided by technologies that belong under the category of geospatial and geo-spatiotemporal AI, such as DL and parallel computing (Ghobadi & Kang 2023). First off, earlier research concentrated on traditional ML models, but the more recently developing DL attention-based models (such as long- and short-term time-series networks, transformers, informers, and conformers) are still in their infancy, especially in the field of water resources management (Agarap 2019). Beyond standalone ML models, hybrid techniques combining AI and soft computing have gained attention for their improved accuracy and robustness. The adaptive network-based fuzzy inference system (ANFIS), particularly the cascaded-ANFIS model, has been utilized for simulating rainfall–runoff relationships, providing reliable hydrological predictions even in data-scarce environments (Rathnayake et al. 2023). Similarly, studies have demonstrated hybrid models incorporating wavelet transformations and feature selection techniques improve model generalization in hydrological forecasting. In ungauged or data-sparse regions, generative adversarial networks have been introduced for data augmentation and streamflow prediction, addressing challenges related to limited historical records (Perera et al. 2024). Explainable AI methods, such as Shapley Additive Explanations, enhance model interpretability, ensuring that predictions align with hydrological processes (Madhushani et al. 2024).

The Central Groundwater Board (CGWB) is India's leading authority for groundwater management. It monitors GWLs nationwide, taking seasonal readings from a network of observation wells. In Chhattisgarh, the CGWB operates 1,098 observation wells, where GWLs are manually recorded four times a year: during the monsoon (June–September), post-monsoon (October and November), winter (December and January), and pre-monsoon (February–May). However, the manual measurement of GWLs at such a large scale is time-consuming and not cost-effective. This highlights the need for more efficient groundwater monitoring methods, especially in a region like Chhattisgarh. Previous studies have focused on regional hydrological assessments, including flood risk mapping and soil erosion analysis in India using geospatial techniques (Jodhani et al.2023a, b). Groundwater fluctuations are not solely driven by climatic variables but also by anthropogenic activities, land use changes, and lithological characteristics. The increasing reliance on groundwater for irrigation, urbanization, and industrial activities contributes to groundwater depletion and water quality deterioration (Wimalagunarathna et al. 2024). Research utilizing remote sensing and geospatial techniques has demonstrated the effectiveness of integrating land cover dynamics into groundwater models to enhance prediction accuracy. Studies have also explored the role of geospatial AI in water resource management, utilizing Google Earth Engine for large-scale hydrological assessments. For instance, integrating geospatial indices such as the normalized difference vegetation index and irrigation water quality index has been beneficial in identifying groundwater suitability for irrigation (Vyas et al. 2024). However, limited research has been conducted on integrating DL models with climate projections and land use dynamics for groundwater forecasting.

This study investigates both direct and indirect drivers for forecasting seasonal GWLs in Chhattisgarh, India. While direct factors like climate variables have been well explored, indirect drivers, such as anthropogenic activities and lithology, remain scarce in groundwater forecasting (Obahoundje et al. 2017). To address this gap, we designed a DL-based model incorporating land cover dynamics, lithological properties, and climatic variables like temperature and precipitation. Using an ANN approach, we accurately predicted seasonal GWLs based on precipitation, temperature, land use/land cover (LULC) percentages, and lithology conditions. The incorporation of indirect impacts ensures a more refined representation of subsurface water movement. The findings from this study offer an innovative approach that can be utilized by farmers and stakeholders for effective groundwater planning and management, ensuring long-term sustainability in the region.

The present study uses five datasets, as summarized in Table 1, collected from various websites and the CGWB. The data includes GWL readings from 267 observation wells spanning 1995–2020. These wells vary in type, including dug wells, bore wells, and tube wells. The CGWB, the central authority for groundwater monitoring in India, takes seasonal readings in these wells four times a year: during the monsoon (June–September), post-monsoon (October–November), winter (December–January), and pre-monsoon (February–May). However, the southern part of Chhattisgarh, known as the Naxalite area, lacks observation wells, resulting in uniform GWLs in that region.

In this study, we used two climatic variables, precipitation, and temperature, as input data. Precipitation is essential for replenishing groundwater reserves and directly influencing water table fluctuations. In contrast, the temperature controls the groundwater recharge and depletion. The historical climate data had a spatial resolution of 12 × 12 km, while the future climate data were available at a resolution of 0.25° × 0.25° (approximately 27.5 × 27.5 km). The future climate data were biased, corrected, and downscaled and are freely available (Mishra et al. 2020).

On the other side, the lithology condition determines the ability of subsurface formations to transmit water, impacting groundwater movement. We considered the physical characteristics of rocks, specifically in the central part of Chhattisgarh, where there are nine types of lithological conditions. These include basalt, charnokite, gneiss, gneissic complex (basement), granite, limestone, sandstone, schist, and shale. Each lithology type has distinct properties and varying hydraulic conductivity values. The number of wells available in each lithology condition is mentioned in Figure 1. Most lithological types are relatively impermeable, resulting in low water-holding capacity. Consequently, the hydraulic conductivity values for these rocks are very low. It is incorporated as a continuous predictor to account for the influence of lithology on groundwater dynamics. Each lithological type is assigned a specific hydraulic conductivity value (Table 2).

The methodology begins by defining five key input variables: rainfall, temperature, hydraulic conductivity of lithology, LULC, and historical groundwater data. The data are arranged in a time-series format suitable for model input. The input data are then divided into two categories: features and labels. Features represent the independent variables (e.g., rainfall, temperature, lithology, and LULC) that the model uses to make predictions, while labels correspond to the output variable (historical GWL). The dataset is split into two periods, i.e. training and testing sets. The training set, which constitutes most of the data (60%), trains the DL model and recognizes the patterns in the input features. This ensures that the model learns valuable relationships between the input variables and the output. Testing ensures that the model generalizes well to unseen data by providing a final validation of its predictive accuracy. The same training and testing sets were used throughout the study to ensure consistency in model evaluation across different SSP scenarios. To maintain comparability, this dataset remained unchanged across all SSP scenarios. Once the model was trained, CMIP6 climate projections were used to forecast GWLs under different SSPs. This modified approach, which incorporates climatic factors and indirect drivers like lithology and LULC, enables robust seasonal GWL predictions. It provides a practical framework for applying DL techniques to address complex groundwater forecasting challenges.

The future climatic data of historical and projected annual average precipitation in Chhattisgarh show significant variations across different temporal periods and SSP scenarios. Historically, from 1951 to 2014, precipitation exhibited considerable interannual variability without a clear increasing or decreasing trend. In the near future (2015–2045), precipitation under SSP126 (a low-emission scenario) will remain relatively stable, closely resembling historical trends. In contrast, SSP245, SSP370, and SSP585 show slight variability increases, indicating the onset of changing rainfall patterns. In the mid-future (MF) (2046–2070), variability becomes more pronounced, especially under SSP370 and SSP585. These scenarios project more frequent and intense rainfall deviations compared to SSP126, which maintains a steady trend, and SSP245, which shows moderate variability. By the far-future (FF) period (2071–2100), the divergence among SSPs is more apparent. SSP126 stabilizes with minimal changes, while SSP245 and SSP370 exhibit moderate variability. SSP585, the high-emission scenario, shows the most significant changes, with sharp spikes in precipitation, indicating a higher likelihood of extreme rainfall events.

Similarly, annual average maximum and minimum temperature projections also show a clear warming trend, with distinct patterns across the NF, MF, and FF periods. Historically, maximum and minimum temperatures gradually increased, reflecting the warming climate. In the near future (2021–2045), maximum and minimum temperatures will rise modestly under SSP126, with trends similar to historical patterns. Under SSP245 and SSP370, the warming is more pronounced, while SSP585 shows the steepest increase, suggesting the onset of significant heat stress in the region. During the MF period (2046–2070), the warming trend intensifies across all scenarios, with SSP126 exhibiting the slowest rise in temperatures due to its low-emission pathway. SSP245 and SSP370 indicate moderate increases, while SSP585 projects rapid warming, with maximum temperatures exceeding 36 °C and minimum temperatures surpassing 25 °C. By the FF (2071–2100), SSP585 shows alarming increases, with maximum temperatures exceeding 37 °C and minimum temperatures nearing 27 °C, signaling severe heat stress. In contrast, SSP126 stabilizes, and SSP245 and SSP370 continue moderate warming trends.

The model accurately captures seasonal and interannual variations in GWLs in both wells, with close alignment between observed (solid blue line) and simulated (dashed green line) values. In well W201920081514501, the model shows high predictive capability during both the calibration and validation, with minor deviations during peak levels. Similarly, in well W215130081550001, the model replicates the observed trends effectively, though slight underestimations occur during specific high-water-level periods in the validation phase. These results demonstrate the robustness of the model in reproducing temporal groundwater fluctuations under varying conditions. However, outliers corresponding to extreme monsoon years also exist, where sudden high-intensity rainfall events lead to short-term groundwater spikes, which are not fully captured by the model's learning process. The model also underestimates groundwater declines in severely drought-affected years, where precipitation was significantly lower than the long-term mean. The observed values may sometimes contain measurement inaccuracies, particularly in shallow wells where fluctuations can be more pronounced.

Hence, the model performance has been checked using statistical metrics, the coefficient of determination (R²), and NSE across all the wells for both calibration and validation. The median R² values for calibration and validation are 0.83 and 0.84, respectively, indicating a strong correlation between observed and simulated data during both phases (Figure 4(c)). Similarly, the median NSE values for calibration and validation are 0.81 and 0.82, respectively, reflecting high model efficiency in reproducing observed GWLs (Figure 4(c)). Also, the narrow interquartile ranges for both R² and NSE highlight consistent performance across wells with minimal variability. This performance consistency across wells underscores the model's applicability and gives confidence to use this model for future scenarios. The accuracy was further evaluated using RMSE and MAE, which quantify the deviation between observed and predicted values (Table 3). The RMSE values of 1.35 m for calibration and 1.05 m for validation indicate that the model maintains a low overall prediction error, with improved accuracy in the validation period. The MAE value for calibration (validation) is 0.92 m (0.87 m). The decrease in RMSE and MAE during validation suggests that the model generalizes well beyond the training dataset without significant overfitting. These results demonstrate the model's effectiveness in accurately capturing GWL variations. Recent studies on groundwater forecasting using DL have primarily focused on climatic factors, overlooking indirect influences. Malakar et al. (2021) used long short term memory (LSTM) models to predict groundwater depletion across India but did not consider land cover or geological variations. Similarly, Feng et al. (2024) achieved high accuracy with convolutional neural network (CNN)-based predictions in Iran, yet their study lacked indirect factors. Sivakumar et al. (2024) forecasted groundwater trends in the Chennai Basin using climate variables, limiting real-world applicability. Wunsch et al. (2021) compared LSTM, CNN, and NARX models for GWL forecasting using meteorological inputs. Their study found NARX models to be effective but sensitive to initialization, while CNNs were computationally efficient. Unlike these studies, our model integrates LULC and lithology, recognizing urbanization's impact on infiltration and groundwater recharge. By incorporating these factors, our model demonstrates strong predictive accuracy, with NSE values of 0.81–0.82 and R² values of 0.83–0.84. The reduced RMSE (1.35 m calibration and 1.05 m validation) confirms its reliability. These results highlight the necessity of including surface and subsurface interactions in groundwater modeling, making our approach more robust for long-term water resource planning.

The analysis of GWL variations under the SSPs provides a comprehensive understanding of the potential impacts of climate change on water resources in Chhattisgarh. Four SSP scenarios, SSP126 (sustainability), SSP245 (middle of the road), SSP370 (regional rivalry), and SSP585 (fossil-fueled development), offer distinct trajectories for socio-economic development and greenhouse gas emissions, which directly influence precipitation, temperature, and, consequently, groundwater recharge and withdrawal patterns (Rogelj et al. 2016; Siabi et al. 2023). Each scenario captures a unique range of GWL variations across seasonal periods and temporal scales. Detailed results are described as follows.

In the MF (2045–2070), for well W201920081514501, seasonal patterns remain similar, but an overall decline in GWLs is evident, particularly during S2 and S4 (September–November). This decline suggests a gradual reduction in recharge potential, possibly linked to precipitation intensity and frequency changes. The interannual variability becomes more pronounced, reflecting potential shifts in climate-induced hydrological processes. For W215130081550001, a noticeable reduction in GWLs is projected, particularly in season 2 and season 4. Seasonal variability remains distinct, but the magnitude of interannual fluctuations increases, suggesting increased susceptibility to climate variability.

In the FF (2071–2100), a significant depletion of GWLs is observed across all seasons, with levels consistently below 10 mbgl during season 2 and season 4. However, season 3 (monsoon season) shows a slight recovery compared to other seasons, emphasizing the role of monsoon rainfall in mitigating depletion. Subsequently, for W215130081550001 season 4 (post-monsoon), critical depletion was shown below 8 mbgl. The monsoon season (season 3) continues to provide some respite, but its recharge potential appears insufficient to counterbalance the cumulative effects of reduced recharge and increased withdrawal.

SSP245 reflects moderate emissions and a balanced approach between sustainability and fossil-fuel dependency (Fricko et al. 2017). In the NF, both wells display seasonal trends with relatively stable GWLs. Well W201920081514501 shows GWLs ranging between 4 and 12 mbgl, with higher levels during season 3 (June–August) due to monsoon recharge and lower levels during season 2 (February–May) and season 4 (September–November), reflecting seasonal withdrawal patterns (Figure 5(b)). Similarly, well ID W215130081550001 demonstrates comparable trends, with GWLs ranging from 2 to 10 mbgl. The monsoon season (season 3) consistently contributes to peak GWLs in both wells, while pre-monsoon (season 2) levels remain the lowest. Interannual fluctuations are moderate, reflecting steady recharge and extraction dynamics under SSP245 (Figure 6(b)).

In the MF, a notable decline in GWLs emerges across both wells, particularly during the non-monsoon seasons (season 2 and season 4). W201920081514501 experiences a reduction in levels, with season 2 and season 4 showing more pronounced depletion compared to the NF, while season 3 retains relatively higher levels due to monsoon recharge. Well W215130081550001 exhibits similar patterns, with GWLs declining significantly, especially during season 2 and season 4, where levels consistently approach 6 mbgl or lower. Interannual variability becomes more pronounced for both wells, indicating increased susceptibility to climatic and anthropogenic pressures during this period.

Subsequently, significant groundwater depletion is evident in both wells under SSP245 by the FF period. For well ID W201920081514501, GWLs consistently remain below 10 mbgl during all seasons, with season 2 (pre-monsoon) and season 4 (post-monsoon) showing critical depletion. Season 3 (monsoon season) provides some recovery, but the recharge potential appears insufficient to counterbalance overall depletion. Well W215130081550001 follows a similar trajectory, with GWLs declining further, particularly during non-monsoon seasons where levels drop to critical lows of 4 mbgl or less. The FF projections highlight the increasing vulnerability of groundwater resources under SSP245, emphasizing the need for proactive management strategies to mitigate long-term impacts.

SSP370 characterizes high emissions, regional conflicts, and limited international cooperation (Meinshausen et al. 2011). Under the SSP370 scenario, GWL projections for both wells illustrate the impacts of high greenhouse gas emissions, limited international cooperation, and regional conflicts. In the NF, both wells will exhibit moderate seasonal oscillations, with monsoon recharge contributing to peak GWLs and pre-monsoon and post-monsoon levels showing lower levels. For well W201920081514501, GWLs range between 4 and 10 mbgl, with higher interannual variability than observed under SSP245 (Figure 5(c)). W215130081550001 also follows a similar trend, with levels varying between 3 and 8 mbgl. The increasing fluctuations indicate a growing sensitivity to climatic pressures even in the early period of SSP370 (Figure 6(c)).

In the MF, the adverse impacts of SSP370 become more evident. GWLs for well W201920081514501 show a pronounced decline, particularly during season 2 and season 4, where levels consistently approach 6 mbgl or lower. S3 (monsoon) continues to provide seasonal recovery, but the recharge potential is visibly reduced compared to the NF. W215130081550001 exhibits similar patterns, with GWLs dropping further during non-monsoon seasons and season 3 providing only partial mitigation. Increased interannual variability during this period underscores the influence of higher emissions and regional climate impacts on groundwater recharge and withdrawal dynamics.

In the FF period, GWLs in both wells show severe depletion under SSP370. Well W201920081514501 consistently experiences levels below 10 mbgl across all seasons, with season 2 and season 4 reaching critical lows near 8 mbgl. Season 3 still offers a slight recovery, but the overall trend is one of significant depletion. W215130081550001 follows a similar trajectory, with GWLs falling below 6 mbgl during non-monsoon seasons and season 3 providing diminishing recharge benefits. The FF projections reflect the compounded effects of climate variability and increased anthropogenic pressures, highlighting the urgent need for robust groundwater management strategies under high-emission scenarios.

SSP585 depicts the most extreme high-emission scenario driven by intensive fossil fuel use and minimal climate mitigation efforts. The scenario highlights the extreme stress on water resources due to intensified climate variability, rapid temperature rise, and uncontrolled anthropogenic pressures. In the NF, GWLs for both wells will exhibit moderate seasonal fluctuations, with the monsoon season showing higher levels due to monsoon recharge. In contrast, others will demonstrate consistent depletion. For well ID W201920081514501, levels range between 6 and 12 mbgl, while well ID W215130081550001 ranges from 4 to 10 mbgl (Figures 5(d) and 6(d)). Although the seasonal patterns are still evident, the amplitude of interannual variability begins to increase, signaling early signs of stress under high emissions.

In the MF, both wells experience noticeable declines in GWLs, with pre-monsoon and post-monsoon showing critical reductions. Well ID W201920081514501 sees levels consistently approaching 8 mbgl during non-monsoon seasons, while well ID W215130081550001 exhibits similar trends with levels nearing 6 mbgl. Seasonal recovery during the monsoon becomes less pronounced as the recharge potential decreases due to climatic variability. The widening interannual fluctuations highlight the increasing vulnerability of groundwater resources under SSP585. In the FF, GWLs in both wells will show significant and widespread depletion. Well W201920081514501 consistently exhibits levels below 12 mbgl across all seasons, with season 2 and season 4 reaching critical lows near 10 mbgl or more. Similarly, well W215130081550001 shows extreme depletion, falling below 8 mbgl during non-monsoon seasons. Even monsoons fail to recover substantially, reflecting the compounded effects of uncontrolled emissions and climate variability. This period underscores the dire need for adaptive strategies to mitigate the catastrophic impacts of SSP585.

The differences in GWLs across SSP scenarios primarily stem from changes in precipitation and evapotranspiration rates, which influence groundwater recharge and depletion. In low-emission scenarios (SSP126 and SSP245), moderate precipitation and relatively stable temperatures contribute to gradual groundwater fluctuations, allowing for a more balanced recharge-depletion cycle. However, in high-emission scenarios (SSP370 and SSP585), increasing temperatures accelerate evapotranspiration, reducing groundwater recharge and intensifying water demand, leading to faster depletion. Furthermore, regions dominated by high-permeability formations exhibit better recharge potential, while low-permeability formations restrict infiltration, making GWLs more susceptible to climatic variations. Similarly, LULC changes significantly affect GWLs, with urban expansion increasing impervious surfaces, leading to reduced infiltration, whereas agricultural intensification leads to higher groundwater abstraction.

In the NF, both low-emission scenarios, such as SSP126 and 245, show slight increases in GWL variations compared to the baseline. Specifically, SSP126 records a median variation of 5.22 mbgl, while SSP245 shows 5.02 mbgl, reflecting moderate deviations under sustainable and middle-of-the-road development pathways. These scenarios indicate that even under controlled emission trajectories, minor impacts on GWLs are expected due to climatic shifts. Conversely, SSP370 and SSP585 project significantly higher variations, with a median of 7.16 and 7.85 mbgl, respectively. These high-emission scenarios exhibit broader interquartile ranges and higher maximum values, suggesting more pronounced and uncertain GWL declines than the historical baseline. In the MF, the predicted variations in GWLs increase across all scenarios, marking a clear intensification of climate change impacts. SSP126 and SSP245 continue to show moderate increases, with median variations of 6.43 and 6.69 mbgl, respectively, indicating that sustainable practices can partially mitigate groundwater stress. However, SSP370 and SSP585 predict substantial deviations, with median variations increasing to 8.10 and 9.42 mbgl, respectively. The higher variability in SSP370 and SSP585 reflects the magnitude of groundwater depletion and the uncertainty associated with high-emission pathways. These trends highlight the growing challenges in managing groundwater resources as emissions rise and climate impacts intensify. Subsequently, in the FF, the gap between low- and high-emission scenarios becomes increasingly apparent. SSP126 and SSP245 project mean GWL variations of 7.42 and 7.58 mbgl, respectively, suggesting that these pathways can prevent the most extreme outcomes. However, SSP370 and SSP585 show drastic increases, with median variations reaching 11.02 and 12.14 mbgl, respectively. These scenarios exhibit the highest median values and the largest ranges of variability, indicating severe and widespread groundwater stress under continued high emissions. The broader interquartile ranges and extreme values further underscore the uncertainty and risks associated with unchecked climate change.

Season 2, which covers the pre-monsoon period, exhibits the highest stress in GWLs. The baseline median variation is 5.00 mbgl, higher than in other seasons. Under SSP126 and SSP245, the mean GWLs increase to 7.05 and 7.16 mbgl, respectively, showing moderate deviations. However, SSP370 and SSP585 demonstrate significant increases. It shows relatively moderate impacts in the monsoon period compared to season 2. SSP126 and SSP245 indicate that sustainable development pathways help maintain GWLs closer to historical conditions. On the other hand, SSP370 and SSP585 predict median variations of 7.96 and 8.46 mbgl, respectively, showing a marked increase in groundwater stress under high-emission scenarios. Season 4, marking the post-monsoon period, shows a baseline mean variation similar to season 3. Under SSP126 and SSP245, GWL variations increase modestly to 6.54 and 6.86 mbgl, respectively, reflecting the relative stability provided by low-emission pathways. However, SSP370 and SSP585 indicate significant increases, reaching 9.65 and 10.43 mbgl, respectively. The broader interquartile ranges and extreme values in these scenarios suggest persistent groundwater stress even after the recharge period, emphasizing the long-term effects of high emissions.

This model gives promising results, but it has some limitations. The model's performance may also change under different hydrological conditions. The accuracy of predictions may be affected by uncertainties in the input data, particularly climate model outputs and groundwater observations. One key limitation arises from the uncertainties in CMIP6 climate model data. Despite bias-correction techniques, global climate models (GCMs) exhibit inherent biases and varying levels of accuracy across regions, which may affect precipitation and temperature projections, two critical variables influencing groundwater recharge and depletion. The spatiotemporal resolution of GWL data makes this model quite unstable. The limited temporal frequency can reduce the ability of ML models to accurately predict groundwater responses to extreme climatic events or localized anthropogenic withdrawals.

This study also opens opportunities for future research to improve groundwater modeling. Future research should also explore applying advanced techniques, such as attention-based DL models and ensemble learning approaches, to enhance predictive capabilities further. A comprehensive sensitivity analysis should be conducted to identify the most influential variables affecting GWL predictions. Additionally, integrating groundwater forecasting models with socio-economic datasets, real-time data integration, and cloud-based tools can provide a more holistic understanding of the interplay between climate change, human activities, and groundwater dynamics.

This study presents a DL-based approach for forecasting seasonal GWLs in Chhattisgarh, India, by incorporating direct (precipitation and temperature) and indirect (lithology and LULC dynamics) drivers. Unlike conventional models that rely solely on climate data, this study integrates additional geospatial and geological parameters, significantly improving predictive accuracy. This study's key findings provide critical insights into GWL fluctuations in four seasons in Chhattisgarh state. Furthermore, the model was trained and validated using historical groundwater data, demonstrating strong performance, with median NSE and R² values of 0.81 and 0.83 in calibration and 0.82 and 0.84 in validation, respectively. Additionally, the RMSE values were 1.35 m for calibration and 1.05 m for validation, while the MAE was 0.92 and 0.87 m, confirming the model's robustness in simulating groundwater variations. It demonstrates that considering indirect drivers enhances predictive performance and provides a more reliable representation of groundwater dynamics. Also, this study further analyzed groundwater variations under different SSP scenarios, highlighting distinct trends. Some key findings are as follows:

• NF (2021–2045): SSP126 and SSP245 exhibit slight groundwater variations, indicating moderate climate impact with manageable depletion levels. In contrast, SSP370 and SSP585 show significant groundwater decline, highlighting increasing stress and vulnerability.

• MF (2046–2070): Groundwater stress intensifies, with SSP126 and SSP245 experiencing moderate depletion. However, SSP370 and SSP585 show severe groundwater depletion accompanied by heightened variability, emphasizing the escalating impact of high-emission scenarios.

• FF (2071–2100): SSP126 and SSP245 continue to demonstrate resilience, with moderate groundwater depletion. Conversely, SSP370 and SSP585 exhibit extreme groundwater loss, indicating critical scarcity risks and long-term unsustainability.

• Winter (season 1): Higher variability in high-emission scenarios (SSP370: 7.96 mbgl, SSP585: 9.08 mbgl).

• Also, the study shows that the pre-monsoon season is the most stressful period, with rapid depletion followed by winter. In monsoon, it shows recovery, but high emissions reduce recharge efficiency.

Both authors contributed to the study at all levels and original draft preparation. M. K. D. developed the methodology, visualized the process, wrote the original draft; C. K. S. conceptualized the work, wrote the reviewed, and edited the final draft.

No funding was used in this research.

Data will be available upon reasonable request.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.