Machine Learning Outperforms Classical Time Series Models in Coastal Aquifer Forecasting

A long-term hydrological investigation in the climate-sensitive coastal district of Barishal, Bangladesh, has established that nonlinear machine learning algorithms markedly outperform traditional statistical time-series techniques when forecasting spatiotemporal groundwater level dynamics. Evaluating continuous monthly hydrological records collected from twelve observation stations between January 1996 and March 2025, researchers benchmarked the predictive capabilities of four distinct mathematical architectures: Seasonal Autoregressive Integrated Moving Average, Long Short-Term Memory deep recurrent neural networks, Extreme Gradient Boosting, and Random Forest ensemble algorithms. Across the district-wide monitoring infrastructure, the data-driven machine learning models consistently surpassed the linear baseline, successfully reproducing seasonal recharge and recession cycles while tracking subtle localized shifts in the groundwater table. Overall predictive accuracy across the entire observation network was spearheaded by the Random Forest model, which recorded the lowest mean Root Mean Square Error of 0.123 meters and the highest mean Nash-Sutcliffe Efficiency coefficient of 0.930, demonstrating exceptional mathematical stability across divergent hydrogeological settings.

The study focused on Barishal District, a low-lying deltaic floodplain encompassing approximately 2,790 square kilometers situated along the south-central coast of Bangladesh between 22.2 to 22.9 degrees North latitude and 90.0 to 90.8 degrees East longitude. Comprising ten administrative upazilas—Agailjhara, Babuganj, Bakerganj, Banaripara, Gaurnadi, Hizla, Mehendiganj, Muladi, Wazirpur, and Barishal Sadar—the territory forms an integral component of the Lower Bengal Basin within the complex Ganges-Brahmaputra-Meghna delta network. The region is characterized by unconsolidated Holocene alluvial and deltaic sediment strata consisting predominantly of interlaced sand, silt, and clay. Nationwide statistics indicate that nearly eighty-seven percent of the Bangladeshi population relies directly upon subterranean aquifers for drinking water, while more than seventy percent of all domestic agricultural irrigation depends upon tubewell abstraction, predominantly mobilized during the arid dry season to cultivate Boro rice crops across low-lying floodplains.

Compounding this intensive anthropogenic extraction, the regional hydrological system faces severe environmental disruption from shifting climatic patterns and maritime exposure. While annual rainfall historically averages between 2,400 and 2,800 millimeters—concentrated almost entirely within the southwest monsoon spanning June through October—natural aquifer replenishment has exhibited alarming long-term alterations over the preceding three decades. Concurrently, regional atmospheric warming has accelerated moisture loss: mean local temperatures expanded by 0.7 degrees Celsius between 1980 and 2020, with climate projections anticipating an additional 1.5 degrees Celsius increase by the middle of the century. Historical field surveys conducted by the Bangladesh Water Development Board documented alarming declines in pre-monsoon groundwater tables, with seasonal dry-period depressions sinking between 2.5 and 4.8 meters below the ground surface across stressed municipal sectors.

To assemble a rigorous, multi-decade database for mathematical modeling, monthly groundwater level measurements from twelve designated Bangladesh Water Development Board monitoring wells were synchronized with comprehensive hydro-climatic datasets. Seven supplementary meteorological predictors were collated from national scientific archives: maximum air temperature, minimum air temperature, precipitation, relative humidity, soil moisture content, evapotranspiration rates, and surface wind speed. Precipitation records were acquired from the Bangladesh Water Development Board; temperature, humidity, and wind parameters were sourced from the Climate Information Management System under the Bangladesh Agricultural Research Council; and soil moisture and evapotranspiration data were extracted from the National Agricultural Drought Watch platform. Because atmospheric conditions across Barishal remain relatively uniform, a unified district-scale climatic forcing dataset was integrated across all monitoring sites, whereas local subsurface sediment characteristics, surface water connectivity, and extraction intensity dictated well-specific response dynamics.

Prior to algorithm training, the combined time series was subjected to strict data quality screening, removal of missing values, and chronological partitioning to safeguard model integrity and eliminate future information leakage. An eighty-to-twenty temporal holdout split reserved data from January 1, 1996 through December 31, 2020 for calibration and parameter training, while observations from January 1, 2021 through March 31, 2025 served as an independent verification benchmark. All continuous predictor variables underwent standardization using the Standard Scaler methodology, normalizing features by subtracting empirical means and dividing by standard deviations. To incorporate physical system memory and antecedent moisture retention, feature engineering generated three discrete lagged groundwater terms reflecting one-, two-, and three-month historical conditions, alongside a one-month lagged rainfall predictor to represent delayed percolation. Cyclic trigonometric sine and cosine transformations based on observation calendar months were applied to preserve recurring seasonal oscillations mathematically without introducing artificial boundary breaks.

The mathematical evaluation tested fundamentally differing computational paradigms to simulate subsurface aquifer storage. The traditional benchmark utilized a Seasonal Autoregressive Integrated Moving Average SARIMA(1,1,1)(1,1,1,12) configuration designed to capture linear autocorrelation and twelve-month cyclical structures. By contrast, the deep-learning Long Short-Term Memory recurrent neural network mobilized two hidden computational layers containing 128 and 64 memory units, sequentially governed by forget, input, and output gating mechanisms that regulate temporal state retention. Overfitting within the deep neural architecture was mitigated through a 0.30 dropout regularization rate, trained using an Adam optimizer, mean squared error loss criteria, a batch size of 16, a ceiling of 120 epochs, and early stopping patience fixed at 10 epochs. Operating over a 60-month historical lookback window, the univariate model utilized recursive autoregressive forecasting to project forward-marching water level trajectories.

The tree-based machine learning framework incorporated Extreme Gradient Boosting and Random Forest ensemble algorithms to process complex, nonlinear environmental interactions. The gradient boosting pipeline deployed twenty-five highly informative engineered variables, comprising lagged hydrological measurements, moving rolling statistics, interactive covariates, and cyclical markers, fine-tuned across 100 iterative trials via the Optuna hyperparameter optimization framework before assembling an ensemble of five distinct models trained with varied random seeds. In parallel, the Random Forest model optimized a forest of 700 independent decision trees configured without depth limitations and constrained to a minimum leaf size of two samples. Predictive performance across all candidate algorithms was evaluated against holdout observations using Root Mean Square Error, Mean Absolute Error, Mean Squared Error, the Coefficient of Determination, and the Nash-Sutcliffe Efficiency metric, with identical values naturally recorded between residual-based determination and efficiency scores due to mathematical equivalence.

Quantitative benchmarking revealed clear performance disparities between classical time-series analysis and nonlinear algorithms. The conventional SARIMA model exhibited the largest systematic errors across the regional network, generating Root Mean Square Errors ranging from 0.194 to 0.380 meters and Nash-Sutcliffe Efficiency ratings between 0.634 and 0.870. Machine learning methodologies substantially improved fidelity across all locations. Extreme Gradient Boosting captured the highest localized accuracy at six distinct monitoring wells—Babuganj, two stations in Bakerganj, Banaripara, Gaurnadi, and Hizla—achieving an absolute lowest single-station Root Mean Square Error of 0.0337 meters at Banaripara station GT0610003 and a peak Coefficient of Determination of 0.9885 at Bakerganj station GT0607009. The Long Short-Term Memory network proved most effective at two monitoring sites, Barishal Sadar and Agailjhara, delivering a network-wide average error of 0.159 meters and mean efficiency of 0.917. Random Forest attained supreme accuracy at four individual stations while demonstrating unmatched consistency across the entire district territory.

To verify whether these observed mathematical discrepancies represented genuine algorithmic superiority rather than stochastic error variation, pairwise Diebold-Mariano hypothesis tests were executed across all twelve stations at a five percent significance threshold. The statistical tests confirmed that Random Forest significantly outperformed SARIMA across every single observation well in the network (12 out of 12 stations, p < 0.05). The most pronounced statistical separation occurred at Banaripara well GT0610003, where the Diebold-Mariano test statistic plummeted to -10.4369 (p < 0.001). Furthermore, Random Forest generated statistically smaller prediction residuals than the Long Short-Term Memory network at eight monitoring locations, while demonstrating statistically superior accuracy compared to Extreme Gradient Boosting across eight wells. Residual analyses performed on Random Forest outputs at representative monitoring stations revealed errors centered tightly around the zero-reference datum without persistent directional skew, affirming the structural reliability of tree-based bagging algorithms for hydrological modeling.

Detailed representative well investigations underscored how localized hydrogeology controls groundwater responsiveness. At Barishal Sadar, groundwater conditions remained comparatively steady throughout the study timeframe, rising modestly from approximately 1.5 meters below surface in the late 1990s to roughly 3.0 meters by 2025. Here, all four algorithms replicated seasonal dynamics successfully, with the Long Short-Term Memory model recording superior precision, achieving an error of 0.1025 meters and efficiency of 0.949. Conversely, Hizla displayed pronounced hydrological vulnerability, characterized by violent seasonal oscillations ranging between 0.8 and 4.0 meters. Extreme Gradient Boosting demonstrated superior tracking of these rapid hydrological recoveries and drawdowns, attaining an error of 0.1114 meters and an efficiency of 0.9822. Predictions for 2026—generated by driving the calibrated groundwater models with future climate parameters derived from preliminary SARIMA meteorological forecasts—indicated that these distinct spatial behaviors will persist, with cyclic fluctuations dominating dynamic coastal reaches while inland sectors remain comparatively buffered.

To translate discrete point forecasts into actionable administrative insights, spatial vulnerability was mapped across Barishal District using geographic information systems, adopting a groundwater depth threshold of 3.05 meters (equivalent to 10 feet) below the surface as the critical indicator of operational water stress. Historical spatiotemporal maps illustrated a progressive expansion of subterranean vulnerability, transforming the district from predominantly low-risk classifications in 1996 into extensive tracts of high vulnerability by 2025. Spatial patterns across all predictive models for 2026 consistently pinpointed Hizla, Mehendiganj, Bakerganj, Muladi, and portions of Gaurnadi as chronically vulnerable zones, whereas Barishal Sadar, Agailjhara, and Banaripara maintained comparatively resilient aquifer reserves.

The investigative team noted several methodological boundaries inherent to their empirical framework. Spatially distributed meteorological stations were unavailable, requiring the application of uniform district-scale climate predictors across all twelve wells, which may obscure hyper-local atmospheric nuances. Furthermore, continuous historical monitoring data regarding physical groundwater pumping volumes, land-use evolution, river tidal stages, subsurface salinity gradients, and specific aquifer hydraulic transmissivity were inaccessible, preventing the explicit mathematical simulation of physical fluid mechanics. Consequently, the established machine learning models capture predictive statistical correlations rather than deterministic mechanical causation.

Subject of Research: Comparative evaluation of machine learning and statistical models for coastal groundwater forecasting

Article Title: Comparative evaluation of statistical and machine learning models for spatiotemporal groundwater level forecasting in a climate-sensitive coastal aquifer

Article References: Taskin, T., Munia, S. A., & Rasel, H. M. (2026). Comparative evaluation of statistical and machine learning models for spatiotemporal groundwater level forecasting in a climate-sensitive coastal aquifer. Environmental Earth Sciences, 85(17), Article 446. https://doi.org/10.1007/s12665-026-13131-9

Image Credits: AI Generated

DOI: 10.1007/s12665-026-13131-9

Keywords: Groundwater Forecasting, Machine Learning, Random Forest, XGBoost, LSTM, SARIMA, Hydrogeology, Coastal Aquifers, Vulnerability Mapping, Barishal District, Comparative, evaluation