Input-data perturbation in groundwater level prediction: a controlled experiment on accuracy–reliability trade-offs among data augmentation and Bayesian-type uncertainty estimation methods

Groundwater is a major water resource for irrigation and domestic use, especially in arid and semi-arid regions where freshwater is scarce and the total amount of rainfall is limited. Hence, monitoring both quality and quantity parameters of this source is necessary for optimal water resources management, environmental studies, and ecosystem modeling, as well as ecological planning. Groundwater level (GWL) variation is influenced by numerous driving factors, both natural and human-induced. Such external impacts introduce considerable uncertainty into GWL variations, which should be taken into account in hydrological modeling processes of a catchment. This study reports a controlled perturbation experiment that assessed how prescribed input-data errors in GWL drivers (e.g., meteorological variables) propagated to GWL predictions. As the imposed perturbations were synthetic rather than derived from the measurement-error characteristics of the source records, the analysis was framed as a robustness and error-propagation assessment rather than as a validated quantification of total predictive uncertainty. Three data augmentation methods (jittering, frequency domain, and test-time augmentation (TTA)) and an IBUNE-type multiplicative stochastic input perturbation, derived from the integrated Bayesian uncertainty estimator (IBUNE) framework but implemented here without its Bayesian inference step, were applied to introduce Gaussian noise at various levels into the meteorological input variables. The Random Forest (RF) model was then used to predict GWL on a daily time step using these augmented datasets, and an unperturbed RF model sharing the same chronological partition and hyperparameter configuration was used as a baseline (test NSE = 0.885, RMSE = 0.034 m). All reported accuracy indices refer to the ensemble-mean prediction series, whereas coverage and interval-width indices refer to the corresponding 95% ensemble intervals. Jittering and IBUNE provided higher point accuracy levels (with NSE and RMSE values of 0.87 and 0.036 m, respectively) while providing narrower uncertainty bands with low 95% coverages (18.0–24.1% for jittering and 34.1–36.8% for IBUNE). Frequency-domain augmentation produced moderate accuracy (NSE = 0.83 and RMSE = 0.041 m) while improving reliability (approximately 61% coverage) through wider bands. TTA presented the highest coverage (60.7% at 10% noise, rising to 83.6% at 40% noise) with wide intervals and reduced accuracy under higher noise levels. The results revealed that the jittering and IBUNE methods presented the highest prediction accuracy, while TTA produced the highest uncertainty coverage (inclusion rate). The frequency-domain method showed the most balanced trade-off between accuracy and uncertainty. Measured against the unperturbed baseline, none of the four schemes improved point accuracy. Each scheme reduced NSE, by as little as 0.004 for jittering at 10% noise and by as much as 0.066 for TTA at 40% noise, so the wider intervals were purchased with accuracy rather than gained alongside it. Because no configuration approached the nominal 95% level, the reported intervals should be read as measures of perturbation-induced spread rather than as calibrated predictive uncertainty.