Uncertainty-Aware Reconstruction of Groundwater-Level Gaps Under Unmonitored Irrigation Stress

Long gaps in groundwater-level time series limit hydrogeological interpretation, especially when the missing period extends over months or years and the aquifer is affected by unmonitored anthropogenic stresses. This short note presents a workflow for reconstructing groundwater-level records when irrigation withdrawals are known to occur but pumping rates are not measured. Missing values in the target series are reconstructed using an ensemble of long short-term memory (LSTM) neural networks conditioned by hydro-meteorological and groundwater covariates. To preserve temporal continuity in the model inputs, missing auxiliary covariates are completed before training, and binary indicators are added to distinguish observed from reconstructed covariate values. To account for unobserved abstraction pressure, an irrigation-stress index is constructed from proxy variables, including air temperature, precipitation deficit, dry-spell persistence, irrigation seasonality, and auxiliary depth-to-water-level observations. The reconstruction is repeated through independent LSTM runs, allowing an empirical assessment of reconstruction variability. The workflow is illustrated using a groundwater-level time series collected at the Science and Technology Campus of the University of Parma, Northern Italy. Artificial-gap validation and benchmark comparison show that conventional interpolation is preferable for short gaps, whereas the LSTM ensemble with the irrigation-stress index provides the clearest benefit for intermediate gaps. Annual-scale reconstructions remain uncertain and should be interpreted with caution. However, repeated LSTM runs allow the construction of an ensemble envelope based on two standard deviations across independent reconstructions, providing an empirical visualization of reconstruction variability. The proposed approach offers a reproducible, data-informed procedure for groundwater-level gap reconstruction in systems affected by unmeasured irrigation stress.