Climate change is altering the hydrological dynamics of climate-sensitive systems at an accelerating rate. In the Arctic, permafrost degradation, glacier retreat, and shifts in precipitation regime are driving rapid changes in runoff generation, subsurface storage, and groundwater-surface water interactions. In Mediterranean environments, increasing drought frequency and intensifying precipitation extremes are modifying aquifer recharge and storage dynamics. Although controlled by different processes, the Arctic and the Mediterranean represent two highly sensitive end-members of climate change impacts, cryosphere degradation on the one hand and drought intensification on the other, whose distinct responses provide complementary testing grounds for hydrological modelling frameworks. Characterizing these responses requires both complete observational records and analytical frameworks that can identify the main controls on water level variability and their temporal evolution. A persistent constraint on robust inference in these settings is the frequent incompleteness of long-term monitoring records: data gaps alter temporal dependence structure, bias the estimation of hydrological variability, and compromise the statistical basis for trend detection, seasonality analysis, and model calibration. This PhD thesis addresses two intrinsically connected methodological challenges in the context of climate change: the development of statistically robust gap filling procedures for hydrological and hydrogeological time series, and the characterization of water level dynamics in climate-sensitive systems through data-driven modelling. The research focuses on a Mediterranean coastal aquifer on Pianosa Island (Tyrrhenian Sea, Italy) and three Arctic catchments on Svalbard (Bayelva, Fuglebekken, and Ariebekken), two systems characterised by different but equally pronounced hydrological responses to climate forcing. For each system, gap filling and data-driven modelling are treated as a single integrated analytical workflow rather than as independent sequential steps. For the Pianosa coastal aquifer, a comparative evaluation of eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) recurrent neural networks for piezometric data imputation shows that ensemble tree-based methods consistently outperform sequential deep learning architectures, reflecting their capacity to capture complex non-linear predictor-response associations without requiring explicit temporal sequence modelling. For the Bayelva catchment, weighted regression and mixed effects models calibrated over three consecutive melt seasons (2022-2024) reveal systematic interannual shifts in predictor importance, with the dominant statistical control alternating between atmospheric thermal forcing and the storage state of the active layer–groundwater system. These shifts are consistent with a progressive, climate-related change in the controls on water level under sustained Arctic warming, and scenario-based projections indicate a continued increase of total water fluxes through 2040. For the Fuglebekken and Ariebekken catchments, multivariate regression and ensemble modelling reveal marked contrasts between non-glaciated and glacier-fed systems: in the former, water level variability is most strongly associated with groundwater dynamics and soil moisture conditions, whereas in the latter it is most strongly associated with proglacial soil moisture conditions and the ground thermal regime. Taken together, the results show that workflows combining regression-based modelling and machine learning algorithms can achieve both predictive skill and physical interpretability when applied to hydroclimatic dynamics. At the same time, the interannual shifts in predictor influence and the progressive decline in explanatory power under sustained climatic forcing expose the structural limits of static empirical formulations and support the development of hybrid frameworks, in which statistical flexibility is coupled with explicit process understanding. This dual-environment approach demonstrates that a common analytical framework can be transferred across fundamentally different climate-sensitive systems, provided that model assumptions and the physical meaning of the predictors are re-examined in each setting, thereby contributing to the adaptation of data-driven hydrological modelling to non-stationary environmental conditions. The main limitations of the work concern the short duration of the Arctic time series, the uncertainty affecting scenario-based projections, and the site-specific nature of the calibrated relationships, which constrains their direct transferability. Continuous, high-quality, long-term monitoring therefore remains a prerequisite for improving hydrological prediction in systems undergoing accelerated climate-driven change.
Water Level Data-Driven Hydrological Modelling in Climate-Sensitive Systems
CHIMENTI, MARCO
2026
Abstract
Climate change is altering the hydrological dynamics of climate-sensitive systems at an accelerating rate. In the Arctic, permafrost degradation, glacier retreat, and shifts in precipitation regime are driving rapid changes in runoff generation, subsurface storage, and groundwater-surface water interactions. In Mediterranean environments, increasing drought frequency and intensifying precipitation extremes are modifying aquifer recharge and storage dynamics. Although controlled by different processes, the Arctic and the Mediterranean represent two highly sensitive end-members of climate change impacts, cryosphere degradation on the one hand and drought intensification on the other, whose distinct responses provide complementary testing grounds for hydrological modelling frameworks. Characterizing these responses requires both complete observational records and analytical frameworks that can identify the main controls on water level variability and their temporal evolution. A persistent constraint on robust inference in these settings is the frequent incompleteness of long-term monitoring records: data gaps alter temporal dependence structure, bias the estimation of hydrological variability, and compromise the statistical basis for trend detection, seasonality analysis, and model calibration. This PhD thesis addresses two intrinsically connected methodological challenges in the context of climate change: the development of statistically robust gap filling procedures for hydrological and hydrogeological time series, and the characterization of water level dynamics in climate-sensitive systems through data-driven modelling. The research focuses on a Mediterranean coastal aquifer on Pianosa Island (Tyrrhenian Sea, Italy) and three Arctic catchments on Svalbard (Bayelva, Fuglebekken, and Ariebekken), two systems characterised by different but equally pronounced hydrological responses to climate forcing. For each system, gap filling and data-driven modelling are treated as a single integrated analytical workflow rather than as independent sequential steps. For the Pianosa coastal aquifer, a comparative evaluation of eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) recurrent neural networks for piezometric data imputation shows that ensemble tree-based methods consistently outperform sequential deep learning architectures, reflecting their capacity to capture complex non-linear predictor-response associations without requiring explicit temporal sequence modelling. For the Bayelva catchment, weighted regression and mixed effects models calibrated over three consecutive melt seasons (2022-2024) reveal systematic interannual shifts in predictor importance, with the dominant statistical control alternating between atmospheric thermal forcing and the storage state of the active layer–groundwater system. These shifts are consistent with a progressive, climate-related change in the controls on water level under sustained Arctic warming, and scenario-based projections indicate a continued increase of total water fluxes through 2040. For the Fuglebekken and Ariebekken catchments, multivariate regression and ensemble modelling reveal marked contrasts between non-glaciated and glacier-fed systems: in the former, water level variability is most strongly associated with groundwater dynamics and soil moisture conditions, whereas in the latter it is most strongly associated with proglacial soil moisture conditions and the ground thermal regime. Taken together, the results show that workflows combining regression-based modelling and machine learning algorithms can achieve both predictive skill and physical interpretability when applied to hydroclimatic dynamics. At the same time, the interannual shifts in predictor influence and the progressive decline in explanatory power under sustained climatic forcing expose the structural limits of static empirical formulations and support the development of hybrid frameworks, in which statistical flexibility is coupled with explicit process understanding. This dual-environment approach demonstrates that a common analytical framework can be transferred across fundamentally different climate-sensitive systems, provided that model assumptions and the physical meaning of the predictors are re-examined in each setting, thereby contributing to the adaptation of data-driven hydrological modelling to non-stationary environmental conditions. The main limitations of the work concern the short duration of the Arctic time series, the uncertainty affecting scenario-based projections, and the site-specific nature of the calibrated relationships, which constrains their direct transferability. Continuous, high-quality, long-term monitoring therefore remains a prerequisite for improving hydrological prediction in systems undergoing accelerated climate-driven change.| File | Dimensione | Formato | |
|---|---|---|---|
|
PhD_Thesis_Marco_Chimenti.pdf
embargo fino al 10/09/2029
Licenza:
Creative Commons
Dimensione
17.44 MB
Formato
Adobe PDF
|
17.44 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/385766
URN:NBN:IT:UNIPI-385766