Abstract
Advancements in big data analytics and high-performance computing have significantly expanded the ability to process and analyse geospatial data cubes using parallel computing and data driven methodologies. This thesis investigates the application of advanced artificial intelligence (AI) techniques to address complex challenges in hydrological modelling and drought propagation assessment, intending to enhance the physical interpretability of black-box models to support more transparent and reliable predictions. The scope of this study focuses on analysing drought dynamics at a 1 km spatial resolution, using soil moisture variability across the Alpine region. It seeks to identify and quantify the geophysical characteristics of areas most sensitive to drought events across both spatial and temporal scales, encompassing historical trends and short-term forecasting. To achieve these objectives, the thesis contributes by addressing key challenges related to multisource data heterogeneity, non-stationarity in time series, and the computational demands of large-scale prediction. It is organised into three main research directions. The first key contribution focuses on enhancing the integrity and usability of climate data to support more reliable forecasting. This is accomplished through a machine learning-based time-lag framework that incorporates a detrending transformation to improve data stationarity and predictive accuracy. To address the issue of missing data, often introduced by time-lag effects, a novel hybrid imputation strategy was developed. This method effectively integrates multiresolution climate datasets, enabling the reconstruction of complete daily records even when more than 50% of the data is missing. It consistently outperforms benchmarked imputation techniques such as GAN and MICE. The overall model framework generates high-resolution gridded climate data that retains seasonal consistency and captures finer spatial variability, outperforming the conventional downscaled climate projection, such as CMIP6. The second key contribution enhances the robustness and physical interpretability of data driven hydrological models by employing a twin-model approach that integrates the physically based model (Wflow) with deep learning. To address spatial heterogeneity, particularly in high-elevation areas with steep slopes and shadow effects, a hybrid Surrogate Deep Learning (SDL) framework is introduced, integrating a Convolutional Variational Autoencoder (CVAE) with an LSTM model to generate synthetic data, improving calibration accuracy and spatial consistency. To reduce computational cost, a parameter calibration transformation is proposed, aligning geophysical inputs with hydrological responses and enabling scalable multi-variable predictions. This framework outperforms baseline models by over 20%, achieving better results in fewer training epochs, especially for actual evapotranspiration and soil moisture. To manage seasonal variability, a multitask SDL approach is incorporated, which includes a seasonality detection method that improves model stability and spatiotemporal accuracy and reduces biases under varying climate conditions. A further focus of the thesis introduces a transfer learning strategy based on a novel regionalisation technique to support accurate large-scale prediction. This method identifies a representative subregion using convex polygon-based centroids for efficient small-scale SDL hyperparameter tuning. The learned parameters are then transferred to larger domains, improving scalability and generalisation, and reducing computational costs. Compared to benchmark methods such as catchment similarity and upscaling, this approach reduces bias in daily actual evapotranspiration and daily soil moisture predictions by 22% and 17%, respectively. The final contribution of this work aims to enhance the explainability of drought dynamics in the Alpine region by analysing geophysical characteristics of areas prone to drought anomalies using high-resolution Standardised Soil Moisture data at 1 km of spatial resolution. To address the computational challenges of generating such fine-scale indices, a conditional generative adversarial network (cGAN) is developed. This model captures nonlinear behaviours, sparse observations, and calibration complexities while preserving statistical integrity and demonstrating improved performance over spatiotemporal GANs and 3D convolutional models in complex terrains. In addition, a mapping approach for dynamic drought events is formulated using latent feature-based reinforcement learning, which improves interpretability and yields physically consistent insights. The integrated framework enables accurate detection and characterisation of drought anomalies, reveals a progressive intensification of drought conditions in recent years, and identifies key landscape features, such as elevation and soil thickness, that regulate vulnerability. These findings support the development of spatially targeted and geophysically informed adaptation strategies.