Unlocking the potential of artificial intelligence in hydrology : Deep learning framework for snow data assimilation in S3M

Time and computational constraints often limit the retrieval of reliable snow model estimates, particularly in large-domain operational contexts. artificial intelligence–based modelling approaches can improve the representation of snow storage and dynamics, while reducing uncertainties in snow-dominated regions and enhancing assessments of water avail- ability, ecosystem functioning, and climate feedbacks, with significant societal and envi- ronmental implications. Despite this potential, the effective integration of artificial intel- ligence–based methods into operational snow hydrological modelling frameworks remains an open research challenge. In this context, the main hypothesis of this doctoral research is that artificial intelli- gence–based techniques can be effectively embedded within snow hydrological models to enhance predictive accuracy while reducing computational effort and execution time. This integration leverages the potential of artificial intelligence to address key challenges in op- erational snow modelling, namely: (i) quality control and quality assurance of in situ snow observations, and (ii) the assimilation of snow-related variables. The proposed Random forest–based quality-control algorithm provides a fully auto- matic, computationally efficient, and scalable approach for quality assurance of snow depth observations. By incorporating expert domain knowledge, it offers an effective alternative to manual screening and substantially reduces the time and resources required for data val- idation. It achieved F1 scores above 90% for snow versus grass or bare-ground detection, even outside the training domain. Despite less consistent error classification performance, the algorithm’s limited sensitivity to variations in snow-season climatology highlights its applicability across heterogeneous environmental conditions. The Ensemble Kalman filter emulator based on a Long Short-Term Memory network achieved snow depth and snow water equivalent estimates comparable to the traditional ensemble approach while reducing computational time by up to 70% by using ensemble simulations only during training. The framework demonstrated strong spatial transferabil- ity, with only a 20% decrease in performance outside the training domain, and highlights the potential for a fast, scalable, and spatially distributed deep data assimilation frame- work, with future developments envisaging the propagation of pointwise corrections across the domain using Gaussian Process interpolation. In my view, while the computational efficiency of artificial intelligence-based emulators offers clear advantages for cryosphere and hydrosphere modelling, it should not come at the expense of process transparency and interpretability. I believe that ensuring explainability is essential in operational and decision-support contexts, where trust, accountability, and physical consistency are critical, and that scientists have a key role in linking efficient artificial intelligence outputs with interpretable, physically meaningful results.