Improving explainable AI in attributing hydrological responses to climate variabilities in snow-dominated watersheds

Explaining the decision-making of machine learning (ML) models, known as interpretation, connects data-driven results to real-world hydrological processes, representing the next major challenge in ML applications for attribution, beyond accurate simulation. To improve ML interpretability in watershed-scale hydrological attribution, this study develops an eXplainable Artificial Intelligence (XAI) framework that incorporates a novel interpretation algorithm, Lagrange Multipliers-Support Vectors (L-SV), within a feature-based, multi-criteria constraint ML framework termed Climate Feature-Bootstrapped Support Vector Regression (CF-BootSVR). SVR simulations have been conducted in two snow-dominated watersheds, showing satisfactory simulation accuracy (average R² and NSE ≥ 0.88). The aggregated features enhance model interpretability with physically meaningful inputs and reduce computational costs by up to 30 times. The multi-criteria-layer design improves robustness and generalizability (declines in R² and NSE ≤ 0.11) while reducing uncertainties compared to single-run models. L-SV ranks feature importance similarly to model-agnostic algorithms, Permutation Feature Importance (Perm) and SHapley Additive exPlanations (SHAP), particularly in identifying the most sensitive features. L-SV also provides additional directions for feature contribution and enhances computational efficiency, being over 2513 and 2023 times faster than SHAP in the watersheds of Greata and 240, respectively. Furthermore, from a physical-based perspective, the XAI-derived attributions align with general hydrological expertise. Consequently, we conclude that CF-BootSVR offers an efficient approach to enhance predictive capabilities and deepen our understanding of climate-runoff relationships. Beyond hydrology, this CF-BootSVR framework establishes a generalizable paradigm for addressing issues related to climate seasonality. Moreover, L-SV demonstrates significant potential for broader applications in interpreting SVR models across diverse research domains.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.