Unsupervised Concept Discovery for Deep Weather Forecast Models with High-Resolution Radar Data

The global climate crisis is creating increasingly complex rainfall patterns, leading to a rising demand for data-driven artificial intelligence (AI) in short-term weather forecasting. However, the black-box nature of AI models act as a critical obstacle against their integration into existing forecasting operations. This study addresses this issue by implementing an explainable AI framework that translates model behavior into human-understandable information. Our proposed framework integrates example-based explanations, which provide interpretations in the form of user-familiar materials such as radar data, and unsupervised concept vector analysis, which identifies semantic concepts captured by the internal vector space of AI models, to interpret a forecasting model's behavior in terms of human-understandable weather concepts. We develop a multi-label self-supervised deep clustering algorithm to derive perceptually meaningful representations from an insufficient embedding space. Our method improves clustering performance over baseline methods, achieving an increase of 0.5358 in terms of silhouette coefficients. We assess the interpretability of the extracted concepts by performing a survey with five forecasters regarding the homogeneity of selected rainfall patterns. The results indicate comparable accuracies between human label-based (80%) and model-based (92%) examples. Furthermore, the proposed method can effectively distinguish between polar low and typhoon cases, successfully capturing the nonlinear weather patterns represented by data-driven models. Our explanation framework may be extended to explore the internal decision behaviors of state-of-the-art multivariable models by extracting nonlinear rainfall development and dissipation mechanisms in a human-interpretable manner.