Explainable hybrid deep learning framework for enhancing multi-step solar ultraviolet-B radiation predictions

Acute exposure effects of short-wavelength solar ultraviolet-B (UV-B) radiation can trigger skin-based diseases and eye health ailments in humans and animals, as well as disrupt photosynthetic or hormonal systems in plants. Within the UV wavebands, high levels of UV-B exposure are particularly severe and the leading cause of skin cancers. Therefore, accurate and explainable short-term UV-B forecasts are essential for effectively providing sun exposure information to the public and UV experts. To address this pressing issue, we developed an explainable hybrid TabNet framework optimized with the Optuna algorithm. The model was trained using predictors derived from satellite products and sky images for the experimental site in Toowoomba, Queensland, Australia. For model training, 3,863 data points were utilized from July 1, 2002 to February 29, 2004. The model development phase entailed dimensionality reduction using recursive feature elimination with cross-validation (RFECV) and principal component analysis (PCA) methods. The proposed model outperformed all competing counterparts, achieving comparatively high correlation coefficients of 0.908, 0.880, 0.868, and 0.868 for hourly, 2-hourly, 3-hourly, and 4-hourly forecast horizons, respectively. Explainable artificial intelligence (xAI) results, based on Local Interpretable Model-Agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP), indicate that the antecedent lagged memory of UV-B radiation and the solar zenith angle contribute significantly to UV-B predictions. Ozone effects and cloud cover conditions are also influential features in this respect. The superior capabilities of the newly designed hybrid explainable TabNet model affirm its potential for UV-B monitoring and mitigating the harmful sun exposure risks for the public and terrestrial life.