Solar ultraviolet radiation predictions under cloud cover effects with artificial intelligence approaches

Solar ultraviolet radiation predictions under cloud cover effects with artificial intelligence approaches

Solar ultraviolet (UV) radiation causes deleterious health effects on the skin and eyes. Skin-based malignant keratinocyte cancers, cataracts, and pterygium can be triggered by prolonged UV radiation exposure. In plants, UV radiation can damage photosynthetic processes. To address these pressing issues, this doctoral thesis developed computationally efficient artificial intelligence (AI) predictive frameworks to deliver early warnings of sun exposure times through short-term forecasting of solar UV radiation in terms of ultraviolet index (UVI) and UV radiation in category A (UV-A), with model explanations. The first objective is to develop an AI model which integrates cloud chromatic properties from sky images to forecast short-term UVI at multiple time-step horizons. The UVI forecast is generated for the four seasons of autumn, summer, spring and winter for the Toowoomba, Queensland study site in Australia that receives a high UV exposure. For the second objective, we design an AI model incorporating cloud statistical properties extracted from sky images to forecast 20-minute ahead UV-A in Toowoomba. The proposed model is calibrated through uncertainty quantification to improve predictive performance, model reliability, and generate interpretable results. In the third objective, explainable AI (xAI) is applied to forecast hourly UVI using satellite-derived predictor variables and to provide model-agnostic explanations for four Australian hotspots affected by solar UV radiation. The model is explained locally with local interpretable model-agnostic explanations (LIME) and globally with Shapley additive explanations (SHAP) and permutation feature importance (PFI). The objective model trained with satellite-derived cloud cover conditions, aerosol effects, precipitation and ozone effects extends its applicability to other geographical locations, especially in remote regions. In order to help mitigate skin and eye health risks under stochastic cloud cover conditions, the outcomes of this PhD research are expected to serve as a promising decision-support tool for delivering more accurate sun-protection times to the public. The findings can also facilitate further studies into the harmful effects of UV radiation on terrestrial plants and animals. The xAI-generated model explanations provide information on the predictive contributions of each satellite-derived variable to improve model predictability, stability, and trustworthiness for end-users. This overcomes the complex problem of UV radiation intermittency. The overall research outcomes of this project can provide a real-time UV predictive framework for the health sector, industry, government, and major stakeholders.

File reproduced in accordance with the copyright policy of the publisher/author.