https://uvadoc.uva.es/handle/10324/83797 2026-04-18T12:24:02Z 2026-04-18T12:24:02Z Gatón Herguedas, Javier Román Díez, Roberto Guzman, César González Fernández, Daniel Longarela, Bruno Toledano Olmeda, Carlos González Caton, Ramiro https://uvadoc.uva.es/handle/10324/83808 2026-03-25T20:02:03Z 2026-01-01T00:00:00Z

This paper presents a comparative study on the impact of input representation on deterministic artificial intelligence models for short-term multi-frame prediction in all-sky images. This work compares a model operating on 8-bit RGB all-sky images with a model that shares the same backbone, but operates directly on semantically segmented masks that encode cloud-related classes. Using an available sky segmentation model, predictions are evaluated in the segmentation label space using segmenter-derived masks as a proxy reference. Within this evaluation framework, the use of semantic masks as input for short-term prediction leads to improved temporal stability and higher agreement across standard segmentation metrics such as intersection over union, Dice coefficient, and categorical cross-entropy. While these results suggest potential relevance for weather and solar energy nowcasting applications, further validation against physical irradiance measurements is required.

2026-01-01T00:00:00Z