Towards Sustainable Heritage Conservation: Transformer-Based temperature forecasting in the city of Florence

Abstract. Urban Heat Islands (UHIs) are intensifying under climate change, posing serious threats to the conservation of cultural heritage in historic cities. This study presents a deep learning framework for forecasting land surface temperature (LST) at high temporal resolution, aimed at supporting thermal risk mitigation in heritage-sensitive urban environments. Using hourly LST data from the Copernicus Land Monitoring Service (CLMS), we developed a custom autoregressive Transformer model capable of predicting 72 future hourly temperature values based on the previous 168 hours of satellite observations. The model follows an encoder-decoder architecture: the encoder processes the full input sequence with spatial and temporal embeddings, while the decoder generates the forecast step-by-step, leveraging past predictions. We applied our approach to the city of Florence, Italy, using available data from the years 2021, 2022, 2023, and 2024. The model was evaluated on the four summer months (June 2021, July 2022, August 2023 and September 2024) achieving a minimum MAE of 0.81 °C and confirming its applicability for extreme heat forecasting.