Urbanization and climate change have posed challenges for managing heatwave events, particularly in cities where the Urban Heat Island effect intensifies extreme temperatures. This study aims to develop a framework for estimating urban air temperature at a local, continuous scale using remote sensing, crowdsourced weather data, and machine learning. The key contributions of this paper to the field include investigating the model’s spatial and temporal transferability. Methods were tested across Stockholm, Sweden as well as Budapest, Hungary.

Three machine learning algorithms were evaluated, trained using a variety of feature subsets to determine the most accurate modelling approach. Results for Stockholm found that a Random Forest model with 50 features, primarily derived from optical Landsat imagery like land surface temperature,urban-, vegetation-, and water-based band indices, performed best with RMSE of 1.62°C. Aggregating predictions from individual dates revealed higher temperatures in urban and industrial areas. Fitting the model on imagery beyond the trained time period maintained consistent performance, indicating its potential for future prediction without retraining. The negligible impact of adding LiDAR data to the model shows promise for its ability to be applied to different cities, demonstrating generalizability where closed source data is unnecessary. 

When applying the same method to Budapest, the model showed slightly lower accuracy with RMSE of2.44°C, attributed to fewer available weather stations and less frequent satellite coverage. However, at a degree of accuracy still comparable to existing research, the study underscores that the methodology, despite being computationally intensive, provides worthwhile and robust across different spatial and temporal contexts. 

Key findings emphasize the value of using open- and crowdsourced data for air temperature modelling and highlight the model's resilience and transferability. Future work should focus on integrating advanced land surface temperature retrieval algorithms and maximizing available training data to further enhance model accuracy and urban resilience against climate-induced heatwaves.The study's methodological approach, including the use of variable spatial aggregation scales andfeature selection, enhances urban resilience against climate-induced heatwaves.