GPU-Accelerated Deep Learning Framework for Urban Heatwave Prediction in Sarajevo
A study available on arXiv (2605.16435) introduces a deep learning framework utilizing GPUs to forecast urban thermal conditions and assess heat risk for the following day, specifically tested in Sarajevo. The research utilized MODIS land surface temperature and Open-Meteo forecast data to analyze various convolutional and spatiotemporal models. Among these, the ConvLSTM model, which employed a mixed loss function, yielded the highest performance metrics (MAE = 0.2293, RMSE = 0.3089, R2 = 0.8877). Findings suggest that extending the temporal series and incorporating more meteorological variables could enhance prediction accuracy. Additionally, the GPU implementation with mixed precision significantly decreased execution time and allows for the integration of hazard information with exposure and vulnerability data for comprehensive risk assessment.
Key facts
- arXiv paper 2605.16435 presents GPU-accelerated deep learning for heatwave prediction
- Study conducted in Sarajevo using MODIS land surface temperature and Open-Meteo forecast data
- ConvLSTM with mixed loss function achieved best results: MAE=0.2293, RMSE=0.3089, R2=0.8877
- Longer temporal series and additional meteorological variables improve predictions
- GPU implementation with mixed precision reduced execution time
- Framework combines hazard information with exposure and vulnerability for heat risk assessment
Entities
Institutions
- arXiv
Locations
- Sarajevo