Titolo della tesi: Designing an Early-Warning System to Forecast Extreme Climate Conditions Using Machine-Learning and Deep-Learning Methods
Extreme weather events exacerbated by climate change present unprecedented challenges to forecasting and disaster management. This thesis addresses the limitations of traditional numerical weather prediction and global climate models by developing an integrated deep learning framework for high-resolution climate forecasting and early warning systems. Central to this work is the novel E-TEPS model, a downscaling approach inspired by SRGAN architectures that incorporates auxiliary inputs, notably digital elevation maps, to enhance the spatial fidelity of temperature and precipitation predictions. By coupling E-TEPS with the global forecasting capabilities of FourCastNet, the proposed system efficiently transforms coarse-resolution outputs into detailed, high-resolution forecasts within seconds—delivering critical insights into extreme rainfall and temperature anomalies that serve as precursors for floods and heatwaves.
The integrated framework was rigorously evaluated using extensive datasets, including ERA5 and the dynamically downscaled CMCC data, and validated against real-world case studies of extreme rainfall events, such as those observed in the Emilia-Romagna and Marche regions of Italy that have historically been associated with flooding. Quantitative assessments based on metrics such as RMSE, MAE, Pearson correlation, and threshold-based event detection scores (POD, FAR, and CSI) demonstrate that the inclusion of elevation data significantly improves forecast accuracy, particularly in regions with complex topography. Furthermore, the integrated system outperforms alternative approaches, such as U-Net-based models, by preserving fine-scale spatial details and enhancing the detection of extreme events.
A key enabler of this research was the extensive use of cloud infrastructure, which provided the necessary computational flexibility and efficiency to overcome hardware limitations. Platforms like Google Cloud Platform and Google Colab allowed for scalable, cost-effective processing across all stages—from data preprocessing and model training to fine-tuning and inference—thus facilitating rapid experimentation and operational deployment of the forecasting system.
In addition to these achievements, this work carefully addresses the challenge of balancing error distributions, particularly for low-intensity precipitation, and effectively mitigates inherent dataset biases through robust data harmonization and innovative model training strategies. The resulting framework demonstrates a reliable capability to capture both moderate and extreme weather events, offering a comprehensive solution that meets the rigorous demands of operational climate forecasting.
Overall, the integrated forecasting system presented in this work represents a substantial step forward in high-resolution climate prediction, offering both scientific insights and practical benefits for early warning and disaster management in an era of increasing climatic uncertainty by providing forecasts of extreme rainfall and temperature anomalies that can serve as critical indicators for potential flood and heatwave events. Moreover, the framework's adaptability and cost-effectiveness open up promising avenues for deployment in developing countries, where enhanced forecasting capabilities can drive proactive disaster management and sustainable development. Future work will focus on tailoring the system to local infrastructural and regulatory contexts, as well as integrating additional hydrological modeling components to translate extreme rainfall forecasts into actionable flood risk assessments, thereby fostering collaborative approaches among stakeholders and further improving resilience to extreme weather events.