Accuracy–Privacy Trade-off in Federated Learning for Joint Fire Risk Model Training Across International Wildfire Datasets

Wildfire risk prediction has become critically important as climate change intensifies fire events worldwide. Wildfire datasets collected by different countries constitute a rich source of geographically and ecologically diverse information; however, data sovereignty regulations and privacy requirements substantially restrict the centralized aggregation of such data. This study examines the accuracy–privacy trade-off provided by federated learning (FL) when training a joint fire risk model across multiple countries' wildfire datasets without centralizing the raw data. Synthetic datasets representing five countries — Turkey, the United States, Australia, Portugal, and Canada — were used to compare FedAvg, FedProx, and differentially private federated learning algorithms. Experimental results demonstrate that federated learning models incur only a 2.3–4.7% accuracy loss compared to centralized training, while data privacy is mathematically guaranteed through differential privacy mechanisms. Furthermore, communication costs can be reduced by 60% through model compression techniques. The findings suggest that federated learning offers a viable solution for international wildfire risk modeling by overcoming data-sharing barriers while maintaining an acceptable accuracy–privacy balance.