A Framework Using Federated Learning for IoT-Based Forest Fire Prediction

Forest fires are a growing threat to human commu-nities. The Canadian Wildland Fire Information System gives realtime information to fire management agencies and the public. However, machine learning use for forest fire ignition classification prediction within the platform and ones like it, is yet to be fully realized. We propose a novel framework that uses federated machine learning combined with Internet of Things technologies, for forest fire ignition classification prediction. The framework incorporates distributed IoT weather stations deployed in an area prone to forest fires. We find comparable prediction accuracy between a federated machine learning system and a central server machine learning system. Our federated system, trained on an imbalanced dataset comprising 5,008,365 non-ignition cases and 45,411 ignition instances, has shown encouraging outcomes. It attained an Accuracy of around 0.76 and a ROC-AUC of about 0.80. The performance is on par with other systems, indicating that our approach is effective in classifying forest fire ignitions with a spatial resolution markedly superior to that of centralized systems.