Wildfire detection is an indispensable component of many resilient platforms, preventing environmental disasters from damaging life. Online detection, which refers to situations in which wildfire events need to be detected in real time, is an essential tool, as the consequences of wildfires might already be irreversible by the time they are detected. Recently, wildfire detection methods based on machine learning and deep learning have been proposed. However, most of these methods take raw data as the input and then aggregate it into predictive features, which requires all of the historical data or a large part of a stream and violates the timeliness requirement of online detection. Moreover, these methods scarcely discuss the explainability (e.g., how the model comes up with the final detection). The lack of explainability reduces human trust in the system and may hinder further applications of the system, especially in high-stakes situations where decisions can have significant consequences. Unlike existing works that treat the timeliness and explainability problems separately, we propose a real-time wildfire detection system that is built upon the streaming capability of complex event processing and the expressiveness of semantic annotation. The system continuously processes raw data streams, transforms them into semantic events, and learns CEP queries that are both predictive and interpretable for humans at the same time. Experiments on four real datasets and one synthetic dataset show that our approach outperforms the baselines in terms of efficiency, accuracy, explainability, and adaptivity.