Recently, Wireless Sensor Networks (WSN) community has witnessed an application focus shift. Although, monitoring was the initial application of wireless sensor networks, in-network data processing and (near) real-time actuation capability have made wireless sensor networks suitable candidate for event detection and alarming applications as well. Unreliability and dynamic (e.g. in terms of deployment area, network resources, and topology) are normal practices in the field of WSN. Therefore, effective and trustworthy event detection techniques for the WSN require robust and intelligent methods of mining hidden patterns in the sensor data, while supporting various kinds of dynamicity. Due to the fact that events are often functions of more than one attribute, data fusion and use of more features can help increasing event detection rate and reducing false alarm rate. In addition, sensor fusion can lead to more accurate and robust event detection by eliminating outliers and erroneous readings of individual sensor nodes and combining individual event detection decisions. In this paper, we propose a two-level sensor fusion-based event detection technique for the WSN. The first level of event detection in our proposed approach is conducted locally inside the sensor nodes, while the second level is carried out in a level higher (e.g., in a cluster head or gateway) and incorporates a fusion algorithm to reach a consensus among individual detection decisions made by sensor nodes. By considering fire as an event, we evaluate our approach through several experiments and illustrate impact of sensor fusion on achieving better results.