A forest fire has long been a severe threat to the forest resources and human life. The threat could effectively be mitigated by timely and accurate detection. In this paper, we propose a novel accurate forest fire detection system using Wireless Sensor Networks (WSNs). In the proposed system, the detection accuracy is increased by applying the multi-criteria detection that an alarm decision depends on multiple attributes of a forest fire. The multi-criteria detection is implemented by the artificial neural network which fuses sensing data corresponding to multiple attributes of a forest fire into an alarm decision. Due to the utilization of the artificial neural network, the proposed system enjoys low overhead and the self-learning capability. Furthermore, we have developed a prototype consisting TelosB sensor nodes and carried out extensive experiments to study the performance of the proposed system. We have also developed a solar battery in order to persistently power the unattended sensor node deployed in the forest.