Opponent modeling is a significant method in imperfect information games. And intention recognition is regarded as the important but difficult in opponent modeling. This paper focuses on the task of tactical intention recognition in computational wargame. We propose an approach to recognize opponents' intention which models the intention as long-term trajectories. The approach consists of situation encoding model and position prediction model. The first model uses attention mechanism to attach the statistic map data with dynamic feature and adopt CNN to learn the representation of battlefield situation. The position prediction model then predicts the long-term trajectories of opponents, based on well-represented situation vectors. Experiment indicates that our approach is proven to be effective on the task of tactical intention recognition in wargame. Meanwhile, a high-quality replay data set for analyzing the actions' characteristics is also provided in this paper.