Def-Use faults are a very important and common type of faults. The state-of-the-art detection schemes for such faults still hardly achieve both preciseness and scalability. This paper applies the idea of combining the sensitive and insensitive detection approaches and deploying the effective range of the two approaches to achieve both high detection scalability and high precision. The study results in a new scene- sensitive detection strategy based on a classification scheme on statements that contain potential faults. The key idea is to classify these statements into different categories based on how a potential fault in these statements might be triggered. It uses polynomial flow-, field- and context-sensitive summary based scene analysis to do the classification and identifies triggering scenes based on program dependence information. Different detection schemes with different amount of overheads are then applied to different categories and thus reducing the overall overhead and achieving a higher scalability. The path-sensitive detection schemes are only performed on the necessary triggering scenes. The proposed approach is implemented in a prototype system, called Minerva. Using null pointer dereference fault detection as an example and verifying the approach through applications whose total code size exceed 2.9 million lines (one application exceeds 2 million lines), the experimental results show that the average detection time of Minerva is 3× and 46× faster than the two state-of-the-art path-sensitive detection tools, Clang-sa and Saturn, respectively. The false positive rate of Minerva is 24%, which is also a third of that of Clang-sa and Saturn's. There is no false negative on the known faults. The results show that the proposed scene-sensitive fault detection approach can achieve both high scalability and high accuracy.