In many ubiquitous intelligent systems, high-level situations are often required to be inferred by fusing several contexts, which is referred to as situation inference. As opportunistic sensing becomes widely accepted, new challenges are brought into situation inference. In the opportunistic sensing paradigm, applications make the best use of the sensors that happen to be available in a certain location, and those sensors do not necessarily need to be pre-deployed. In this way, opportunistic sensing effectively expands the scope of ubiquitous intelligent applications, but meanwhile brings uncertainty of sensed contexts to the situation inference as well. In this paper, we propose a learning-based approach for situation inference by fusion of opportunistically available contexts. In the offline training phase, in order to reduce the computation load, it only pre-computes some reduced-feature models (RFMs) with higher utility for situation inference, rather than training all possible ones. In the online classification phase, if the input context combination matches one of the pre-computed RFMs, then the model is used to infer the situation, otherwise a less accurate but more general method, the imputation-based method is applied. We evaluate our approach using an open dataset with various degrees of incompleteness and inaccuracy introduced.