In this paper, taking Shunyi district hyperpectral imagery as an example, we probe into the keypoints and principal procedures for multi-level classification approach and apply this object-oriented method to extracting and mining crops information. Because multi-level classification approach can simplify the complicated information extraction procedure into several relatively easy sub-process, we can intentionally select different characteristic parameters and data mining methods for information extraction in each sub-process according to the feature of the crop we want to discriminate. Consequently, we can make the most of the abundant information in hyperspectral data, and improve the precision of extracted information.