Applicability of remotely sensed data in heterogeneous land use/cover mapping has been greatly be restricted, given the fragmented distribution and spectral confusion of land cover features. In addition, ALOS/AVNIR-2 data, which show great potential for land use/cover mapping considering the high-resolution but low-cost characteristics, have not received too much attention. A new hybrid methodology to address these issues is proposed. This approach combines traditional supervised classifiers and unsupervised classifiers and integrates multi-temporal spectral and structural information from ALOS/AVNIR-2 images. The multi-temporal spectral and structural information are then used as auxiliary data through a rule-based decision tree approach to generate a final product with enhanced land use classes and accuracy. A comprehensive evaluation of derived products of the northern part of Shangyu City in eastern coastal China is presented based on official land use/cover map (1:10000) as well as inter-classification consistency analyses. Overall accuracy of 86.5% and Kappa statistics of 0.84 have been achieved, which are significantly higher than those obtained from the Maximum Likelihood classifier and ISODATA classifier. The hybrid approach presented here is straightforward and flexible enough to be generalized so the approach can be applied to interpret similar fragmented land use/cover using various remotely sensed source data.