Feature extraction is an essential task in the classification of high-resolution remote sensing images, with the primary technique being the object-oriented classification method. Current research describes object-oriented classification methods by using remote sensing data, wherein how to reduce the redundant feature information to achieve good classification results is the most challenging problem. The high-resolution remote sensing image is characteristic of a large amount of data and high feature dimensions, which also exist particularly in the forestry remote sensing. Feature information redundancy can reduce the extraction accuracy and make the classification results worse. To address this problem, in this paper we propose a framework that uses the rough set theory and the membership function to establish the classification rule set. In our approach, we first select an optimal segmentation scale to segment the remote sensing image with multi-scale and apply the rough set theory to reduce the feature dimensions of objects. We then use the selected features to establish classification rule set and classify image objects. This paper also presents a detailed study of the proposed framework for species classification with ALOS images, wherein 13 most effective feature parameters are selected from 34 feature parameters of objects, such as band ratio, brightness value, and average gray value. Our experimental results demonstrate that the proposed framework, applied to classify tree species, achieves a classification accuracy of 80.4509%, which is an improvement over both the classification accuracy of 77.2408% achieved with the traditional supervised classification and that of 75.5068% achieved with the nearest neighbor classification. The research proves that the proposed framework can effectively take advantage of tree species information in remote sensing images, and provides an auxiliary means for forest resources investigation and monitoring.