Remotely sensed imagery classification based on object-oriented image analysis plays an important role in mapping land cover. The object-oriented classification method is more useful than that based on pixel classification. Texture, shape and other features can be included in the object, which is generated after the segmentation. For a large area to be classified, Landsat Thematic Mapper (TM) remotely sensed imagery can be used as the data source. Therefore, we used TM images for object-oriented classification here. After selection of parameters for segmentation, we investigated how to optimize the TM temporal resolution, thereby improving the classification accuracy. In the study area (the city of Yantai, China), pixel-based classification of vegetation can be a challenge. The use of object-oriented classification combined with ancillary data such as multi-temporal characteristics, digital elevation model, slope, and slope direction can be a better solution to this problem. This study is organized as follows. First, a segmentation algorithm, the multiresolution segmentation based on the Fractal Net Evolution Approach (FNEA), is applied to the images. The shape parameter was set to 0.1 to highlight the homogeneous pixels for imagery segmentation. The compactness parameter was set to 0.5 to equally balance the compactness and smoothness of objects. Image layer weights of band1, band2, band3, band4 and band5 were all 1. We then tested the segmentation results to evaluate whether the scale parameter was suitable for classification. Ten objects of varying scale were visually selected from each category, and we then developed statistical spectral information of each band to obtain the mean as spectral values of each category for small variances. Ten pure pixels of each corresponding category were selected in the original image, the mean of which represented the spectral values of each band. We used linear regression analysis in which y was the mean spectral value of the objects and x was mean spectral value of pure pixels. If there was a good fit of y and x, the scale was considered reasonable. We chose L20, L40, and L80 scales (the scale parameters were set to 20 m, 40 m and 80 m) for classification based on the above results. According to spectral characteristics, geographic origin, shape, and location, an interpretation signs library was established. We distinguished categories by the characteristics of various types of objects. A decision tree was created, and we then used the membership function for classification. When the membership of a segment was less than the default threshold values for all relevant land-use categories, the segment was marked as unknown. Finally, a support vector machine was used to classify these unknown category objects. We obtained 216 sample points from the Yantai study area. We assessed the accuracy of results by the methods described above. This assessment gave an overall accuracy of 82.7%. We classified the same imagery using the maximum likelihood method, together with the same ancillary data and expert knowledge. Overall accuracy of the results was 64.2%. Using the object-oriented classification method with multi-temporal images can significantly improve classification accuracy compared with traditional pixel-based classification, especially in vegetation classification to distinguish shrubs and grasses.