Abstract: As a comprehensive combination of natural and artificial structures on the surface, land coverage is an important foundation for the development of land science related research. In the background of large remote sensing data, accurate, fast and automatic land coverage extraction technology is focused in remote sensing research all the time. Based on eCognition software, this paper take object-oriented multi-scale segmentation method. Taking account of the spectral, shape and texture characteristics of land objects comprehensively in remote sensing images, established a variety of rules to extract the types of land coverage in the study area by using fuzzy function, support vector machine (SVM) and threshold method. A comparative analysis was also made to compare with FROM-GLC10 data and land use change data in the study area. The results showed that: ① The overall accuracy of land coverage classification in the study area was 97%, Kappa coefficient was 0.96, and the classification accuracy was high. ② Based on 10 m resolution image, the comprehensive use of shape, texture and spectral information had a good effect on the extraction of roads. The Kappa coefficient of road extraction was 0.84. ③ The classification results were better than the FROM-GLC10 data both in area and spatial distribution, as great consistency with the land change data in the study area. The object-oriented and rule-based classification method for feature extraction can effectively utilize a variety of remote sensing image features, and the classification accuracy is high, which has a good advantage for processing high-resolution remote sensing data.

Key words: land coverage, object orientation, multi-scale segmentation, rules, Sentinel-2A