Image classification is one of the key issues of remote sensing technology and selecting training data is an essential process in supervised image classification. Dramatically increasing imagery data require more effective and automated classification techniques. The traditional process of selecting training data requires intensive manpower and, as a result, it has been costly and time-consuming. This study proposed an automatic training data extraction technique using outdated geographic information system (GIS) data and its applicability was tested. We used a high-resolution KOMPSAT-3A satellite image taken on July 7, 2018, and the land cover map in 2015 for the test of automated training data extraction based on the iterative trimming algorithm. First, the training data were extracted based on the polygon of the land cover map. Then, the probability distributions of each land cover class were estimated using kernel density estimation. The outliers were removed in the order of low probability. The bootstrap technique was used to determine the ratio of removing outliers. The ratios were different among the land cover classes. The removing ratio was 0.08 for the urbanized area, 0.16 for agriculture/land, 0.04 for forests, 0.16 for bare soil and 0.04 for water. With the refined training data, image classification was conducted. This approach allows automatic extraction of training data based on GIS data without manual digitizing. It is expected to contribute to an automatic and timely update of the urban land cover map with high-resolution imagery.



Keywords: Training Sample, Land Cover Map, Bootstrap, Image Classification, KOMPSAT-3A