Journal of the Korean Geographical Society ISSN:1225-6633(Print) 대한지리학회지

Land cover map best reflects land surface phenomena and change so it is used in various research. However, the land cover maps provided by the Ministry of Environment have limitation in terms of accuracy and temporal resolution. This study proposes the possibility of deep learning based alternative land cover classification method for shortening the renewal cycle of land cover maps through automation and development of a land cover classification that is suitable for Korea. Land cover classification is conducted using Convolutional neural network, which is the deep learning architecture specialized in image analysis. In order to confirm the heterogeneity of the land cover according to region even though the same land cover type, land cover classification of Korea was experimented using the model trained by EuroSAT land cover data composed of Europe region’s land cover data. Using Sentinel-2 satellite images the training and test data are constructed for land cover classification of Korea. The convolutional neural network model trained by the data constructed in this study showed better performance with high accuracy(98.28%) than the model trained by EuroSAT data. In addition, The learning speed and accuracy(99.66%) of the convolutional neural network model was improved by using ImageNet pre-trained parameters and it showed the possibility of using ImageNet parameters in land cover classification of Korea. We hope that this study will prompt the research on land cover classification using convolutional neural network which has not been illuminated in Korea.