Deep learning classification and change detection analysis of land cover using aerial imagery

This study analyzed land cover classifications and change detection anaylsis in the Namhan and Bukhan River Basins using aerial orthophotos and level-3 land use land cover maps in 2010 and 2012 as input data to SegNet, which is a deep learning-based semantic segmentation technique. The classified land cover types consist of urban, crops, forest, and water and the training dataset were constructed from the region except Namyangju city, which is located to the west of the study area. Twenty validation datasets were constructed based on five dominant areas for each land cover type in Namyangju city, the test dataset were composed of 49 datasets from Namyangju city. To optimize the classification performance of SegNet, hyperparameters were selected based on the validation accuracy according to parameter changes. The overall accuracy using validation dataset was very high, at 91.54%. In order to land cover change detection analysis, test datasets constructed in 2010 and 2012, the overall accuracies were 86.70% and 77.44%, respectively. The overall accuracy was 84.03% in case of two-classification scheme (change and no-change). Thus, SegNet showed excellent performance in terms of land cover classification and change detection, and expected to show highly accurate results for studies of sustainable land monitoring.