Rural facilities such as factories and barns near residential areas cause pollution and civil complaints. They are managed through legal registration, but there are still conflicts and inconsistencies. In Korea, the Rural Spatial Restructuring Act will come into effect in March this year. It will promote long-term rural area construction plans and support new public businesses. In this situation, it is essential to have a means to accurately and efficiently identify the status and changes in rural facilities. In this study, we constructed a dataset containing the Korea Multi-purpose Satellite (KOMPSAT) optical images and associated rural facility masks for four cities in South Korea in 2019 and 2020. A deep learning-based segmentation method was then applied to the dataset. Satellite images only show roofs, making it inherently difficult to completely distinguish between building types. Non-target buildings often look almost identical to the target. The higher the complexity of the deep neural network architecture, the more likely these inconsistencies are to cause overfitting problems. For better universal performance, we constructed redundant models from different combinations of data. Redundant models produce different inference results for the same validation sample. Averaging this gives more reliable results. Finally, we also performed a performance comparison between the original model and the new model optimized within the trust region guaranteed by the redundant models.



KeywordsSatellite optical imagery, KOMPSAT, Deep learning, Rural area, Overfitting, Semantic segmentation