Semantic Segmentation of Landsat Satellite Imagery

Currently, practitioners of remote-sensing and geographic information system use mathematical and statistical-based methods in making land cover segmentation. However, this method requires a lengthy process and staff expertise. Urban planners need applications that are not only fast in performing spatial analysis but also do not require special skills. This study proposes a land cover classification application with one of the deep learning methods for semantic segmentation, namely DeepLabV3+. This method will be compared with Iterative Self-Organizing Clustering (ISOCLUST) and Object Based Image Analysis (OBIA) in Karawang, Indonesia, as the case study. The results showed that the DeepLabV3+ accuracy was 95% which is higher than OBIA (80%). Although ISOCLUST more accurate that is usually used as ground truth dataset, this method takes a lot of time as it is semi-automatic compared to DeepLabV3+ which only takes about one minute.