Semantic Segmentation with Remote Sensing Data and Reference Labels Based on Simulation Methods

Kurzfassung

Deep learning provides more opportunities for image segmentation. Meanwhile, neural network becomes a popular method for ground surface classification. Manually selected label in training data needs investment in both cost and time. However, this problem could be well solved by the program, named SimGeoI[1]. SimGeoI simulates optical image, SAR (Synthetic Aperture Radar) image and ground-surface labels from DSM (digital surface model) data. In this thesis, a batch processing part of SimGeoI and the semantic segmentation based on data set generated by SimGeoI are implemented. Through SimGeoI batch processing, the satellite dataset for semantic segmentation can be significantly expanded. Semantic segmentation based on simulation-methods generated dataset is also implemented. The case studies in Munich area, with WorldView-2 imagery and TerraSAR-X data, confirms the opportunity of semantic segmentation using dataset generated by SimGeoI.