Global Urban Detection Based on High-Resolution and Multi-temporal Sentinel-l Big Data

In this work, a deep learning framework is proposed to achieve global urban detection from the high-resolution and multitemporal Sentinel-1 image on Google earth engine platform. Relative to the massive amount of SAR image collected, the completeness of training dataset hardly can be ensured when the sampling is limited. Therefore, different from other existing works, the building of training dataset is also taken into consideration of designing the deep learning framework, and some dynamic programming and transfer learning strategies are adopted to improve its classification ability in this cloud-based platform. Taking Mumbai, Beijing and Stockholm as example, experiment results on real data illustrate the feasibility of the proposed method.