Abstract:
In the recent times satellite images have become one of the most efficient and effective tool to observe the earth’s geographies. These images can be used in many application like early warnings of natural disasters etc. In order to do so we need to have effective methodologies that can use the information presented in these images. One such application of satellite images is the semantic segmentation. Semantic segmentation of satellite images can be very tricky due to the nature of these images. To tackle these challenges we proposed a methodology in which we used U-Net enhanced architecture with VGG 16 as a backbone network. To improve the segmentation performance of our proposed model we have modified the structure of U-Net with modified skip connections and the network was trained on a very large dataset to cover as much details as needed. Our proposed methodology performed very well as compared to some sate of the art methodologies
