Semantic Segmentation of Satellite Images using 2 various U-Net Architectures: A Comparison Study

Image segmentation has been a challenging issue in computer vision for years. In contrast to image classification and object detection, semantic segmentation is considered the top tier of the image analysis approach, which gives detailed details of the scene for a given input image. Analysis of aerial images without human intervention has developed a keen interest in research because of its vital importance in various domains. Different applications like disaster response and urban planning depend mostly on semantic segmentation of aerial imagery for their analysis. In a wide range of image processing tasks, convolutional neural networks (CNNs) have manifested their tremendous performance, and the transformation of the computer vision domain is achieved by deep learning. Amongst multiple varieties of CNN, U-Net has proved its efficiency in segmenting aerial images and the medical domain. Nonetheless, U-Net can’t extract potential spatial features from aerial images because of insufficient layers and may output inaccurate boundaries, particularly for objects with compound structures. To circumvent these deficiencies, different varieties of U-Net are experimented with for aerial image segmentation using U-Net, Attention U-Net, Attention Res U-Net, and Recurrent Residual U-Net. We evaluated all these models on a publicly available dataset named semantic segmentation of aerial imagery. Extensive experimental results conclude that Attention Res U-Net and Recurrent Residual U-Netperform better than other U-Net architectures.