M-unet: Modified u-net segmentation framework with satellite imagery

Abstract

In recent years, convolution neural network (CNN) has emerged as a dominant paradigm in machine learning for image processing application. In remote sensing, image segmentation is a very challenging task, and CNN has shown its worth over traditional image segmentation methods. U-Net structure is one of the simplified architectures used for image segmentation. However, it cannot extract promising spatial information from satellite data due to insufficient numbers of layers. In this study, a modified U-Net architecture has designed and proposed. This new architecture is deep enough to extract contextual information from satellite imagery. The study has formulated the downsampling part by introducing the DenseNet architecture, encourages the use of repetitive feature map, and reinforces the information propagation throughout the network. The long-range skip connections implemented between downsampling and upsampling. The quantitative comparison has performed using intersection over union (IOU) and overall accuracy (OA). The proposed architecture has shown 73.02% and 96.02% IOU and OA, respectively.