Semantic Segmentation of Urban Area Satellite Imagery Using DensePlusU-Net

The primary and most significant region for human activity is an urban area with a high population. The most distinctive characteristics of urban environments are their dense area stock and high-intensity land use as compared to rural and other natural settings. For urban areas, imbalance is a fairly prevalent issue in the task class for remote sensing image segmentation, where large object classes predominate over small object classes. Since small object classes are typically suppressed, methods centered on improving overall accuracy are frequently insufficient. A model is designed with an up-sampling blocks for gathering the context information and a down-sampling blocks for recovering the information in order to alleviate the problem of class imbalance. In this study, we suggested an end-to-end deep convolutional network to carry out semantic segmentation of satellite imagery using aerial images of Dubai collected by MBRSC satellites dataset. DensePlusU-Net is one of thedynamicvariationof U-net network architectureand it is based on Dense Network. The experiment is based on a dataset which consists of aerial imagery of Dubai obtained by MBRSC satellites dataset and annotated with pixel-wise semantic segmentation in 6 classes; building, land, road, vegetation, water and unlabeled image data. After the experiment, it was found that, DensePlus U - Net's overall accuracy was higher than U-Net model. DensePlusU-Net achieved the overall accuracy of 86.11 % and dice coefficient score of 81 %.