Segmentation of Impervious Areas from High Altitude Aerial Images using a Regularized UNet Framework: * Note: First two authors contributed equally and are sorted by last name

In this paper we introduce a novel dataset of high altitude aerial images with manual annotations for impervious surfaces from a predominantly urban area. The dataset also contains a small amount of hand annotations for man-made pervious surfaces, such as unpaved road and driveways. Each aerial image also has unlabeled regions that are neither impervious nor man-made pervious. A pervious surface is defined as one that allows water to pass through whereas an impervious one does not allow water to percolate. After introducing the dataset, we study the problem of automatically categorizing the regions in an aerial image into their respective categories. To this end we explore three directions: 1) predicting impervious, pervious and unlabeled (background), 2) predicting impervious and the combined class of unlabeled and man-made pervious and 3) predicting impervious and background excluding man-made pervious regions. For each of these problems, we used a custom built UNet architecture with dropout layers. We report the results from the same in the form of classification masks. The results were verified by manually aligning the predicted mask with the ground truth mask for impervious, man-made pervious, and background regions. Our experiments and results establish the efficacy of using deep learning techniques for the problem of automatically identifying impervious surfaces from high-resolution aerial imagery, especially in complex urban areas.