ES-05. Fully Convolutional Neural Network for Impervious Surface Segmentation in Mixed Urban Environment

Abstract
The need to create appropriate, high-resolution data products such as impervious cover information has increased urgency as cities face rapid growth as well as climate change and other environmental challenges. This work explores the use of fully convolutional neural networks (FCNN), specifically UNet, in mapping impervious surfaces at the pixel level from WorldView-2 in a mixed urban-residential environment. We have investigated 3-band, 4-band, and 8-band multispectral inputs to the FCNN. Resulting maps are promising in both qualitative and quantitative assessment when compared to automated land-use-land-cover (LULC) products. Accuracies were assessed by F1 and average precision (AP) scores, as well as receiver-operator-characteristic curves, where area under the curve (AUC) was used as an additional accuracy metric. The 4-band model shows highest average test set accuracies with higher AP and AUC than the automated LULC products, indicating the utility of the blue-green-red-infrared channels for the FCNN. Increased performance was seen in residential areas with lower performance seen in more densely developed areas.