BUILDING EXTRACTION IN VHR REMOTE SENSING IMAGERY THROUGH DEEP LEARNING

Abstract

Land-use changes generally are becoming through to urban classes from natural land cover classes. Remote sensing is a standard technology for such monitoring systems. On the other hand, automatic or semi-automatic applications are becoming broader and broader in most fields with development artificial intelligence over time. Building extraction from aerial photos and satellite imagery through deep learning is a new era for observing planned and unplanned urbanization, land-use changes. Besides traditional image-process methods such as supervised and unsupervised classification, deep learning presents a robust and needs less operator process. In this experiment, a commonly shared dataset is used and the sensors are QuickBird, Gaofen-2, WorldView2 with a resolution of 0.6, 0.8, 0.5 meters. Bands are Red, Green, Blue, and Near-Infrared. The dataset was split into train, test, validation, and test parts. Deep learning approaches were conducted to the dataset of different Deeplabv3+ CNN models based on Xception, Resnet-50, ResNet-18, MobileNetv2 networks. The optimized parameters were determined by selecting variations of the training options (different batch sizes, epoch numbers, backbone types, etc.). In the training phase, validation info as training and loss functions were considered. After the training process, the test phase was conducted, and evolution metrics were calculated as mean F1 score, precision, recall, IoU, and global accuracy. According to the results, the models based on ResNet-18 network were yielded better among others. Results were shown as tables and evaluated quantitatively also with time performances of the models.