Use of deep learning models to extract buildings from remote sensing images

In this study, buildings were automatically extracted from remote sensing images using a deep learning model. The open access Inria dataset was used for training and testing the model. While creating the model to be used in the study, the performance of different network backbones and segmentation algorithms were evaluated on the same data set and the model with the highest success rate was selected. The tested models are the model consisting of Resnext50 network backbone and FPN segmentation architecture, the model consisting of VGG16 network backbone and U-Net architecture, the model consisting of ResNet50 network backbone and LinkNet architecture, the model consisting of ResNext50 network backbone and PSPNet architecture. Among these 4 models, the ResNext50 backbone and FPN segmentation architecture, which gave good results with an IoU value of 78.14%, was chosen to be used in the study. The selected model is trained using ReLU activation function, Cross entropy and Dice loss functions for 3 learning periods and results are obtained by using Softmax classifier for the output of the network. The results obtained were compared with the results obtained in the literature using different models with the same data set.