Abstract:

In remote sensing applications, segmentation of input satellite images according to semantic information and estimating the semantic category of each pixel from a given set of tags are of great importance for the automatic tracking task. It is important in situations such as building detection from high resolution satellite images, city planning, environmental preparation, disaster management. Buildings in metropolitan areas are crowded and messy, so high-resolution images from satellites need to be automated to detect buildings. Segmentation of remote sensing images with deep learning technology has been a widely considered area of research. The Fully Convolutional Network (FCN) model, a popular segmentation model, is used for building detection based on pixel-level satellite images. In the U-Net model developed for biomedical image segmentation and modified in our study, its performances during training, accuracy and testing were compared by using customized loss functions such as Dice Coefficient and Jaccard Index measurements. Dice Coefficient loss score was obtained 84% and Jaccard Index lost score was obtained 70%. In addition, the Dice Coefficient loss score increased from 84% to 87% by using the Batch Normalization (BN) method instead of the Dropout method in the model.

Keywords: