Data Augmentation Investigating in Improving the performance of Deep Learning Model for Building Segmentation using Aerial Images

Abstract

Deep learning is a modern method of image processing and data analysis that has entered the field of urban management with promising results and high potential. The purpose of this study is to investigate data augmentation techniques in improving the results of segmentation of building using aerial images with high spatial resolution and deep learning method. For this purpose, MSB building data set and MapNet model were used. The model was trained and evaluated in three stages without data augmentation, with data augmentation of geometric transformations and with data augmentation of geometric and photometric transformations. The results of model evaluation showed that using geometric transformations as data enhancement techniques, F-1 and IoU score evaluation criteria have increased by 0.5 and 0.55%, respectively, and using data techniques Incremental geometric and photometric transformations increased by 1.41 and 1.57 percent. This increase was visually observed in the improvement of the segmentation of dense areas of the building and the discontinuity of large-scale buildings.

Keywords

• photometric transformations
• geometric transformations
• aerial images
• data augmentation
• segmentation
• deep learning