Semantic segmentation of historical aerial photographs using deep learning

Historical image archives provide valuable information about the regions of the world before the satellite era. These archives have been an indispensable source for obtaining information about land cover and use and observing temporal changes. This situation necessitated the semantic interpretation of historical aerial photographs with high accuracy. In the study, U-Net backbone DeepLabV3 model was adapted and used for semantic segmentation and the performance of this architecture on historical aerial photographs was evaluated as 5-class and 3-class to cover ground and above-ground objects. The data labeling process was performed manually with LabelMe software. Historical aerial photographs were divided into training, validation and test sets after being divided into 120 sub-segments with a total size of 256×256 pixels. In addition, Digital Elevation Model (DEM) produced by evaluating the historical aerial photographs in 3D was added to the architecture and the results were compared. Intersection over Union (IoU) and F1 score accuracy metrics were used in the performance evaluation phase. For 5-class data, F1 Score was calculated as 60.94% and IoU as 46.90%, for 3-class data, F1 Score was calculated as 76.19% and IoU as 66.85%, for 5-class data created by adding SYM, F1 Score was calculated as 61.48% and IoU as 47.53%, for 3-class data, F1 Score was calculated as 77.59% and IoU as 68.24%.