Semantic Segmentation of Aerial Images Using U-net for Archaeological Prospection

Archaeologists occasionally capture a few aerial photos during planned surveys. Aerial archaeology is utilized for a variety of archaeological tasks, including site recording, site location, excavation planning, site mapping, site monitoring, and more. The process of monitoring and extraction of each class of objects in aerial images is difficult. To simplify the process, semantic segmentation is used in this work. Semantic Segmentation is a technique that enables us to differentiate different objects in an image. It is able to differentiate the classes in the aerial images like buildings, land, water, vegetation, road, and unlabeled. In the process of implementation, certain stages are involved such as pre-processing, building a U-net model, training and evaluation. Pre-processing includes cropping and extraction of patches of images and get all the images into equal sizes. The U-net model is built to carry out the semantic segmentation of aerial images. The U-net has contraction and expansion paths, in between those 2 paths skip connections are present. The experimental results are carried out on a dataset of Dubai taken by the MBRSC(Mohammed Bin Rashid space center) satellite. The Mean Iou corresponding to all classes is 0.54122. Finally, Iou, Precision, Recall, and f1-scores are calculated with respect to the classes.