Detecting Ancient Mayan Construction Activity by U-Net

This paper deals with the issue of semantic segmentation of geographic data. The work includes applying the acquired knowledge to accurate data, specifically the design and verification of the correct procedure for processing LiDAR data from aerial surveillance. The research objective is to design relevant normalization algorithms for the digital elevation model and its modification to a suitable input to other processes. The main goal of this paper is to use a convolutional neural network U-net for the segmentation of objects of interest on selected data. The method of reverse reconstruction into the original space representation of the landscape area is compiled for partial outputs from the semantic-segmentation model. The research also demonstrates the choice of the correct evaluation metrics to verify the functionality and accuracy of our previous experiments. The prerequisite for the results was the implementation of adequate methods and the right choice of error functions that would successfully semantically segment a digital elevation model. Indicators of the correctness of verified implemented procedures were implemented on the selected metrics.