The Slovenian landscape is wide, diverse and above all offers cultural as well as agricultural functionalities to the population. Terraced landscapes contain an abundance of cultural as well as functional importance and must be identified and preserved. Since Slovenia does not yet have an adequate criterion for the identification of these terraces, we present in a way of classifying them by means of semantic segmentation with deep convolution neural network. For detection, we used the publicly available digital elevation model (LIDAR), where a simple detection model with U-Net architecture was trained. The reference data have been annotated by experts, but the annotations are inaccurate and flawed. The model managed to learned to locate terraced landscapes, which offered insights about terraced landscapes and reference data that will benefit Slovenian geographers. The results also show that, despite the noise data, detection of such landscapes is possible and represents a stepping stone for further research towards improving their detection.
|
