<abstract xmlns="http://www.w3.org/1999/xhtml">
<p>Cultivated terraces are phenomena that have been protected in some areas for both their cultural heritage and food production purposes. Some terraced areas are disappearing but could be revitalised. To this end, recognition techniques need to be developed and terrace registers need to be established. The goal of this study was to recognise terraces using deep learning based on Lidar DEM. Lidar data is a valuable resource in countries with overgrown terraces. The U-net model training was conducted using data from the Slovenian terraces register for southwestern Slovenia and was subsequently applied to the entire country. We then analysed the agreement between the terraces register and the terraces recognised by deep learning. The overall accuracy of the model was 85%; however, the kappa index was only 0.22. The success rate was higher in some regions. Our results achieved lower accuracy compared to studies from China, where similar techniques were used but which incorporated satellite imagery, DEM, as well as land use data. This study was the first attempt at deep learning terrace recognition based solely on high-resolution DEM, highlighting examples of false terrace recognition that may be related to natural or other artificial terrace-like features.</p>
</abstract>

Cultivated terraces are phenomena that have been protected in some areas for both their cultural heritage and food production purposes. Some terraced areas are disappearing but could be revitalised. To this end, recognition techniques need to be developed and terrace registers need to be established. The goal of this study was to recognise terraces using deep learning based on Lidar DEM. Lidar data is a valuable resource in countries with overgrown terraces. The U-net model training was conducted using data from the Slovenian terraces register for southwestern Slovenia and was subsequently applied to the entire country. We then analysed the agreement between the terraces register and the terraces recognised by deep learning. The overall accuracy of the model was 85%; however, the kappa index was only 0.22. The success rate was higher in some regions. Our results achieved lower accuracy compared to studies from China, where similar techniques were used but which incorporated satellite imagery, DEM, as well as land use data. This study was the first attempt at deep learning terrace recognition based solely on high-resolution DEM, highlighting examples of false terrace recognition that may be related to natural or other artificial terrace-like features.

Towards the deep learning recognition of cultivated terraces based on Lidar data: The case of Slovenia

Faculty of Computer and Information Science, University of Ljubljana

Moravian Geographical Reports

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Cultivated terraces are phenomena that have been protected in some areas for both their cultural heritage and food production purposes. Some terraced areas are...

Towards the deep learning recognition of cultivated terraces based on Lidar data: The case of Slovenia

Publié en ligne: 03 avr. 2024

Pages: 66 - 78

Reçu: 29 août 2023

Accepté: 10 févr. 2024

DOI: https://doi.org/10.2478/mgr-2024-0006

Mots clés
cultivated terraces, deep learning, landscapes, digital elevation model, feature detection, Slovenia

© 2024 Rok Ciglič et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Towards the deep learning recognition of cultivated terraces based on Lidar data: The case of Slovenia

Publié en ligne: 03 avr. 2024

Pages: 66 - 78

Reçu: 29 août 2023

Accepté: 10 févr. 2024

DOI: https://doi.org/10.2478/mgr-2024-0006

Mots cléscultivated terraces, deep learning, landscapes, digital elevation model, feature detection, Slovenia

© 2024 Rok Ciglič et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Mots clés
cultivated terraces, deep learning, landscapes, digital elevation model, feature detection, Slovenia