Original scientific paper
https://doi.org/10.32909/kg.24.44.2
Application of Image Segmentation and Classification Techniques to Sentinel-2A Data for Monitoring Desertification: A Case Study in Southeastern Morocco
Otman TAMRI
orcid.org/0009-0000-1151-1656
; Geosciences and Natural Resources Laboratory, University of Ibn Tofaïl, Kenitra, Morocco
*
Mouhssine EL ATILLAH
orcid.org/0000-0002-3431-8143
; Computer Systems Engineering, Mathematics and Applications (ISIMA), Polydisciplinary Faculty of Taroudant, University Ibn Zohr, Taroudant, Morocco
Saïd CHAKIRI
; Geosciences and Natural Resources Laboratory, University of Ibn Tofaïl, Kenitra, Morocco
Allal LABRIKI
orcid.org/0009-0002-3328-740X
; Department of Geology, College of Geosciences and Applications, University of Sciences Ben M'Sik, Casablanca, Morocco
Mohammed Amine ZERDEB
; Geosciences and Natural Resources Laboratory, University of Ibn Tofaïl, Kenitra, Morocco
* Corresponding author.
Abstract
Monitoring land cover in arid and semi-arid zones through remote sensing is essential for understanding desertification processes and guiding sustainable land management strategies. This study assesses the performance of several unsupervised image segmentation and classification algorithms, namely K-means and ISODATA (integrated into QGIS), as well as SLIC, Mean Shift, and Felzenszwalb (implemented in Python), for land cover mapping in southeastern Morocco, a Saharan region highly vulnerable to land degradation. The analysis is based on Sentinel-2A remote sensing imagery acquired during the dry season (April), using three RGB band combinations (11/8/2, 12/11/4, and 8/4/3) to better distinguish key land cover units: sand dunes, desert crusts, sparse vegetation, wadi beds, and rocky plateaus.
The results indicate that each algorithm presents specific strengths and limitations depending on landscape complexity and segmentation parameters. K-means and ISODATA allow for rapid and easily interpretable classification but tend to confuse intermediate classes. However, advanced methods like SLIC (especially with 500 to 1 000 segments), Mean Shift (which uses adjusted bandwidths), and Felzenszwalb (at a medium scale) provide better spatial delineation and thematic superiority, especially for vegetation and linear elements, although they require more computational resources.
These results highlight the importance of choosing methods according to the analysis objectives and the available resources. The use of these remote sensing techniques is an effective way to enhance desertification monitoring systems and support national projects, particularly those orchestrated by the High Commission for Water, Forests, and the Fight against Desertification.
Keywords
remote sensing; image segmentation; classification; Sentinel-2A; desertification; unsupervised algorithms
Hrčak ID:
344540
URI
Publication date:
31.12.2025.
Visits: 407 *