Monitoring oasis desertification in southern Morocco: a comparative analysis of machine learning algorithms

Desertification is a serious environmental challenge that takes place in arid and semi-arid regions, posing serious threats to oasis ecosystem and the socio-economic condition of the local population. This work uses remote sensing and machine learning to assess the severity of desertification in the Tissint oasis in southern Morocco. The analysis utilized nine spectral indices obtained from Sentinel-2 imagery, reflecting the condition of vegetation, exposure of soil, albedo, moisture stress and sand encroachment. Using the interpretation of high-resolution imagery and field observations, 170 training and validation samples defined as small polygons were used to define reference data. To assess the desertification severity maps, three machine learning algorithms were executed, namely random forest (RF), support vector machine (SVM), and decision tree (DT). According to SVM model, the predominant class with area of 48.24 % is desertified land as compared to RF (37.16%) and DT (26.82%). Areas affected by sand encroachment are particularly visible in DT outputs, while stable vegetation is confined to the oasis core. According to model validation, the SVM outperformed other classifiers with accuracy = 91.76% and F1 = 93.22%. The robust and balanced performance of RF was similar to that of SVM, while DT had a lower capacity for generalization. The findings indicate that the strength of combined Sentinel-2 spectral indices and machine learning can effectively assess desertification severity. Furthermore, it can provide spatial insights for monitoring and mitigation processes.