Field accessibility may jeopardize the assessment and monitoring of land degradation and habitat restoration. This is the case for the Sharaan Nature Reserve in AlUla (Saudi Arabia), which covers a large area (1500 km²), where field assessment of the habitat by experts would be tedious. Remote sensing techniques offer to explore vast and difficult to access areas at lower cost. Deep learning in particular has expanded remote sensing analysis, first with convolutional networks and more recently with vision transformers. The main drawback of deep learning methods is their reliance on large labeled datasets, which are often lacking in ecological studies. However, recent studies demonstrate that Vision Transformers can be trained in a self-supervised manner to take advantage of large amounts of unlabeled data to pre-train models. These models can then be fine-tuned to learn a supervised classifier with few labeled data, or used directly as visual feature extractors for unsupervised tasks. Here, we train Vision Transformers on Very High Resolution satellite images of the Sharaan Nature Reserve in a self-supervised way: We first use the unsupervised trained backbone to generate a cluster map of the area and confront it to expert intuition. Afterwards, the model is trained in a supervised way using points where ecological habitats and land degradation have been assessed by experts on the ground. The inferences of the model can then be used on the full scale of the Sharaan Nature Reserve to generate a map of the habitats in the area. This method could be used by experts to monitor an area, design a sampling strategy and optimize field work in areas difficult to access.