Self-supervised learning of Vision Transformers for digital soil mapping using visual data

• Remote sensing visual data can be used to map soil properties in arid lands. • Self-supervised learning enables to leverage deep learning with few data points. • The feature space of Vision Transformers can be used to interpret model behaviour. In arid environments, prospecting cultivable land is challenging due to harsh climatic conditions and vast, hard-to-access areas. However, the soil is often bare, with little vegetation cover, making it easy to observe from above. Hence, remote sensing can drastically reduce costs to explore these areas. For the past few years, deep learning has extended remote sensing analysis, first with Convolutional Neural Networks (CNNs), then with Vision Transformers (ViTs). The main drawback of deep learning methods is their reliance on large calibration datasets, as data collection is a cumbersome and costly task, particularly in drylands. However, recent studies demonstrate that ViTs can be trained in a self-supervised manner to take advantage of large amounts of unlabelled data to pre-train models. These backbone models can then be finetuned to learn a supervised regression model with few labelled data. In our study, we trained ViTs in a self-supervised way with a 9500 km 2 satellite image of dry-lands in Saudi Arabia with a spatial resolution of 1.5 m per pixel. The resulting models were used to extract features describing the bare soil and predict soil attributes (pH H 2 O, pH KCl, Si composition). Using only RGB data, we can accurately predict these soil properties and achieve, for instance, an RMSE of 0.40 ± 0.03 when predicting alkaline soil pH. We also assess the effectiveness of adding additional covariates, such as elevation. The pretrained models can as well be used as visual features extractors. These features can be used to automatically generate a clustered map of an area or as input of random forests models, providing a versatile way to generate maps with limited labelled data and input variables.