Comparative analysis of remote sensing and machine learning approaches for mapping susceptibility to water erosion in the Moroccan High Atlas

Water erosion is one of the main processes of soil degradation in semi-arid mountain watersheds, due to its role in accelerated sediment export, loss of soil fertility, and disruption of water resources. In the upper Tassaoute watershed, located in Morocco's High Atlas Mountains, these dynamics result in increased suspended sediment flows and the development of gullies, revealing the increased vulnerability of slopes to climatic and anthropogenic stresses.

In this context, this study proposes a comparative analysis of two complementary approaches—remote sensing and machine learning—to assess and map spatial susceptibility to water erosion. The first approach is based on the use of Sentinel-2A optical images and statistical analysis of spectral indices describing soil condition and vegetation cover. Four vegetation indices and nine soil indices are calculated and then aggregated to construct a composite explanatory variable. Regression analyses are performed between this variable and the individual indices to estimate the correlation and determination coefficients (R²), allowing their relative contribution to erosion to be assessed. Principal Component Analysis (PCA) is then applied to reduce redundancy between indices and structure the multispectral information. The first component is mainly associated with soil signatures (moisture, roughness, minerality), while the second reflects more the condition and vigor of the vegetation. On this basis, a predictive model is developed by weighting the indices according to their explanatory power and factorial contribution, leading to the development of a map classifying soils into four levels of susceptibility to degradation.

The second approach uses machine learning techniques to map susceptibility to gully erosion. An inventory of 400 occurrences, comprising 200 gullied sites and 200 non-gullied sites, was compiled based on field observations and interpretation of satellite images. These occurrences are correlated with 21 predisposing factors grouped into topographical, geological, climatic, pedological, anthropogenic, and land use variables. Five models (GLM, GBM, ANN, Random Forest, and SVM) are evaluated according to different data partitioning scenarios, with hyperparameter optimization by cross-validation. Performance is assessed using AUC-ROC and classification indicators, before producing probabilistic maps reclassified into four levels of vulnerability.

The results show that remote sensing provides a consistent and easily updatable reading of surface conditions, while machine learning significantly improves predictive capacity by integrating non-linear relationships and multiple environmental factors. Their combination provides a robust decision-making framework for targeting priority areas, guiding anti-erosion actions, and supporting land-use planning in fragile mountain environments.

Keywords: Water erosion, remote sensing, machine learning, upstream Tassaoute watershed, Morocco.