Mapping of Water-Induced Soil Erosion Using Machine Learning Models: A Case Study of Oum Er Rbia Basin (Morocco)

• Published in: Earth systems and environment Vol. 7; no. 1; pp. 151 - 170
• Main Authors: Barakat, Ahmed, Rafai, Mouadh, Mosaid, Hassan, Islam, Mohammad Shakiul, Saeed, Sajjad
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 2023; Springer Nature B.V
• Subjects: Agriculture; Algorithms; Climate; Climate change; Climate Change/Climate Change Impacts; Decision trees; Deforestation; Earth and Environmental Science; Earth Sciences; Earth System Sciences; Environmental Science and Engineering; Erosion rates; Geographic information systems; Geography; Hydrology; Land degradation; Land use; Learning algorithms; Machine learning; Mapping; Methods; Model testing; Modelling; Monitoring/Environmental Analysis; Original Article; Performance prediction; Remote sensing; Sediment deposits; Sensitivity analysis; Soil degradation; Soil erosion; Soil erosion control; Soil water; Morocco; Soil erosion modeling; Oum Er Rbia Basin; Geo-environmental factors; Accuracy analysis; Machine learning; Susceptibility mapping
• ISSN: 2509-9426; 2509-9434
• DOI: 10.1007/s41748-022-00317-x

The basin of Oum Er Rbia River (Morocco) has been greatly affected by water-related erosion leading to loss of soils, land degradation, and deposits of sediment in dams. With this motivation, we estimated the soil erosion vulnerability using three machine learning (ML) techniques, namely random forest (RF), k-nearest neighbor (kNN), and extreme gradient boosting (XGBoost). From a total of 3034 known soil erosion locations, identified from google earth and other data archives and published works, 80% were used for soil erosion model training, with the remaining 20% used for model testing. The Boruta algorithm identified 17 most relevant environmental and geological factors, selected as the main contributors for modeling the soil erosion by water. The performance of the ML models was evaluated based on sensitivity, specificity, precision, and the Kappa coefficient. This evaluation revealed that RF, kNN and XGBoost are very good to excellent models for water-based soil erosion prediction in the study area. Soil erosion susceptibility (SES) maps were generated for all models, compared, and subsequently validated using the receiver-operating characteristic (ROC) curves and area under the curve (AUC). According to ROC results, all derived maps are reliably good predictors of potential soil erosion rates by water. The AUC values attest that all models performed comparably well, with very high accuracies, although RF had a better predictive performance (AUC = 92%) than the others (kNN AUC = 90%, XGBoost AUC = 91%). Hence, the methodology adopted in this study, based on ML algorithms, can be a helpful tool for soil erosion modeling and mapping in similar settings elsewhere. Moreover, our results provide beneficial information for decision-makers to propose appropriate measures to avoid soil loss in the Oum Er Rbia Basin.