Mapping of salt-affected soil using machine learning and remote sensing in Raya Kobo Valley, Ethiopia

In the Abuarie-Addisalem irrigation scheme within the Raya Valley, many productive irrigated lands have become unproductive due to the effects of salinity and sodicity. This study aimed to evaluate the predictive performances of four machine learning regression models: Random Forest (RF), Gradient Boosting Trees (GBT), Decision Tree (DT), and Support Vector Machine (SVM). The specific objectives were to (1) identify the most effective spectral indices derived from Landsat 8 OLI imagery for soil salinity mapping, and (2) determine which model provides the most accurate predictions when integrated with these indices. A total of 33 surface soil samples (0–30 cm) were collected to represent 939.46 hectares of land. The analysis was performed trough Google Earth Engine and R software statistical and graphical techniques. Results indicated that GBT and RF models were the most efficient, particularly for the Modified Soil Adjusted Vegetation Index (MSR), achieving a high coefficient of determination (R2) value of 0.93 with GBT and 0.902 with RF. RF and GBT produced similar results, classified the study area into slightly saline (31.2%), moderately saline (49.9%), and strongly saline (18.9%). Both models demonstrated lower Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and efficient for spatial mapping, while DT and SVM exhibited limited performance. Therefore, RF and GBT models are recommended for practical application.