Prediction of Spatial Variations of Soil Moisture Using Random Forest Method and Environmental Features derived from Satellite Images in Marghab Basin of Khuzestan

Preparation of soil moisture map with high spatial resolution and appropriate quality is important in land management. Due to the lack of meteorological stations in watersheds, especially in mountainous areas, field measurement to study changes in soil moisture is time-consuming, costly and error-prone. To achieve a suitable model for spatial prediction of soil moisture in low rainfall season in Marghab Basin of Khuzestan province, 683 km2 area, field sampling was performed in 174 points at four standard depths (0-5, 5-15, 15-30, and 30-60 cm) correspond to the global digital soil mapping project. The spatial distribution of soil moisture was mapped by a machine learning model using two sets of remote sensing data, including surface biophysical features derived from Landsat-8 and Sentinel-2 satellite images, and topographic features derived from the digital elevation model. The most suitable auxiliary variables for predicting soil moisture were selected via the Recursive Feature Elimination (RFE) method. The results of the trend of mean changes in soil moisture from the first to the fourth layer were observed to be 2.2, 3.24, 3.41, and 4.6%, respectively. At the surface depths (0-5 cm), biophysical covariates had more impact on spatial variations of soil moisture, and at the lower depths (5-15, 15-30, and 30-60 cm), topographic attributes showed higher importance. The evaluation of RF model in relation to the type of image used for the production of biophysical features showed that based on the concordance correlation coefficient (CCC), the model performance increased between 1.28 to 3.66 in standard soil depths when using Sentinel-2 images compared to Landsat 8. Generally, the RF model and biophysical features were extracted from the Sentinel-2 satellite along with topographic attributes at the watershed scale are able to provide soil moisture prediction maps with acceptable accuracy.