Estimation of soil moisture from remote sensing products using an ensemble machine learning model: a case study of Lake Urmia Basin, Iran

This study investigated the capability of remote sensing soil moisture (SM) datasets to estimate in-situ SM over the Lake Urmia Basin in Iran. A novel meta-estimating approach, called Voting Regression (VR), was used to combine the Gradient Boosting (GB) and Support Vector Regression (SVR) algorithms for developing a new hybrid predictive model named GB-SVR. Six SM products from the Global Land Data Assimilation System (GLDAS), Advanced Microwave Scanning Radiometer 2 (AMSR2), and Soil Moisture Active Passive (SMAP) were used to predict SM at 40 in-situ SM sampling locations. The performance of the proposed novel forecasting technique was evaluated using Correlation Coefficient (CC), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). The results showed the superiority of GB-SVR compared to GB and SVR, with an average improvement of 17%, 10%, and 13% in CC, RMSE, and MAE, respectively, in predicting in-situ SM. The model performance in different climates, soil textures, and land covers showed its better prediction accuracy in croplands \(({R}^{2}=0.86)\), loam soil \(({R}^{2}=0.74)\) and cold climate \(\left({R}^{2}=0.71\right)\), while the least in clay soil and barren lands. Besides, the in-situ SM prediction using remote sensing SM data performed better than that obtained using in-situ air and soil temperature. The proposed methodology can be used for accurate SM prediction in regions lacking in-situ SM data.