Ensemble Learning for Soil Moisture Estimation: A Case Study of the Tibetan Plateau

Soil moisture (SM) plays a critical role in agriculture, water management, climate modeling, ecosystem health, and hazard prediction. The majority of SM estimation methods rely on individual machine learning models, which are less robust and more prone to underfitting or overfitting. To overcome the issue, we design SM estimation models based on ensemble learning. The land-surface temperature (LST) has been identified as the most influential indicator of SM in prior research. However, the MOD11A1 LST product suffers from significant data gaps, rendering it unsuitable for direct application in SM estimation. To incorporate the crucial indicator of LST in SM estimation, we employ Random Forest (RF) to impute the missing data in the MOD11A1 LST product. Subsequently, we develop an RF-based LST reconstruction and estimation by utilizing the imputed MOD11A1 LST product and in situ soil temperature measurements. Next, we construct separate SM estimation models based on Extra Tree (ERT) and Light Gradient Boosting Machine (LightGBM), respectively, using the reconstructed LST and in situ measured SM. Finally, to enhance the accuracy of SM estimation, we construct an EL-Stacking-based SM estimation model by combining predictions from ERT or LightGBM with in situ SM measurements. The results substantiate the superior performance of our proposed EL-Stacking methods compared to other models. The root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) of EL-Stacking were 0.0349-0.0657m³m^-3, 0.0227-0.0525m³m^-3 and 0.7305-0.9657, respectively. Furthermore, EL-Stacking effectively captures spatial dynamics and exhibits stronger consistency with in situ measurements.