Retrieving soil moisture from grape growing areas using multi-feature and stacking-based ensemble learning modeling

Soil moisture (SM) is an essential parameter for crop growth and development, and temporal and spatial variation in SM in agricultural fields varies by crop type due to corresponding crop growing characteristics and cultivating patterns. Few studies have performed SM retrieval in grape growing areas, and SM estimation using only spectral reflectance (SR) or derived drought indices may not be wholly accurate. In this study, seven features based on evapotranspiration (ET), land surface temperature (LST), and SR were derived from Moderate Resolution Imaging Spectroradiometer (MODIS) data (8-day temporal resolution and 500 m spatial resolution) and integrated with topography feature to determine SM during the main growth stages of grape in the eastern foothills of the Helan Mountains from 2009 to 2018. A total of 25-period models covering April–October (8-day temporal resolution) were constructed. Finally, the models were achieved to retrieve SM, and the spatiotemporal distribution pattern was analyzed. The stacking ensemble algorithm can integrate multiple models for better retrieval accuracy and overcome the limitations of individual machine learning models. We also compared the SM estimation accuracy of three single machine learning models (Category boosting, random forest, and gradient boosting decision tree) with the ensemble model using the stacking algorithm. The results indicated that the stacking-based ensemble model could retrieve SM more accurately and stably than any individual machine learning algorithm, and the stacking-based ensemble model showed good applicability during the main growth stages of grape. The average coefficient of determination (R2) and root mean square error (RMSE) of the 25-period multi-feature stacking-based models were 0.7504 and 0.0245 m3/m3. Overall, the application results showed that spring and summer droughts were more severe than autumn droughts in the study area during the main growth stages of grape. Our findings indicate that using multiple features and the stacking-based ensemble model could improve SM estimation accuracy in grape growing areas.