Ensemble learning based on remote sensing data for monitoring agricultural drought in major winter wheat-producing areas of China

Drought is mainly triggered by the lack of precipitation, which can lead to insufficient water supply for crops thus affecting their growth and development. Reliable drought monitoring is crucial to understanding drought risk and avoiding drought-induced crop yield losses. Based on the Stacking regression method and multiple remotely-sensed drought factors from 2001 to 2017, this study developed an ensemble learning framework for monitoring agricultural drought in major winter wheat-producing areas in China. Stacking used five machine learning algorithms, namely, extreme gradient boosting, support vector regression, extra trees, and multi-layer perceptron, as the base learners to model the relationship between remote sensing drought factors and 1-, 3-, and 6-month standardized precipitation evapotranspiration index (SPEI). In this study, county-level winter wheat yield records and drought maps provided by the Global SPEI database (SPEIbase) were adopted to assess the suitability of Stacking-predicted SPEI drought maps in agricultural drought monitoring. The results show that Stacking outperformed other machine learning algorithms in terms of estimation accuracy, with the highest R2 value of 0.77 and the lowest root mean square error (RMSE) of 0.47. The longer the time scale of model-predicted SPEI, the higher its correlation with detrended winter wheat yields. The comparison with the drought maps of SPEIbase shows that the Stacking-predicted drought maps successfully captured the spatial pattern and intensity change of drought events. The approach presented in the study has good applicability for agricultural drought monitoring and could be extended to the rest of the areas.