Drought is a frequent natural hazard in the Huaihe River Basin (HRB). Traditional agricultural drought monitoring methods have defects in spatial continuity, so developing an accurate agricultural drought monitoring model at regional scale is necessary. As a popular method, random forest (RF) is widely used due to its high prediction accuracy. However, RF may have significant bias in regression at times, especially for extreme values. The standardized precipitation evapotranspiration index for the 3-month time scale (SPEI3) was used as the dependent variable, and the multi-source satellite product from tropical rainfall measure mission (TRMM) and moderateresolution imaging spectroradiometer (MODIS) was fused by RF to construct agricultural drought monitoring model in two regions of the HRB from April to October in 2001—2014. The accuracy of four bias-correcting methods, including simple linear regression (SLR), bias corrected method (BC), residual rotation method (RR) and best-angle residual rotation method (BRR) were assessed by determination coefficient (R2), root mean square error (RMSE) and correct percentage of drought grades. The best bias-correcting method was used to establish agricultural drought monitoring model, which was called bias correcting random forest drought condition model (BRFDC). The relative soil humidity data and drought records were applied to test the monitoring capacity of BRFDC model. The results showed that all of four bias-correcting methods improved the performance compared with original RF. The BRR method performed better with R2 were 0.897 and 0.874, and RMSE were 0.335 and 0.362, which reduced the residuals efficiently. Additionally, the BC method performed better by the accuracy rate of different ranks of drought, especially the accuracy of extreme drought was between 33.3% and 50.0%. The BC method was applied to construct BRFDC at last. Compared with SPEI3, the outputs of BRFDC model had more significant correlation with soil relative humidity at most stations. Finally, the drought maps during the period from May to October in 2001 were produced by inverse distance weighting method (IDW), original RF and BRFDC model, and all of them showed a strong visual agreement. In particular, the extreme drought conditions were successfully monitored by BRFDC model.