Leveraging multisource data for accurate agricultural drought monitoring: A hybrid deep learning model

Accurate monitoring of agricultural droughts in data-scarce areas remains a challenge due to their intricate spatiotemporal patterns. Deep learning represents a promising approach for developing efficient drought monitoring models. In this study, a hybrid deep learning model, combining convolutional neural network and random forest (CNN-RF), is proposed to monitor agricultural droughts in a mountainous region located in Southwest China. The model integrates multisource data obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite sensor, Global Land Data Assimilation System (GLDAS), Climate Hazards Group InfraRed Precipitation with Station Data (CHIRPS) and digital elevation model (DEM) to reproduce a station-based 3-month Standardized Precipitation Evapotranspiration Index (SPEI-3) during 2001–2020. Performance evaluation of the proposed model utilized an in situ soil moisture dataset and grain yields as benchmarks. The results demonstrated the superiority of the CNN-RF model over both the CNN and RF models in terms of estimating SPEI-3 and forecasting drought categories, as quantified by the lowest root mean square error (RMSE<0.4), the highest correlation coefficient (CC>0.9) and the multi-class receiver operating characteristic (ROC) based area under curves (AUC) (AUC=0.86). The CNN-RF model successfully reproduced the spatial heterogeneity of the drought pattern while maintaining temporal and spatial consistency with actual drought conditions. Notably, strong consistency was observed between the simulated SPEI-3 and the 3-month Standardized Soil Moisture Index (SSMI-3) (CC=0.42, p < 0.01). Moreover, the model-estimated drought areas of cropland in the winter and early spring months exhibited a significant correlation with summer harvest grain yields (CC<−0.45, p < 0.05). Another advantage of the CNN-RF model is its ability to generalize well with limited training samples. This study introduces a scalable, simple, and efficient method for reliably monitoring agricultural droughts over large areas by leveraging freely available multisource data, which can also be easily adapted for monitoring agricultural droughts in other vegetated regions with limited ground observations.