Real-time prediction of corn yield from single-phase SAR and optical remote sensing data using deep learning

Crop yield is related to human survival and development. Early yield prediction and estimation in disaster scenarios are key tasks in addressing the risks posed by climate change. The main objective of this study was to use deep learning to train multisource remote sensing data to explore an instantaneous single time-phase yield prediction method and apply it before and after disasters. In this work, we utilized a convolutional neural network combined with Gaussian process regression (GPR) to train MODIS data (including surface reflectance, land surface temperature, and leaf area index), synthetic aperture radar (SAR) data and yield data to construct a crop yield prediction model and complete real-time yield prediction based on single time-phase images. We selected 802 counties in the United States Corn Belt as the study area to train the model and realize yield forecasting. The results showed that integrating SAR data, particularity during critical growth stages, enhanced model accuracy with an average R 2 exceeded 0.65, outperforming MODIS-only models. The R 2 improved by approximately 17 %, and the root mean square error decreased by approximately 8.5 % during certain time phase. The model with SAR data achieved more accurate results at least one week earlier than did the model with only MODIS data and reached a maximum R 2 of 0.74 at the middle of August. GPR can suppress the spatial estimation error. When the trained optimal model was applied to flood disasters, the minimum yield error was 38.09 kg/ha. The model provides instantaneous yield prediction for single-phase data, and the incorporation of SAR data enhances its effectiveness even in extreme weather conditions. • Deep learning model provides instantaneous yield prediction for single-phase data. • Incorporating synthetic aperture radar data promotes model for early prediction. • Gaussian process regression suppresses the spatial estimation error. • Applying the trained model to flood events, the minimum yield error was 38.09 kg/ha.