Improving yield prediction based on spatio-temporal deep learning approaches for winter wheat: A case study in Jiangsu Province, China

Accurately and timely predicting yield and identifying its drivers are crucial for adjusting agricultural interventions, thereby responding to climate change and ensuring food security. Winter wheat phenology exhibited pronounced temporal and spatial variations at the county level in Jiangsu province. This study constructed a robust long short-term memory (LSTM) model that incorporates multiple data sources, including meteorological parameters, soil attributes, terrain characteristics, remote sensing data, and heterogeneous wheat phenology, aiming at forecasting the county-level wheat yield from year 2005 to 2020, with an emphasis on different precipitation year-types. The results demonstrated that the LSTM model could explain 71% of yield variations because of its ability to deal with time cumulative effects and high-dimensional data. Its performance outperformed that of machine learning algorithms, even in a normal year (RMSE = 0.26 t/ha) and an extreme year (RMSE = 0.29-0.33 t/ha). Time-series data (meteorological factors and vegetation indices), soil attributes, as well as topographic data that capture spatial and temporal heterogeneity, can achieve accurate wheat prediction approximately three months before harvest. Feature importance analysis showed meteorological factors, especially precipitation, are the most important drivers for yield prediction. This study developed a promising and scalable framework for spatio-temporal yield prediction by integrating available multi-source data to respond to climatic variations and provided insights for policymakers and farmers to adjust agricultural interventions.