Predicting Weather Station-Scale GPP and ET with Deep Learning for Climate-Resilient Corn Production in the U.S.

Over the past two decades, extreme climate and weather events have become increasingly frequent in the United States, and the carbon–water cycle of corn ecosystems has shown high sensitivity to climate change. However, traditional simulation methods that rely on coarse-scale reanalysis data are unable to reflect changes in local water and heat conditions accurately. This study combines in situ meteorological observations with remote sensing, using a long short-term memory model to simulate the daily gross primary productivity (GPP) and evapotranspiration (ET) of 684 corn-growing meteorological stations in the United States. In summer, GPP and ET showed a spatial pattern of gradual decrease from the humid eastern region to the arid western region, and the multi-year daily averages at meteorological stations showed a single-peak pattern. The sensitivity of GPP and ET changes is mainly influenced by leaf area index (LAI) and shortwave radiation downward changes, which together explain more than 90% of the main variation in GPP and ET at the meteorological stations. The 2012 drought caused a general decline in GPP and ET, with the peak occurring approximately 15 days earlier than usual. Water use efficiency (GPP/ET) decreased at 85% of the sites (p < 0.05), but photosynthesis per unit leaf area (GPP/LAI) increased at 63% of the sites (p < 0.05). This study demonstrates the importance of meteorological station-scale data for understanding carbon–water flux dynamics in cornfields. Integrating the models developed in this study with medium-to-long-term climate projections will further guide climate-informed agricultural water management and provide reliable accounting and pricing tools for agricultural land carbon markets and carbon trading.