A generalized model to predict large-scale crop yields integrating satellite-based vegetation index time series and phenology metrics

Convenient and reliable large-scale crop yield prediction is needed when formulating administrative plans and ensuring food security, especially under changing climate and international conditions. In this study, we explored Moderate Resolution Imaging Spectroradiometer (MODIS) vegetation indices-and phenology-based yield prediction generalization model taking the US Corn Belt as an example. We calculated the normalized difference vegetation index (NDVI) and 2-band enhanced vegetation index (EVI2) time series, which were adjusted using greenup dates derived from the Land Cover Dynamics product MCD12Q2. Based on the adjusted VI (NDVI, EVI2) time series, the VI change rate (dVI) time series was calculated, which represents crop growth rate. The first step was to cluster the adjusted VI and dVI time series, called 'greenup groups', according to corresponding greenup dates with a five-day interval. Then in different greenup groups, we constructed empirical univariate models with VI having the maximum correlation with crop yield, and multivariate models with VI and dVI, which were also used to construct the generalized model. After clustering, the days with maximum VI correlation gradually decreased as greenup days increasing, and the univariate VI model and multivariate VI and dVI model performances in different groups improved. The generalized models with specific VI and dVI variables in each group predicted yields of corn and soybean with R-2 values mainly ranging from 0.55 to 0.75 and 0.55 to 0.70, while RMSE mainly ranging from 1000 to 1500 kg/ha and 300 to 400 kg/ha for both NDVI and EVI2 from 2008 to 2018 with leave-one-year-out cross-validation for all groups. The model using MODIS data was convenient and scalable with limited data requirements and date-determined variables after greenup, and offered a generalized method to predict crop yields at a large scale before harvest with good performance.