Time-series NDVI and greenness spectral indices in mid-to-late growth stages enhance maize yield estimation

Abstract

Accurate and timely satellite-based yield estimation is crucial for agricultural policy and food security. The Normalized Difference Vegetation Index (NDVI), which represents canopy greenness, is widely used to predict crop yields. However, most studies have developed empirical models based only on the relationship between instantaneous spectral information and final yield, ignoring the canopy physiological changes that can be captured by spectral monitoring during crop yield formation, which limits the applicability of these models. In this study, based on time-series NDVI data, we found that the trend of NDVI changes in the mid-to-late stages of maize is closely related to yield, and accordingly constructed four greenness spectral indices: NDVI decline rate (DR), average daily accumulation of NDVI (ADA), time-series NDVI standard deviation (STD), and leaves greenness duration (LGD). We assessed the validity of the GSIs under two scenarios: (1) When the time-series data were consistent across years, using three strategies, namely utilizing the complete NDVI datasets, estimating yield in advance, and accounting for missing data due to meteorological conditions; (2) We ask whether the GSIs remain valid when the time-series data are inconsistent across years. Results under time-series consistency showed that combining these GSIs derived from the complete NDVI dataset with the third-period NDVI achieved the highest model accuracy (R² = 0.7, rRMSE = 12.3 %). Approximately one month before harvest, GSIs improved estimation accuracy (R² = 0.661, rRMSE = 13.2 %), increasing R² by 0.023 and reducing rRMSE by 0.4 %. When NDVI data were incomplete due to meteorological conditions, GSIs still enhanced yield estimation, increasing R² by 0.007–0.077 and reducing rRMSE by 0.1 %-1.1 %. Even with the inconsistency of time-series data across years, the accuracy of yield estimation improved by 28 % after integrating GSIs. These results demonstrate the adaptability and reliability of GSIs under different conditions.