NDVI joint process-based models drive a learning ensemble model for accurately estimating cropland net primary productivity (NPP)

Abstract

The accurate estimation of cropland net primary productivity (NPP) remains a significant challenge. We hypothesized that incorporating prior information on NPP simulated by process-based models into normalized difference vegetation index (NDVI) data would improve the accuracy of cropland ecosystem NPP estimations. We used NDVI, MNPP (NPP of process-based model), and SNPP (statistic-based NPP) data estimated by nine process-based models and yield statistics to build a learning ensemble of the random forest model (LERFM). We used the new model to re-evaluate the cropland NPP in China from 1982 to 2010. Large spatial discrepancies among MNPPs, which indicate uncertainties in cropland NPP estimation using different methods, were observed when compared to SNPP. The LERFM model showed a slightly underestimation of only −0.37%, while the multi-model average process-based model (MMEM) strongly underestimated −15.46% of the SNPP. LERFM accurately estimated cropland NPP with a high simulation skill score. A consistent increasing trend in the LERFM and MMEM NPP during 1982–2010 and a significant positive correlation (r = 0.795, p < 0.001) between their total NPP indicate that the LERFM model can better describe spatiotemporal dynamic changes in cropland NPP. This study suggests that a learning ensemble method that combines the NDVI and process-based simulation results can effectively improve cropland NPP.