Global gross primary production (GPP) product generated by data fusion based on random forest

The results of the site-level data validation show that the DFRF-GPP estimates are closest to the site observations, and the performance of the data accuracy is better than the mainstream GPP products, such as MTE-GPP (statistical model), MODIS-GPP (empirical parametric model driven by remote sensing data) and MsTMIP-GPP (process-based land surface model). As a high-accuracy time-series product of global gross primary production, DFRF-GPP can provide a more accurate reference for the optimization calculation of global carbon sources and sinks. The DFRF-GPP fusion model also incorporates drought/high-temperature climate indicators (SPI/STI), which enhanced the sensitivity of GPP simulation results to climate anomalies and made its accuracy on climate anomaly samples higher than other GPP products. This result indicates that DFRF-GPP can better explain the impact of extreme climate on vegetation growth and development and is suitable for the study of the response of GPP to high temperature and drought. Another advantage of DFRF-GPP is that it takes into account the differences among vegetation types. Different vegetation types have different model parameters; therefore, the output results for different vegetation types with better estimation can be directly selected for use. In addition, DFRF-GPP also considers the land cover change in a given pixel among different years so that the estimation results of GPP are as close as possible to reality, which provides more potential application value for high-accuracy time-series products of global gross primary production. DFRF-GPP has two modes: daytime (DFRF-GPP_D) and nighttime (DFRF-GPP_N). DFRF-GPP_D only contains positive values, which is more consistent with other GPP products in the low latitudes of the Northern Hemisphere and Southern Hemisphere, where the GPP estimation errors are generally high. It is applicable to the estimation of terrestrial carbon flux, global carbon balance simulation and other work with high requirements for GPP simulation values. The global trend distribution and the proportion of pixels with different trends of DFRF-GPP_D are also closer to the other GPP products, indicating that DFRF-GPP_D can provide a valuable contribution to the short- or medium-term trend analysis of global GPP. DFRF-GPP_N contains negative values, but its accuracy at the site level is higher. Its reduction and explanatory ability of the spatial and temporal variations in GPP are better than those of DFRF-GPP_D. DFRF-GPP_N is suitable for work that focuses on GPP variation rather than using raw GPP sequences directly. For example, Zhu et al. (2021) analyzed the compound impact of drought and heat on GPP by calculating GPP anomaly time series. Overall, DFRF-GPP is highly reliable and can be used as a valuable data source for various applications, especially in high-temperature and drought-related studies.