Reducing spatial resolution increased net primary productivity prediction of terrestrial ecosystems: A Random Forest approach.

Net primary production (NPP) is a pivotal component of the terrestrial carbon dynamic, as it directly contributes to the sequestration of atmospheric carbon by vegetation. However, significant variations and uncertainties persist in both the total amount and spatiotemporal patterns of terrestrial NPP, primarily stemming from discrepancies among datasets, modeling approaches, and spatial resolutions. In order to assess the influence of different spatial resolutions on global NPP, we employed a random forest (RF) model using a global observational dataset to predict NPP at 0.05°, 0.25°, and 0.5° resolutions. Our results showed that (1) the RF model performed satisfactorily with modeling efficiencies of 0.53-0.55 for the three respective resolutions; (2) NPP exhibited similar spatial patterns and interannual variation trends at different resolutions; (3) intriguingly, total global NPP varied greatly across different spatial resolutions, amounting 57.3 ± 3.07 for 0.05°, 61.46 ± 3.27 for 0.25°, and 66.5 ± 3.42 Pg C yr-1 for 0.5°. Such differences may be associated with the resolution transformation of the input variables when resampling from finer to coarser resolution, which significantly increased the spatial and temporal variation characteristics, particularly in regions within the southern hemisphere such as Africa, South America, and Australia. Therefore, our study introduces a new concept emphasizing the importance of selecting an appropriate spatial resolution when modeling carbon fluxes, with potential applications in establishing benchmarks for global biogeochemical models.