Estimating GPP in China using different site-level datasets, vegetation classification and vegetation indices
Abstract Background Machine learning is widely used to estimate gross primary productivity (GPP) on large scales. Usually, models are trained at site level using eddy flux observations and remote sensing based vegetation indices. However, how to more effectively utilize the gradually increasing site...
Saved in:
| Main Authors: | , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
SpringerOpen
2025-08-01
|
| Series: | Ecological Processes |
| Subjects: | |
| Online Access: | https://doi.org/10.1186/s13717-025-00617-w |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Background Machine learning is widely used to estimate gross primary productivity (GPP) on large scales. Usually, models are trained at site level using eddy flux observations and remote sensing based vegetation indices. However, how to more effectively utilize the gradually increasing site observations and select different vegetation indices to improve large-scale estimations remains to be further studied, as there is currently no widely recognized optimal solution. In recent years, flux observations in China have expanded rapidly, and a new batch of publicly shared data has provided opportunities for further research. Results We tested the random forest model at the site scale and found that the model which accounts for vegetation types, using data from FLUXNET2015 and ChinaFLUX sites, was the best for estimating GPP in China (R 2 = 0.73). However, models based on different vegetation indices (leaf area index (LAI) and near-infrared reflectance of vegetation (NIRv)) showed no major difference in accuracy. Using these indices separately, we simulated monthly GPP for China from 2001 to 2022 at a 0.05° resolution. The datasets were consistent in annual totals and spatial distribution between 2001 and 2018, reporting totals of 7.52 Pg C yr−1. However, significant differences were found in spatiotemporal trends, particularly in southern China, where the linear regression coefficients were 0.04 Pg C yr– 1 and 0.07 Pg C yr– 1. Compared to other GPP datasets, our results showed slightly higher totals and trends, but they were still within a reasonable range. Conclusions The study confirms the effectiveness of differentiating between different vegetation types and adding site observations for increasing accuracy of GPP estimates. However, the difference of vegetation index does not affect the accuracy of the model, and more influences are mainly reflected in the regional simulation. These findings will help improve GPP estimates and further highlight sources of uncertainty in regional GPP simulations (input vegetation index datasets). |
|---|---|
| ISSN: | 2192-1709 |