Improving the Gross Primary Production Estimate by Merging and Downscaling Based on Deep Learning

Abstract

A reliable estimate of the gross primary productivity (GPP) is crucial for understanding the global carbon balance and accurately assessing the ability of terrestrial ecosystems to support the sustainable development of human society. However, there are inconsistencies in variations and trends in current GPP products. To improve the estimation accuracy of GPP, a deep learning method has been adopted to merge 23 CMIP6 data to generate a monthly GPP merged product with high precision and a spatial resolution of 0.25°, covering a time range of 1850–2100 under four climate scenarios. Multi-model ensemble mean and the merged GPP (CMIP6DL GPP) have been compared, taking GLASS GPP as the benchmark. Compared with the multi-model ensemble mean, the coefficient of determination between CMIP6DL GPP and GLASS GPP was increased from 0.66 to 0.86, with the RMSD being reduced from 1.77 gCm−2d−1 to 0.77 gCm−2d−1, which significantly reduced the random error. Merged GPP can better capture long-term trends, especially in regions with dense vegetation along the southeast coast. Under the climate change scenarios, the regional average annual GPP shows an upward trend over China, and the variation trend intensifies with the increase in radiation forcing levels. The results contribute to a scientific understanding of the potential impact of climate change on GPP in China.