Deep transfer learning for spatiotemporal mapping of PM_2.5 nitrate across China: Addressing small data challenges in...

The proportion of nitrate in PM_2.5 is increasing in China, leading to rising health risks. However, due to the lack of a publicly accessible nationwide monitoring network, small data challenges persist in research on the spatiotemporal distribution of nitrate concentrations. Here, we employed a novel transfer learning method to compensate for data scarcity in reconstructing PM_2.5 nitrate concentrations across China. Firstly, a deep neural network (Source model) was pre-trained with extensive nitrate observations from the conterminous United states. Then, the Source model was fine-tuned with limited nitrate observations from China to yield the Transfer model. Compared to previous machine learning models constrained by insufficient nitrate observations, the Transfer model demonstrated improved generalizability through various validation strategies, with cell-based cross-validation R²= 0.72 and RMSE= 7.5 μg/m³. The results suggested that the prior knowledge learned from the large United States dataset has enhanced the generalizability of the Transfer model. Through the Transfer model, this study generated a 10 km gridded daily nitrate dataset for 2005-2020. The predictions revealed the most severe nitrate pollution in the North China Plain and major economic zones. Overall, nitrate concentrations across China showed a fluctuating upward trend from 2005 to 2014 by 0.4 μg/m³/year, followed by reductions from 2015 to 2020 by 0.6 μg/m³/year, corresponding with the implementation of air pollution control policies. This study demonstrates the effectiveness of transfer learning for addressing small data challenges and data disparity, which is particularly valuable for developing countries and regions with limited resources for environmental management and research.