Shi Su, Wang Weidong, Li Xinyue, Hang Yun, Lei Jian, Kan Haidong, Meng Xia
School of Public Health, Key Laboratory of Public Health Safety of the Ministry of Education, Key Laboratory of Health Technology Assessment of the Ministry of Health, Fudan University, Shanghai 200032, China.
Gangarosa Department of Environmental Health, Rollins School of Public Health, Emory University, Atlanta, GA 30322, USA.
Sci Total Environ. 2023 Jan 1;854:158624. doi: 10.1016/j.scitotenv.2022.158624. Epub 2022 Sep 8.
Including data of different time intervals during model development influences the predicting accuracy of PM but has not been widely discussed. Therefore, we included modeling data with multiple time windows to identify optimized modeling time windows for capturing the long-term variation of PM in China during 2005-2019. In general, we incorporated PM measurements, aerosol optical depth (AOD), meteorological parameters, land use data, and other predictors to train random forest models. The study period was separated into two phases (2013-2019 and 2005-2012) according to the availability of PM measurements. First, we trained models with two strategies of choosing time windows to compare model performance in predicting PM from 2013 to 2019, when measurements were available. Strategy 1a (ST1a) refers to training one model with all available data, and Strategy 1b (ST1b) refers to training multiple models each with one-year data. Second, we trained models with additional ten strategies (ST2a-ST2j) based on data from different time windows during 2013-2019 to compare the accuracy in predicting PM before 2013, when measurements were unavailable. The internal and external cross-validation (CV) indicated that the model performance of ST1b was better than ST1a. Predictions based on ST1a tended to underestimate PM levels in 2013 and 2014 when PM concentrations were high, and overestimate after 2017 when PM dropped dramatically. The external CV of predicting historical PM was the most robust in ST2i (averaged predictions from two models developed by 2013 and 2014 data, respectively). Models with data closer to historical years and PM levels performed better in predicting historical PM concentrations. Our results suggested that training models with data of current-years performed better during 2013-2019, and with data of 2013 and 2014 performed better in predicting historical PM before 2013 in China. The comparison provided evidence for choosing optimized time windows when predicting long-term PM concentrations in China.
在模型开发过程中纳入不同时间间隔的数据会影响PM的预测准确性，但这一点尚未得到广泛讨论。因此，我们纳入了具有多个时间窗口的建模数据，以确定用于捕捉2005 - 2019年中国PM长期变化的优化建模时间窗口。总体而言，我们纳入了PM测量值、气溶胶光学厚度（AOD）、气象参数、土地利用数据和其他预测变量来训练随机森林模型。根据PM测量值的可用性，研究期分为两个阶段（2013 - 2019年和2005 - 2012年）。首先，我们采用两种选择时间窗口的策略训练模型，以比较在有测量值时预测2013年至2019年PM的模型性能。策略1a（ST1a）指用所有可用数据训练一个模型，策略1b（ST1b）指用每年的数据分别训练多个模型。其次，我们基于2013 - 2019年不同时间窗口的数据采用另外十种策略（ST2a - ST2j）训练模型，以比较在无测量值时预测2013年之前PM的准确性。内部和外部交叉验证（CV）表明，ST1b的模型性能优于ST1a。基于ST1a的预测在2013年和2014年PM浓度较高时往往低估PM水平，而在2017年之后PM大幅下降时则高估。预测历史PM的外部CV在ST2i中最为稳健（分别由2013年和2014年数据开发的两个模型的平均预测值）。使用与历史年份和PM水平更接近的数据的模型在预测历史PM浓度方面表现更好。我们的结果表明，在2013 - 2019年期间用当年数据训练模型表现更好，而在中国预测2013年之前的历史PM时，用2013年和2014年的数据表现更好。该比较为在中国预测长期PM浓度时选择优化时间窗口提供了依据。
