Estimating high-resolution PM1 concentration from Himawari-8 combining extreme gradient boosting-geographically and temporally weighted regression (XGBoost-GTWR)
Estimating high-resolution PM1 concentration from Himawari-8 combining extreme gradient boosting-geographically and temporally weighted regression (XGBoost-GTWR)
复制标题
结合极端梯度增强地理和时间加权回归估计 Himawari-8 的高分辨率 PM1 浓度 (XGBoost-GTWR)
DOI：
10.1016/j.atmosenv.2020.117434
复制
发表时间：
2020
影响因子：
5
通讯作者：
Guo Jianping
中科院分区：
文献类型：
--
作者：
Li Rui;Cui Lulu;Fu Hongbo;Meng Ya;Li Junlin;Guo Jianping
作为一种更细的颗粒，小于 1 μm 的颗粒物 (PM1) 对雾霾的形成和人类健康起着重要作用。然而，粗时间分辨率和低估计精度严重损害了绘制 PM1 浓度的能力，这很大程度上是由于忽略了 PM1 的空间或时间自相关性。为了改进高分辨率 PM1 的估计，我们使用 Himawari-8 气溶胶光学深度 (AOD)、气象因素和地理协变量开发了一种名为极端梯度增强 (XGBoost) 的新型时空模型 - 地理和时间加权回归 (GTWR)。浙江省PM1的估算表明，与其他11个模型相比，XGBoost-GTWR方法具有更强的预测能力（10倍交叉验证R2=0.83，均方根误差（RMSE）=10.72μg/m3）。此外，还进行了外推测试来验证混合模型的稳健性，结果表明 XGBoost-GTWR 可以准确预测带外 PM1 浓度（R2= 0.75 (0.60)，RMSE = 12.71 (12.58) μg/m3）。 PM1浓度表现出明显的空间异质性，衢州最高（34.72±1.77μg/m3），舟山最低（26.39±1.56μg/m3）。从季节上看，冬季PM1浓度最高（39.06±3.08μg/m3），其次为春季（32.54±3.09μg/m3）和秋季（30.97±4.50μg/m3），夏季最低（25.57±5.22μg/m3）。气溶胶排放量高和不利的气象条件（例如边界层高度低和降水不足）是造成冬季PM1浓度峰值的关键因素。此外，PM1浓度呈现显着的日变化，在当地太阳时（LST）1500时达到峰值，但在当地太阳时（LST）1000时达到最低值。该方法增强了我们从太空估算每小时PM1的能力，为改进细颗粒物相关健康影响的评估奠定了坚实的数据基础。
As a much finer particle, particulate matter less than 1 μm (PM1) plays an important role on the haze formation and human health. However, the capability of mapping PM1concentration is severely impaired by coarse temporal resolution and low estimation accuracy, largely due to the neglect of spatial or temporal autocorrelation of PM1. In order to improve the estimation of high-resolution PM1, here we developed a novel spatiotemporal model named extreme gradient boosting (XGBoost)-geographically and temporally weighted regression (GTWR) using Himawari-8 aerosol optical depth (AOD), meteorological factors, and geographical covariates. The estimation of PM1over Zhejiang province showed that XGBoost-GTWR method was characterized by greater predictive ability (10-fold cross-validation R2= 0.83, root mean squared error (RMSE) = 10.72 μg/m3) compared with other 11 models. Additionally, the extrapolation test was performed to validate the robustness of the hybrid model and the result demonstrated that XGBoost-GTWR can accurately predict the out-of-band PM1concentration (R2= 0.75 (0.60), RMSE = 12.71 (12.58) μg/m3). The PM1concentration displayed pronounced spatial heterogeneity, with the highest value in Quzhou (34.72 ± 1.77 μg/m3) and the lowest in Zhoushan (26.39 ± 1.56 μg/m3), respectively. In terms of the seasonality, the highest PM1concentration was observed in winter (39.06 ± 3.08 μg/m3), followed by those in spring (32.54 ± 3.09 μg/m3) and autumn (30.97 ± 4.50 μg/m3), and the lowest one in summer (25.57 ± 5.22 μg/m3). The high aerosol emission and adverse meteorological conditions (e.g., low boundary layer height and lack of precipitation) were key factors accounting for the peak PM1concentration observed in winter. Also, the PM1concentration exhibited significant diurnal variation, peaking at 1500 local solar time (LST) but reaching the lowest value at 1000 LST. This method enhances our capability of estimating hourly PM1from space, and lays a solid data foundation for improving the assessment of the fine particle-related health effect.