Author Affiliations
1School of Earth Sciences, Zhejiang University, Hangzhou 310027, China1浙江大学地球科学学院,地理与空间信息研究所,杭州 3100272Zhejiang Provincial Key Laboratory of Geographic Information Science, Hangzhou 310028, China2浙江省资源与环境信息系统重点实验室,杭州 310028show less
Fig. 1. Spatial distribution of the PM2.5 monitoring stations of China in 2017 and the spatial partitions of the training, validation, and testing datasets
Fig. 2. Pre-processing of the experimental dataset
Fig. 3. Definition of the GNNWR model for PM2.5 estimation
Fig. 4. Structure design of the spatial weighted neural network
Fig. 5. Training and validation procedures of the GNNWR model
Fig. 6. Performance variations for the training and validation datasets of the GNNWR model
Fig. 7. Estimates of the annual mean PM2.5 across China in 2017
Fig. 8. Details comparison of the annual mean PM2.5 estimates between the GWR and GNNWR models
Fig. 9. Spatial distribution of absolute estimation errors of the annual mean PM2.5 of China in 2017 for the OLR, GWR, and GNNWR models
数据类型 | 数据名称 | 变量名称 | 时间分辨率 | 空间分辨率 | 数据来源 |
---|
站点 | PM2.5监测站点 | PM2.5 | h | - | 中国国家气象局 | 遥感 | 气溶胶 | AOD | d | 3 km、10 km | LAADS | 气象 | 2 m温度 | TEMP | h | 0.5° | ERA5 hourly data | 降水量 | TP | h | 0.5° | ERA5 hourly data | 10 m风速 | WS | h | 0.5° | ERA5 hourly data | | 10 m风向 | WD | h | 0.5° | ERA5 hourly data | 地理 | 地形 | DEM | - | 1弧分 | NOAA |
|
Table 1. Data sources and description
模型名称 | 带宽优化准则 | 核函数 |
---|
类型 | 结构 |
---|
GWR-AFG | AICc | 固定型 | Gaussian | GWR-AAB | AICc | 适应型 | Bi-square |
|
Table 2. Settings of GWR models
输入层 | 隐含层1 | 隐含层2 | 隐含层3 | 输出层 |
---|
1004 | 512 | 256 | 128 | 7 | 学习率 | epoch最大值 | 批处理大小 | Dropout | | 0.2 | 20 000 | 64 | 0.8 | |
|
Table 3. Settings of architectures and hyper-parameters for the GNNWR model
变量 | PM2.5/ | AOD | DEM | 温度/ | 降水量/ | 风速/ | 风向/° |
---|
相关系数 | - | 0.564 | -0.345 | 0.135 | -0.234 | -0.373 | -0.461 | 显著性水平 | - | 0.001 | 0.001 | 0.001 | 0.001 | 0.001 | 0.001 | 方差膨胀因子 | - | 3.005 | 1.636 | 4.639 | 2.107 | 2.357 | 4.184 | 平均值 | 46.07 | 0.540 | 393.460 | 288.000 | 8.95E-05 | 7.080 | 147.990 | 标准差 | 15.56 | 0.180 | 658.460 | 5.170 | 4.82E-05 | 0.940 | 43.150 | 最小值 | 8.34 | 0.070 | -5.250 | 271.810 | 1.06E-06 | 4.430 | 80.110 | 最大值 | 103.89 | 1.140 | 4539.960 | 297.720 | 2.34E-04 | 11.500 | 236.040 |
|
Table 4. Exploratory analysis and descriptive statistics of the experimental dataset across China in 2017
模型 | 训练集(拟合精度) | | 测试集(预测精度) |
---|
R2 | RMSE | MAE | MAPE/% | AICc | F1 | p-value | R2 | RMSE | MAE | MAPE/% |
---|
OLR | 0.517 | 10.578 | 8.132 | 20.2 | 7579.202 | - | - | | 0.477 | 12.023 | 8.838 | 19.7 | GWR-AFG | 0.759 | 7.478 | 5.797 | 14.5 | 6999.491 | 0.538 | 0.010 | | 0.683 | 9.359 | 6.963 | 16.1 | GWR-AAB | 0.898 | 4.860 | 3.478 | 8.7 | 6673.537 | 0.297 | 0.010 | | 0.675 | 9.484 | 5.423 | 12.6 | GNNWR | 0.914 | 4.452 | 3.235 | 7.9 | 5833.002 | 0.137 | 0.010 | | 0.831 | 6.837 | 4.688 | 11.0 |
|
Table 5. Fitting and prediction performances of the PM2.5 estimates for the OLR, GWR, and GNNWR models
变量 | GWR-AFG | | GWR-AAB | | GNNWR |
---|
F2 | p-value | F2 | p-value | F2 | p-value |
---|
常量项 | 2407.9 | 0.001 | | 1269.1 | 0.001 | | 590.5 | 0.001 | AOD | 1379.7 | 0.001 | 1199.4 | 0.001 | | 243.7 | 0.001 | DEM | 1983.9 | 0.001 | 1272.2 | 0.001 | | 146.7 | 0.001 | 温度 | 3783.7 | 0.001 | 1229.6 | 0.001 | | 127.7 | 0.001 | 降水量 | 1237.4 | 0.001 | 1170.1 | 0.001 | | 214.7 | 0.001 | 风速 | 1626.0 | 0.001 | 1016.2 | 0.001 | | 159.0 | 0.001 | 风向 | 2379.8 | 0.001 | 1081.6 | 0.001 | | 73.6 | 0.001 |
|
Table 6. Spatial nonstationarity diagnosis of each variable in the GNNWR model