基于卫星数据和深度学习方法预测京津冀地区地表NO_(2)浓度

Deep learning architecture based on satellite remote sensing data for estimating ground-level NO_(2) across Beijing-Tianjin-Hebei Region

二氧化氮(NO_(2))对人类健康和气候变化有着诸多负面影响,随着中国城镇化和工业化进程加速,NO_(2)污染成为人们日益关注的问题。相关研究表明传统的单个站点监测结果只能代表数平方公里内的污染物水平,无法提供大尺度的污染物分布信息。相比于站点监测,卫星遥感可以提供大尺度且时空连续的数据,为研究大气污染提供了新的角度。基于哨兵5P卫星的NO_(2)柱浓度数据和气象、人口密度等其他辅助数据,构建了对地表NO_(2)进行预测的深度神经网络(DNN)模型。并使用两种交叉验证方法对该模型进行验证。在基于样本的验证中,模型的决定系数R^(2)、均方根误差(RMSE)和平均预测误差(MAE)分别为0.80、7.72μg/m^(3)和5.31μg/m^(3);在基于站点的验证中,模型的R^(2)、RMSE和MAE分别为0.74、8.95μg/m^(3)和6.01μg/m^(3),两种验证结果都表明DNN模型具有较好的整体预测能力和空间泛化性。此外,与经典的地学统计和机器学习算法对比结果表明DNN预测性能优于其它方法。最后用训练好的模型获得了京津冀地区0.1°的NO_(2)分布图。

Nitrogen dioxide(NO_(2))has many adverse impacts on human health and climate change.With the acceleration of urbanization and industrialization in China,NO_(2) pollution has become a growing concern.However,releveant research shows that the traditional monitoring results of a single site can only represent the concentration of pollutants within a few square kilometers,and cannot provide large-scale pollutant distribution information. Compared with site monitoring, satellite remote sensing can provide large-scaleand spatiotemporal continuous data. Based on NO_(2) column densities of Sentinel-5 Precursor and otherauxiliary data such as weather and population density, a deep learning model (DNN) to predict groundlevelNO_(2) concentration is built in this work, and then the model is verified by two cross-validationstrategies. In the sample-based cross validation, the determination coefficient R^(2), root mean square error(RMSE) and mean absolute error (MAE) of the model are 0.80、7.72 μg/m^(3) and 5.31 μg/m^(3), respectively,while in the site-based cross validation, they are 0.74、8.95 μg/m^(3) and 6.01 μg/m^(3), respectively. Both ofthe two cross-validation results indicate that the DNN model has excellent overall predictive performanceand spatial generalization ability. In addition, the comparisons with the other classic geostatistics andmachine learning algorithms also show that the predictive performance of the deep learning algorithm isbetter than that of the other methods. Finally, the trained model is applied to generate NO_(2) distribution with0.1° spatial resolution across Beijing-Tianjin-Hebei region.