• 关键词：PM_2.5  Himawari-8  表观反射率  大气贡献值  6S模型  地气解耦  深度神经网络

• 基金项目：国家自然科学基金（No.41604028）；安徽省科技重大专项（No.18030801111）

• 伍万祥
• 安徽大学资源与环境工程学院, 合肥 230601

• 吴艳兰
• 1. 安徽大学资源与环境工程学院, 合肥 230601;2. 安徽省地理信息智能技术工程研究中心, 合肥 230000

• 江鹏
• 安徽大学资源与环境工程学院, 合肥 230601

• 宁海涛
• 安徽大学资源与环境工程学院, 合肥 230601

• 摘要：现阶段大气PM_2.5遥感反演方法大多数都基于卫星气溶胶光学厚度（aerosol optical depth，AOD）产品，这些产品通常是从表观反射率（top-of-atmosphere reflectance，TOA）中反演而来.直接建立TOA产品和地面站点监测的PM_2.5浓度间的反演模型能够有效降低由AOD反演所带来的误差传递，但是现阶段反演PM_2.5所用到的TOA同时耦合了地表反射率和大气贡献值，想要进一步提升反演精度则需要设法将二者分离.基于此，本文利用Himawari-8（H8）卫星数据，由6S模型进行大气校正，继而统计得到H8前6个波段之间的地表反射率关系式，再运用卫星第六波段表观反射率与地表反射率接近的特性，估算得到前5个波段的地表反射率，并扣除地表反射率得到大气贡献值，以此来达到地气解耦的目的.随后，本文基于深度神经网络构建了PM_2.5、大气贡献值、卫星亮温数据、观测角等之间的关系.以安徽省为例，反演结果表明，与不考虑地气解耦的TOA-PM_2.5方法相比，本文提出的ATM-PM_2.5方法精度更高，在未参与训练的验证站点上，ATM-PM_2.5的R²和RMSE值为0.87和13.77 μg·m^-3，相对于未经过地气解耦的TOA-PM_2.5，R²提高了20%，RMSE值降低了5.24 μg·m^-3.另外，利用H8卫星时间分辨率较高的特点，本文对安徽省域范围内进行了逐小时的PM_2.5监测，显示本文方法有潜力为PM_2.5实时监测提供数据支撑.

• Abstract：At present, the inversion methods of remotely sensing the atmospheric PM_2.5 using satellites are mainly based on the products of aerosol optical depth (AOD), which are usually retrieved from the top-of-atmosphere (TOA) reflectance. Directly modeling the relationship between the TOA and the ground-level PM_2.5 concentrations can effectively reduce the errors transferred from the AOD inversion processes. However, the TOA observation employed in the inversion of ground-level PM_2.5 concentrations is usually coupled by two parts: the surface reflectance and the atmospheric contributions. Therefore, it is necessary to separate the two reflectance parts in TOA for improving the inversion precisions of ground-level PM_2.5 concentrations. In this paper, we firstly made corrections to the observation data from Himawari-8(H8) satellite of Japan by 6S model. Then, the relationships between the surface reflectance at the first six bands of H8 satellite were obtained. Since the apparent reflectance of the H8's sixth band is very close to its surface reflectance, the surface reflectance at the other five bands can be estimated. Deducting the surface reflectance from the TOA, we finally decoupled the TOA and obtain the atmospheric reflectance. Based on the retrieved atmospheric reflectance, we established a deep neural network to model the relationship between the ground-level PM_2.5 concentrations, the atmospheric reflectance, the satellite brightness temperatures and the observation angle. We carried out a real experiment over Anhui Province region. The results indicated that the accuracy of our proposed ATM-PM_2.5 method, which decoupled the TOA, is higher than that of TOA-PM_2.5 method which employed the TOA reflectance. At the validation sites, the R² and RMSE of ATM-PM_2.5 are respectively 0.87 and 13.77 μg·m^-3. Compared with the results of TOA-PM_2.5 method, the R² dropped by 20% and the RMSE reduced 5.24 μg·m^-3 by using the atmospheric reflectance. At last, we hourly monitored the ground-level PM_2.5 concentrations using the high-temporal-resolution H8 data based on the ATM-PM_2.5 method over Anhui Province, which demonstrated that our method's potential to provide data support for the PM_2.5 real-time monitoring.

• Key words：PM_2.5  Himawari-8  top-of-atmosphere reflectance  atmospheric contribution value  6S model  ground air decoupling  deep neural network

• 
  摘要点击次数： 437 全文下载次数： 552