基于深蓝算法的Sentinel-2数据气溶胶光学厚度反演

Retrieval of AOD from Sentinel-2 Data based on Deep Blue Algorithm

中高分辨率气溶胶信息对于高精度地表反射率反演以及城市空气环境质量监测具有重大意义,但在城市及稀疏植被等高亮地表区域,气溶胶光学厚度(AOD)的高精度反演一直是定量遥感领域的难点之一。以北京城市区和包头沙漠区为例,利用MODIS地表反射率产品构建先验知识约束条件,基于深蓝算法实现了13景Sentinel-2高亮地表的AOD反演。为验证算法精度,将反演结果与全球气溶胶自动观测网(AERONET)站点实测值、Sentinel-2官方插件Sen2Cor处理结果、Landsat-8反演值作对比。结果表明:①采用深蓝算法反演的AOD值与AERONET实测值具有显著的相关性(R^2>0.9,RMSE=0.056);②无论是沙漠高亮区还是植被较少的城市高亮区,Sen2Cor插件反演的AOD值整景均为固定值,无空间分布,不符合实际情况;③Sentinel-2深蓝算法反演结果与准同步过境的Landsat-8反演的AOD产品在空间分布上具有高度一致性,较好地反映了人类活动特征。相比于目前官方产品,深蓝算法适合Sentinel-2数据高亮区域的气溶胶反演,在绝对精度和空间分布趋势方面均具有明显优势。

Medium-to-high resolution aerosol information is of great significance for surface reflectance inversion and urban ambient air quality monitoring.However,the high-precision aerosol optical thickness(AOD)retrieval in bright areas,such as cities and sparse vegetation areas,has long plagued the quantitative remote sensing applications.Taking Beijing urban area and Baotou desert area as examples,using MODIS surface reflectance products to construct prior knowledge constraints,the AOD inversion of 13 scenes Sentinel-2 images in bright areas was realized based on the deep blue algorithm.To verify the accuracy of the algorithm,the result were compared with the Sentinel-2 official algorithm processing result,the Landsat-8 official aerosol products and the ground-measured AOD data from the Global Aerosol Automated Observing Network(AERONET).The results indicate that the retrieved AOD values from deep blue algorithm is significantly correlated with the measured value of AERONET(R^2>0.90,RMSE=0.0560),and the AOD spatial distributions are also well consistent with those from Landsat-8,which reflects the characteristics of human activities.But,whether in desert bright area or urban bright area with less vegetation,the AOD values retrieved by Sen2 Cor plug-in are fixed,no spatial distribution and do not conform to the actual situation.In general,compared with the current official products,the deep blue algorithm is suitable for aerosol retrieval in high-brightness areas of Sentinel-2 data,and has obvious advantages in terms of estimation accuracy and spatial distribution trend.