深度学习与辐射传输模型协同的气溶胶反演

Aerosol Retrieval Using Deep Learning and Radiative Transfer Model

传统气溶胶反演方法通常先基于模型假设确定地表反射率,但反演结果会受到假设的影响;而深度学习方法基于数据驱动,能在气溶胶定量反演中得到更加准确、高效的结果,但模型训练需要充足的优质样本数据支持。为此,使用大气辐射传输模型构建模拟样本,支持深度学习方法实现气溶胶定量反演,旨在解决当前训练数据代表性不足、数据获取困难的问题。利用辐射传输模型模拟不同参数条件下传感器获得的辐射信息,考虑概率组合及筛选标准限制进行模拟数据构建,并使用深度置信网络(DBN)对模拟样本进行训练,获得气溶胶反演模型。将模型应用于Landsat-8数据,在中国北京地区开展气溶胶反演实验。最后使用AERONET地面站点的实测数据对反演结果进行精度验证。验证结果表明,模型估算的气溶胶与站点测量值吻合良好,相关系数为0.8989,均方根误差为0.1029,约74.05%的估算值在误差标准内。本文提供了一种基于辐射传输方程构建样本数据集的思路,可减少样本质量与数量导致的局限性,实现深度学习方法对气溶胶光学厚度的高精度反演。

Objective Atmospheric aerosols are an important component of the earth-atmosphere system,exerting significant influence on various aspects,such as climate change,air quality,and human health.Traditional aerosol retrieval methods,such as dark target(DT)and deep blue(DB)algorithms,assume surface reflectance parameters during atmospheric radiative transfer processes and construct Look-up tables to retrieve aerosol optical depth(AOD).However,these methods simplify retrieval factors based on prior knowledge,resulting in error accumulation and propagation.Additionally,Look-up tables are usually constructed based on blue or red bands to greatly limit the image information utilization.With the emergence of various machine learning algorithms,deep learning algorithms have become the preferred AOD retrieval method.The quality and quantity of the deep learning training dataset determine the accuracy and applicability of the final model.Meanwhile,current dataset construction faces the problem of biased and insufficient training datasets due to the sparsity of ground station(AERONET)data,the limitation of image cloud cover,and satellite replay cycle,which greatly affects the accuracy of the retrieval model.Therefore,we propose an atmospheric radiative transfer model to construct a simulated dataset that conforms to real scenes,supporting deep learning methods to achieve quantitative aerosol retrieval.We hope that this method can solve the difficult data acquisition,ensure the dataset comprehensiveness,and reduce the aerosol retrieval error to assist with AOD retrieval of high-resolution image aerosols.Methods We propose an atmospheric radiative transfer model to construct a simulated dataset and support the implementation of aerosol retrieval using deep learning methods.Firstly,the apparent reflectance of satellite bands in different conditions(observation geometry,surface reflectivity,atmospheric conditions,different AODs,etc.)is simulated by the atmospheric radiative transfer model.Then,based on the statistical relationship among parameters in the study area and the probability distribution, we combine different bands and parameters to construct a simulated dataset that conforms to real scenes. Next, we employ this dataset to perform aerosol retrieval with deep learning methods, and apply the trained retrieval model to Landsat-8 OLI sensor data and retrieve high-resolution (30 m) aerosol images above the urban surface of Beijing. To evaluate the model performance, we validate the results with AERONET ground stations and adopt four metrics including mean absolute error (MAE), root mean square error (RMSE), the Pearson correlation coefficient (R), and expected error (EE) to perform analysis and evaluation.Results and Discussions We combine the radiative transfer model with machine learning methods, and take Beijing as an example to achieve the urban aerosol AOD retrieval. This method has higher accuracy than ground-based measurements to avoid the disadvantage of insufficient training datasets from ground-based measurements in the past. The AOD values inverted from the four stations shown in Fig. 5 (a) - (d) have a strong correlation with AERONET measurements (R = 0. 8397-0. 9283), and more than 72% of the points are within the expected error with relatively stable retrieval results.The overall results of the study area [Fig. 5 (e)] show that the error of the retrieval values is relatively small. The MAE and RMSE are 0. 0685 and 0. 1029 respectively, and have a high correlation with AERONET measurements, with an Rvalue of 0. 8989. 74. 05% of the points are within the expected error, while some results still show the underestimation phenomenon, which is mainly manifested in regions with AOD values above 0. 5. In the spatial aerosol distribution under different levels of atmospheric pollution [Fig. 7 (a) - (h)], the overall trend of the retrieval results is reasonable with a continuous distribution, providing full-space coverage of the aerosol retrieval results. Additionally, the aerosol image at 30 m spatial resolution can provide local details and provide more detailed information on the spatial AOD distribution in urban areas, which is of significance in pollution source monitoring and other aspects.Conclusions We propose an atmospheric radiative transfer model to construct a dataset to support the retrieval of high-resolution aerosol data from Landsat-8 using deep learning algorithms. The 6S atmospheric radiative transfer model is adopted to simulate the apparent reflectance of different Landsat-8 OLI bands in different conditions. By traversing the parameters in the study area and considering the joint probability of different parameters, different real scenes are constructed to ensure the unbiased dataset. To address the ill-posed problem during the retrieval, we leverage geometric angle information and statistical information between adjacent band apparent reflectance to screen the sample data and limit the parameter combination. An aerosol retrieval model is trained using the simulated dataset, and aerosol retrieval experiments are conducted in Beijing to verify the accuracy against AERONET ground-based aerosol measurement data.The results show that the method of employing simulated data to support deep learning algorithms has high accuracy, small errors, and high correlation with the measured data (R=0. 8989). The RMSE and MAE are 0. 1029 and 0. 0685respectively, and about 74. 05% of the data falls within the expected error line. The proposed method addresses the bias and data volume problems of deep learning training samples using simulated data to better employ large amounts of training data and then learn more complex and accurate relationships between AOD and observation parameters, thereby achieving more accurate retrieval results.