为了解决现有陆地气溶胶光学厚度(AOD)反演算法精度和空间分辨率较低的问题,基于深度学习的思想,使用深度置信神经网络(DBN),实现了具有30 m空间分辨率的陆地气溶胶光学厚度反演。算法的训练样本包括全球长时间序列的AERONET站点数据以及在时空上与之对应的Landsat8 OLI的观测几何数据和表观反射率数据。为了保证反演的精度和稳定性,研究了AERONET站点数据的处理方法、卫星与站点数据的时空匹配方法以及DBN结构的设置。使用独立于训练样本的AERONET站点数据,对不同地表类型的550 nm处的AOD估算结果进行了整体验证,并对研究区域进行小尺度精度验证。结果表明,该方法的均方根误差和平均绝对误差分别为0.11与0.072。该方法打破了现有的气溶胶光学厚度反演方法依赖于其他遥感产品或者其他时相数据的局面,有效提高了气溶胶光学厚度反演的效率和空间分辨率。

To solve the problem that there exist low precision and spatial resolution in the retrieval algorithm of land aerosol optical depth (AOD), a deep learning-based deep belief network (DBN) is proposed to realize the retrieval of land AOD with a spatial resolution of 30 m. The training samples for the algorithm include the AERONET site data with global long time series as well as the observation geometric data and apparent reflectivity data from Landsat 8 OLI which are corresponding to the former in space and time. To ensure the estimation accuracy and stability of retrieval, the process method for the AERONET site data, the spatial-temporal matching method for satellite and site data, and the setting of the DBN structure are investigated. The AERONET site data, independent of the training samples, are used to test the AOD estimation results at 550 nm for different surface types as a whole. In addition, the small-scale accuracy verification is conducted in the study area. The results demonstrate that the root mean square error and the mean absolute error of the proposed method are 0.11 and 0.072, respectively. The proposed method can break the situation in which the retrieval of AOD based on the existing methods relies excessively on other remote sensing products or time-phase data, and it effectively improves the efficiency and spatial resolution in the retrieval of AOD.