DecSolNet: A noise resistant missing information recovery framework for daily satellite NO2 columns

Abstract A novel statistical method (hereafter referred to as DecSolNet) for reconstructing satellite NO2 columns is introduced. The method has been developed and evaluated by comparing its performance with four benchmark models in three scenarios. When the amount of satellite data is limited, DecSolNet outperforms the benchmark models and its performance does not degrade with noisy inputs. The implementation of DecSolNet consists of: (1) feature extraction, sequential data decomposition in both temporal and frequency domains; (2) NO2 columns reconstruction by training a deep neural network. In three cross-validations, the averaged R 2 score of DecSolNet reaches 0.9, which is better than that of the most benchmark models. The multi-layer perceptron (MLP) has a higher R 2 score, but it degrades greatly with noisy inputs, while the performance of DecSolNet remains robust with an R 2 of ~ 0.8. The bias of DecSolNet is small with an average of 1.61 μ g / m 3 . In addition, DecSolNet is a reliable learning machine, the averaged loss and standard deviation are 0.42 μ g / m 3 and 0.04 μ g / m 3 , respectively.