Filling Cloud Gaps in Satellite AOD Retrievals Using an LSTM CNN-Autoencoder Model

Satellite imagery enables spatially-temporally continuous monitoring and understanding of global environmental factors for a range of applications. This data, however, often suffers from gaps in retrieval from sensor malfunction or atmospheric interference, particularly dense clouds that obscure parts or all of an area. For dynamic datasets such as the MCD19A2 Aerosol Optical Depth (AOD) dataset, gap filling is especially challenging. The difficulty lies in the often large, continuous blocks of cloudy pixels with missing data that limit the ability of spatial filling and the daily fluctuation in features such as AOD that incur high difficulty in gap filling from temporal trends. In this study, we propose a spatiotemporal long short-term memory (LSTM) convolutional autoencoder method that effectively reconstructs missing data resulting from thick cloud interference for MODIS AOD data. The proposed method outperforms previous methods of reconstructing data lost to thick cloud interference in AOD retrievals with a generalized network achieving a weighted average PSNR, SSIM, and R
2
of 47.2, 0.992, and 0.941, respectively, between original, cloud-free days and those same days masked with simulated thick cloud interference without the need for additional covariates.