For-backward LSTM-based missing data reconstruction for time-series Landsat images

Reconstructing the missing data for cloud/shadow-covered optical satellite images has great significance enhancing availability and multi-temporal analysis. In this study, we proposed a deep-learning-based method reconstruction time-series Landsat images. Central to is combined use of autoencoder, long-short-term memory (AE-LSTM)-based similar pixel clustering backward LSTM-based prediction. First, manually delineated masks were overlaid onto produce pixel-wise with masking values. Second, these time series clustered by an AE-LSTM-based unsupervised into multiple clusters, searching pixels. Third, each cluster target images, for-backward-LSTM-based model was established restore values in data. Finally, reconstructed merged cloud-free (image) regions The applied three datasets Landsat-8 OLI results, showing improvement greater than 10% normalized mean-square error compared state-of-the-art methods, demonstrated effectiveness