Time series prediction of multi-spectral satellite images and its application for cloud removal

In recent years, tremendous advancements in remote sensing techniques have enabled us to study almost any region of interest and obtain its relevant information as and when required. In this work, the focus is to develop self-supervised deep learning models that take advantage of the information conveyed by the Satellite Image Time Series (SITS), and extract a set of spatial and temporal features in the data. We aim to capture this complex spatio-temporal behavior in geo-spatial time series data which in turn can be used as an indispensable aide in several tasks. We use Convolutional Auto Encoders in combination with Multi-Headed Self-Attention Based Transformers. We have demonstrated the use of this model to predict future images. Further this is used to recover areas covered by clouds. The model is able to condense the information contained in SITS data into small vectors and recover them.