With the advent of advanced remote sensing technologies in past few decades, acquiring higher resolution satellite images has become easier and cheaper in recent days. However, on the other hand, it has offered a big challenge to the remote sensing community in smart image interpretation from such huge volume of data. Deep learning, which offers efficient algorithms for extracting multiple levels of feature abstractions, may be suitable to serve the purpose. This letter presents a deep learning approach (Deep-STEP) for spatiotemporal prediction of satellite remote sensing data. The proposed learning architecture is derived from a deep stacking network, consisting of a stack of multilayer perceptron, each of which models the spatial feature of the associated region at a particular time instant. The proposed method has been demonstrated on normalized difference vegetation index (NDVI) data sets, derived from satellite remote sensing imagery, containing several thousands to millions of pixels/records. The experimental results (related to NDVI prediction) reveal that the proposed architecture exhibits fairly satisfactory performance with promising learning capabilities.

The problem of missing data in remote sensing analysis is manifold. The situation becomes more serious during multitemporal analysis when data at various a-periodic timestamps are missing. In this work, we have proposed a deep-learning-based framework (Deep-STEP_FE) for reconstructing the missing data to facilitate analysis with remote sensing time series. The idea is to utilize the available data from both earlier and subsequent timestamps, while maintaining the causality constraint in spatiotemporal analysis. The framework is based on an ensemble of multiple forecasting modules, built upon the observed data in the time-series sequence. The coupling between the forecasting modules is accomplished with the help of dummy data, initially predicted using the earlier part of the sequence. Then, the dummy data are progressively improved in an iterative manner so that it can best conform to the next part of the sequence. Each of the forecasting modules in the ensemble is based on Deep-STEP, a variant of the deep stacking network learning approach. The work has been validated using a case study on predicting the missing images in normalized difference vegetation index time series, derived from Landsat-7 TM-5 satellite imagery over two spatial zones in India. Comparative performance analysis demonstrates the effectiveness of the proposed forecasting ensemble.