Time series analysis of satellite data using ConvLSTM for spatio-temporal feature extraction and prediction

Abstract

The rapid changes in climate of a particular place can effect the lives of local peoples and the area on which they are living. If we are able to detect those changes by mapping the spatial and temporal features of the high resolution satellite image and able to predict the changes before, then we can save ourselves from calamities. In this paper we have used two version of ConvLSTM to capture the spatio-temporal features of high resolution multi-spectral time series satellite images(Landsat-8 image data) and predict the next frame. In the first model(basic ConvLSTM) we simply use the ConvLSTM and predict the next image. The second model we have used is ConvLSTM with additional layer of 3D convolution and 3D Trans-convolution with extract more information about temporal and spatial features. The second model is fast in compare to first basic ConvLSTM model. The predicted result are shown in this paper after conducting experiments demonstrate that second model performs better.