Comparison between Multitemporal Graph Based Classical Learning and LSTM Model Classifications for Sits Analysis

Very High Resolution (VHR) multispectral Satellite Image Time Series (SITS) enables the production of temporal land cover maps, thanks to high spatial, temporal and spectral resolution of modern earth observation programs. Besides, statistical learning methods applied to SITS monitoring and analysis have created relatively efficient semi-automatic classification techniques. It would therefore be natural to think that the use of deep learning methods on SITS would lead to advances comparable to those known in the field of computer vision. However, when applied to concrete cases, the results are not as convincing. This paper proposes a comparison between a SOTAG (Spatial-Object Temporal Adjacency Graphs) SVM based spatio-temporal classification approach and the Recurrent Neuronal Network (RNN), LSTM (Long Short-Term Memory) model which is trained by historical SITS. The trained LSTM networks are then used to predict new time series data. Both methods perform a spatio-temporal map indicating the temporal profiles of cartographic regions. The proposed approaches will be applied on real and simulated SITS data. We will demonstrate that both results are comparable despite computational times and algorithms complexity.