Deep recurrent neural networks for land-cover classification using sentinel-1 insar time series

To date, the potential of multitemporal interferometric SAR (InSAR) data in land-cover mapping has not been fully explored despite suitable time series increasingly acquired from SAR sensors. Here, we suggest to use an LSTM (Long Short Term Memory) based land-cover classifier to address this problem. Spatial context is preserved by using grey-level spatial dependencies and morphological profiles. Further, a 4-LSTM-based model was trained to capture the temporal dynamics of InSAR coherence. Altogether 39 Sentinel-1 interferometric coherence pairs acquired over Donana in Spain were used to evaluate the method performance. Achieved more than 90% overall accuracy indicates the strong potential of developed InSAR recurrent approach in improving differentiation between various land cover classes.