A Hybrid CNN-RNN Deep Learning Network for Deriving Cyclonic Change Map from Bi-Temporal SAR Images

In this paper, the earth observation data is interpreted using machine learning techniques for analyzing the changes happened in water and vegetation area due the cyclone aftereffects. In particular, Synthetic Aperture Radar (SAR) data is considered as it is transparent to weather, cloud and dust. Thus SAR images could be effectively used for cyclonic areas to analyze the aftereffects. In the proposed methodology, a Convolutional Neural Network (CNN) is used to extract the feature map of the input sequence, and then a Recurrent Neural Network (RNN) is used to analyze the temporal dependencies of the bi-temporal input. Binary change map is generated as the output. Case studies are performed on the SAR images of areas affected by Gaja cyclone. The change map is used to analyze the destruction caused by the cyclone. Both qualitative and quantitative results of analysis demonstrate that the hybrid architecture (CNN + RNN) performs better than the existing state-of-the-art methodologies.