A 3D Dynamical Hybrid Model: Coupling Statistical And Machine Learning Techniques To SAR Time Series Prediction

Multitemporal satellite images can be represented as a three-dimensional cube. This remote sensing data type require modeling techniques comprising spatial and temporal dependence altogether. This work aims at developing a hybrid framework combining the three-dimensional autoregressive (3D-AR) statistical model and machine learning algorithms to accommodate spatial and temporal correlations in a stack of temporal synthetic aperture radar (SAR) images. We propose using the 3D-AR to model the linear dependence among the voxels and one among artificial neural networks (ANN), support vector regression (SVR), and random forest (RF) for modeling the nonlinear component. The resulting models, labeled as 3D-AR-ANN, 3D-AR-SVR, and 3D-AR-RF hybrid models, respectively, can help to predict missing voxels or to forecast a one-step-ahead image of a remote sensing data sequence in the presence of linear and nonlinear patterns. The procedure is validate by applying the proposed hybridization approach to predict different parts of a SAR image stack from the urban area of Cordoba, Argentina. The results show the usefulness of our approach in predicting multitemporal SAR data, being more accurate than the state-of-the-art 3D-AR model.