A Full 4-Way Tensor Network for Crop Classification with Time-Series Fully Polarization SAR Image

Time-Series Polarimetric Synthetic Aperture Radar (PolSAR) image is extensively applied in real-world crop classification tasks due to its ample characterization of crop growth stages and electromagnetic properties. However, this abundance information often leads to “dimension disaster” problem. Traditional methods attempt to mitigate the impact of high dimensional features by employing dimensionality reduction techniques in advance, which typically result in diminished classification accuracy due to the hard coupling between the dimensionality reduction model and the classifier. In this paper, a full 4-way tensor network is proposed, which is designed to efficiently extract the high-level time-series information and structural features directly from the original data, enabling classification without the need for dimensionality reduction procedure beforehand. Moreover, the proposed model has a simple structure and holds less parameters than the state-of-art methods. The performance of the proposed method is evaluated on simulated Sentinel-1 data provided by European Space Agency (ESA). The results validate the merits on agricultural applications of the proposed method.