Covariance Meets Context: Transformer-Based SAR Covariance Prediction Across Frequencies and Time

Abstract

Synthetic Aperture Radar (SAR) data differ significantly from optical images, and their statistics are strongly affected by the acquisition context, including geometry, polarization, frequency, and the temporal separation between acquisitions. We introduce COCONUT, a SAR-centered transformer-based architecture that leverages the entire SAR observation space by operating on second-order signal statistics captured by covariance matrices and accounting for the acquisition context. To demonstrate the proposed method, we develop a model capable of processing multifrequency polarimetric time series and translating the signal across frequencies and time. The predictions retain the SAR signal properties, including the phase, and are fully compatible with downstream methods operating on SAR covariance matrices. We analyze the effect of temporal separation between acquisitions and show that the model can produce dense time series that integrate information across several sensors with irregular sampling intervals. Considering the differences in frequency, we find that the X- and C-bands are easier to translate between each other than the longer-wavelength, deeper-penetrating L-band. The proposed method seamlessly handles missing data and supports joint information extraction across different SAR data dimensions, including polarization, frequency, and time. The implementation source code is available at https://github.com/nbasargin/sarcoconut.