Compressive Learning for the Classification and Reconstruction of Synthetic Aperture Radar Data

Compressive learning (CL) for synthetic aperture radar (SAR) aims to reduce the volume of data required for effective SAR image processing while preserving classification performance and minimizing reconstruction loss. This study introduces a novel CL framework comprising three distinct scenarios: (I) direct classification from compressed measurements, (II) image reconstruction from compressed measurements, and (III) joint classification and reconstruction using a trainable compression layer. The proposed network includes a linear transformation layer that performs data compression, followed by multilayer perceptrons (MLPs) tailored for classification and reconstruction tasks. In the joint scenario, end-to-end training enables the compression layer to learn task-specific representations that improve both inference and data recovery. We evaluate our approach on the MNIST and MSTAR datasets across various compression ratios. Experimental results show that joint training significantly improves classification accuracy and reconstruction quality compared to fixed compression schemes. These findings highlight the potential of adaptive compressive learning for enhancing SAR data processing efficiency.