A Benchmark for Learned SAR Data Compression On-Board

Synthetic-Aperture Radar (SAR) images are becoming more and more popular due to their resilience against adverse weather conditions and clouds. However, the rapid growth of SAR data places a significant burden on its storage and transmission. Consequently, efficient SAR data compression algorithms are needed, particularly to optimize bandwidth and downlink time after spaceborne acquisitions. In the last decade, numerous compression algorithms for SAR images have been proposed, some of them being based on optical image compression standards, such as JPEG, JPEG2000 or SPIHT. In order to perform compression, these algorithms rely on transformations such as the Discrete Cosine Transform (DCT) or the Discrete Wavelet Transform (DWT) to achieve spatial decorrelation. Subsequently, in case of lossy compression, the generated decorrelated coefficients are quantized before being encoded in a bit-stream to be downloaded to the ground. With the rise of Machine Learning methods to tackle remote sensing image processing problems, researchers have proposed various Convolutional Neural Network (CNN) architectures to perform SAR data compression. The structure of autoencoders, with their latent space, naturally complies to the spatial decorrelation step necessary to compress the images. The SAR image compression can be performed on-board, with a forward pass through the Encoder followed by the quantization and encoding of the latent space to further reduce the bit-rate. The generated bitstream is then transmitted to the ground, where the original image is reconstructed with the Decoder. While these models demonstrate promising performance, they are designed for ground-based processing with millions of parameters and resource-intensive operations. On the other hand, on-board data compression must meet the limited hardware resource constraints, be real-time and should minimize energy consumption. With this regard, this work presents a benchmark of an autoencoder for SAR data compression. The model is constrained to fit in space-qualified hardware, especially FPGA boards that are commonly deployed on-board satellites. Comparison is made with traditional compression methods, such as JPEG, JPEG2000 or SPIHT, using several image quality metrics and taking into account the particularities of SAR signal. In future work, this light-weighted autoencoder will be tested on Commercial-Off-The-Shelf (COTS) components suitable for space application