Rate-distortion trade-off for learned semantic compression for remote sensing platforms

Remote sensing platforms such as satellites, UAVs, and HAPs generate massive volumes of imagery data, leading to a downlink bottleneck. On top of that, transmission of data is often restricted to certain time-intervals with limited bandwidth, making the downlink process even for compressed image data to a long-term obstacle. Deep learning-based approaches like learned semantic compression, enable the extraction and transmission of representative features instead of raw imagery, thus reducing data volume downlink time. In this work, we extend learned semantic compression pipelines to multispectral images and further include learned efficient quantization. The proposed pipeline integrates four components: a linear compressor, quantization, unrolled reconstruction network, and downstream semantic task. End-to-end training with semantic loss ensures that compression is aligned with the performance of tasks such as classification. We study the trade-offs among bit rate and downstream task accuracy and experiments on Earth observation applications demonstrate the effectiveness of the approach for land use and land cover classification. thus reducing data volume downlink time. In this work, we extend learned semantic compression pipelines to multispectral images and further include learned efficient quantization. The proposed pipeline integrates four components: a linear compressor, quantization, unrolled reconstruction network, and downstream semantic task. End-to-end training with semantic loss ensures that com- pression is aligned with the performance of tasks such as classification. We study the trade-offs among bit rate and downstream task accuracy and experiments on Earth observation applications demonstrate the effectiveness of the approach for land use and land cover classification.