Compressed learning based onboard semantic compression for remote sensing platforms

Protim Bhattacharjee,PEter Jung

2024-09-03

Abstract:Earth observation (EO) plays a crucial role in creating and sustaining a resilient and prosperous society that has far reaching consequences for all life and the planet itself. Remote sensing platforms like satellites, airborne platforms, and more recently dones and UAVs are used for EO. They collect large amounts of data and this needs to be downlinked to Earth for further processing and analysis. Bottleneck for such high throughput acquisition is the downlink bandwidth. Data-centric solutions to image compression is required to address this deluge. In this work, semantic compression is studied through a compressed learning framework that utilizes only fast and sparse matrix-vector multiplication to encode the data. Camera noise and a communication channel are the considered sources of distortion. The complete semantic communication pipeline then consists of a learned low-complexity compression matrix that acts on the noisy camera output to generate onboard a vector of observations that is downlinked through a communication channel, processed through an unrolled network and then fed to a deep learning model performing the necessary downstream tasks; image classification is studied. Distortions are compensated by unrolling layers of NA-ALISTA with a wavelet sparsity prior. Decoding is thus a plug-n-play approach designed according to the camera/environment information and downstream task. The deep learning model for the downstream task is jointly fine-tuned with the compression matrix and the unrolled network through the loss function in an end-to-end fashion. It is shown that addition of a recovery loss along with the task dependent losses improves the downstream performance in noisy settings at low compression ratios.

Image and Video Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to achieve efficient semantic compression on remote sensing platforms to deal with the problem of limited downlink bandwidth. Specifically, the paper focuses on how to achieve efficient compression and transmission of image data through the compression learning framework on remote platforms such as satellites and drones, while maintaining the ability to perform downstream tasks (such as image classification) on these data. It is mentioned in the paper that traditional compression methods may not be suitable for resource - constrained remote platforms. Therefore, a new compression method needs to be developed, that is, encoding data by performing fast and sparse matrix - vector multiplication through a linear compressor. This method can not only reduce the amount of required data, but also maintain high downstream task performance in a noisy environment.

The paper proposes a model named GloCal (GLODISMO + Classifier), which consists of three parts: a compression matrix \(\Phi\), a recovery network \(f: \mathbb{R}^m\rightarrow\mathbb{R}^N\) and a classifier \(F\). Among them, the compression matrix and the recovery network together form the GLODISMO sub - network, and the classifier serves as the subsequent sub - network. The entire model is trained in an end - to - end manner, aiming to learn a compression matrix that can directly extract useful features from compressed data, thereby achieving efficient semantic compression. In addition, in order to improve the robustness of the model in low - compression - ratio and noisy environments, the author introduces recovery loss, which is used together with task - related losses (such as classification loss) to optimize the model.

Overall, the core problem of this paper is to design a new compression learning framework to achieve efficient and robust image data compression and transmission on resource - constrained remote sensing platforms, so as to support the high - quality completion of downstream tasks.