Towards Efficient Neural Compression for Earth Observation Data

The rapid growth of Earth observation (EO) data presents a challenge in data management, data processing, and data storage. Missions like Sentinel-2 generate petabytes of multi-spectral imagery, contributing to the massive volume of information from diverse EO missions such as Landsat, and MODIS. In combination with climate-related satellites, e.g. MetOp, commercial and future satellites, a significant contribution to a climate-responsible and sustainable future is ensured [1,2]. The constantly growing volume of data has implications for hosting infrastructure: from satellite specifications, e.g. bandwidth and storage capacity, to the ground segment where the data is processed and managed. Any design choice in this flow of data impacts the corresponding CO2-footprint [3] and costs associated. As part of the “Embed2Scale” project [4] co-funded by the European Union (Horizon Europe contract No. 101131841), the Swiss State Secretariat for Education (SERI), and UK Research and Innovation (UKRI), we develop and evaluate compression techniques leveraging deep neural network models [5,6]. Our end-to-end pipeline leverages current progress in the area of pre-training with masked autoencoder (MAE) reconstruction tasks, contrastive encoders for feature extraction, quantization, and entropy models for latent distribution modeling [7]. Based on the foundations of the SSL4EO-S12 data set [6], these methods are designed to drastically reduce data size while preserving relevant information for a diverse set of EO downstream tasks. Efficient neural compression improves downlink efficiency and data storage on the ground, as the temporal, spatial and spectral dimensions have subtle correlations, and our neural compressor architectures exploit these correlations. We validate our methods in a range of real-world use cases, from agriculture to maritime awareness. The diverse set of downstream tasks allows us to verify the degree to which the compressed data retains relevant information. The long-term goal of our study is the implementation of an embedded neural compressor for satellites, such as ESA's Φsat-2 [8,9], where the benefits of initial on-board applications have been demonstrated.