Neural Embedding Compression for Earth Observation Data – an Ablation Study

The exponential growth of Earth Observation data presents challenges in storage, transfer, and processing across fields such as climate modeling, disaster response, and agricultural monitoring. Efficient compression algorithms—either lossless or lossy—are critical to reducing storage demands while preserving data utility for specific applications. Conventional methods, such as JPEG and WebP, rely on hand-crafted base functions and are widely used. However, Neural Compression, a data-driven approach leveraging deep neural networks, has demonstrated superior performance by generating embeddings suitable for high levels of entropy encoding, enabling more accurate reconstructions at significantly lower bit rates.

In our prior work, we developed a Neural Compression pipeline utilizing a masked auto-encoder, embedding quantization, and an entropy encoder tailored for satellite imagery [1]. Instead of reconstructing original images, we evaluated the reconstructed embeddings for downstream tasks such as image classification and semantic segmentation. In this study, we conducted an ablation analysis to quantify the contributions of individual pipeline components—encoder, quantizer, and entropy encoder—toward the overall compression rate. Our findings reveal that satellite images achieve higher compression rates compared to ImageNet samples due to their lower entropy. Furthermore, we demonstrate the advantages of learned entropy models over hand-crafted alternatives, achieving better compression rates, particularly for datasets with seasonal or geospatial coherence. Based on these insights, we provide a list of recommendations for optimizing Neural Compression pipelines to enhance their performance and efficiency.

This work was conducted under the Embed2Scale project, supported by the Swiss State Secretariat for Education, Research and Innovation (SERI contract no. 24.00116) and the European Union (Horizon Europe contract no. 101131841).

[1] C. Gomes and T. Brunschwiler, “Neural Embedding Compression for Efficient Multi-Task Earth Observation Modelling,” IGARSS 2024, Athens, Greece, 2024, pp. 8268-8273, doi: 10.1109/IGARSS53475.2024.10642535.