Leveraging Neural Compression for Earth Observation

Abstract

The exponential growth of satellite data marks a new era in Earth observation (EO) and enables a better understanding of our planet, with applications such as crop mapping and the detection of natural hazards. However, the sheer volume of this data poses challenges for transmission, storage, and accessibility, ultimately limiting its usability. Image compression offers ways to efficiently store and transfer data. In recent years, data driven neural compression approaches have demonstrated improved performance in compressing images, compared to handcrafted algorithms (e.g., JPEG). Our work builds on that progress and focuses on the application of neural compression specifically on satellite images. We explore adaptations of neural compression models that leverage EO-specific characteristics including location and timestamp information. This metadata may offer potential to tailor and improve compression performance. Our research reveals fundamental differences in the input pixel distribution and entropy of satellite images compared to standard natural image datasets. Interstingly, EO data reveils a substantially lower entropy compared to ImageNet samples. We demonstrate that applying neural compression to EO data improves compression performance within a few hours of training, requiring lower bit rates and fewer parameters than models trained on natural images. We further investigate the effect of ingesting encoded metadata information on neural compression techniques. Our results suggest that neural models extract enough image features to make additional spatial and metadata inputs redundant. Finally, we compare specialized neural compression models, trained on specific seasons or geolocations, with general neural EO compressor trained on the entire EO data set. Our results indicate that although the specialized models are learning on different input distributions, general neural EO compressors are still beneficial in many cases. In low-entropy, strongly skewed distribution scenarios, specialized model outperform general neural EO compressors. These results underline that the specific nature of EO data may benefit from individual processing for certain parts of the data, while the majority can be compressed with a general neural EO compressor. Overall, we demonstrate superior perfromance of neural compressors relative to classical methods. Further, our findings suggest that tailoring neural compression to EO data partly requires specialization, as significant differences in input distribution and entropy can enable more specialized compression. We show that the extent to which neural compression models benefit from dataset diversity versus specialization is an essential trade-off, which requires further research from the EO domain. While tailoring neural compression to EO is generally challenging, we show that low entropy image samples allow for lightweight, specialized compressors, which can be very helpful for different EO scenes such as oceans, deserts or forests.