Learned Spectral and Spatial Transforms for Multispectral Remote Sensing Data Compression

As more and more multispectral and hyperspectral platforms are deployed for Earth observation (EO), limited downlink capacity increases the pressure for more efficient data compression algorithms. Machine learning (ML) has been successfully applied to produce highly competitive compression models though this performance has typically been at the cost of high computational complexity, a crucial limitation for on-board remote sensing data compression. To address these issues, a reduced-complexity multispectral and hyperspectral data compression architecture is proposed. Using separate spectral and spatial transforms, the complexity of the proposed models is scalable on the number of bands, regardless of the compression ratios. This proposal outperforms state-of-the-art ML compression models as well as established lossy compression methods such as JPEG 2000 prepended with a spectral Karhunen-Loève transform (KLT) on a variety of remote sensing data sources. The performance improvement is achieved with a lower complexity than said ML models. To reproduce our results, training and test data are publicly available at https://gici.uab.cat/GiciWebPage/datasets.php and source code at https://github.com/smijares/mbhs2025.