MIDNet: Neural-based Efficient Delivery of Multispectral Satellite Image Data

Yiyun Zhang, Zijian Wang

Dec 2024 · Australasian Database Conference (ADC)

View Full Text

Abstract

The transmission of multispectral satellite images in remote sensing is challenging due to the large data volume resulting from additional spectral bands and a higher dynamic range. To ensure the effective delivery within the limited bandwidth (e.g. from satellites to ground stations), image compression and neural delivery techniques can be applied to keep the transmission size at a minimum. However, since most learned image compression and neural delivery models are optimised for traditional RGB images, adapting these models may lead to suboptimal performance due to the high dynamic range of multispectral data. To address these limitations, this study introduces a novel neural-based delivery method named MIDNet. MIDNet combines both the compression efficiency of traditional codecs, such as JPEG 2000, with the reconstruction capabilities of neural networks. Unlike the conventional super-resolution models used in neural delivery, MIDNet employs fast compression algorithms to first generate intermediate-quality images, after which a proposed MID-RN resolving model is applied to enhance image reconstruction via residual learning. Through the extensive experiments, MIDNet with MID-RN achieves higher PSNR than JPEG 2000 and the state-of-the-art learned image compression models under the same bpp at lower compression rates. These results underscore MIDNet’s effectiveness for delivering multispectral satellite imagery, and indicate its potential for improving efficient image delivery in real-world applications.

BibTeX

@inproceedings{zhang2024midnet,
  author="Zhang, Yiyun and Wang, Zijian",
  editor="Chen, Tong and Cao, Yang and Nguyen, Quoc Viet Hung and Nguyen, Thanh Tam",
  title="MIDNet: Neural-Based Efficient Delivery of Multispectral Satellite Image Data",
  booktitle="Databases Theory and Applications",
  year="2025",
  publisher="Springer Nature Singapore",
  address="Singapore",
  pages="434--446",
  isbn="978-981-96-1242-0"
}