Neural Compression for Multispectral Satellite Images

Woojin Cho^*, Steve Andreas Immanuel^*, Junhyuk Heo, Darongsae Kwon

TelePIX
NeurIPS 2024 Workshop on Machine Learning and Compression
^*Equal Contribution

Overall architecture of ImpliSat
The left part shows the INR backbone, which takes spatial coordinates as input and is conditioned on resolution ψ and channel information η. The right part demonstrates the Fourier modulation process, which effectively represent the MSI data.

Abstract

Multispectral satellite images are essential for applications in agriculture, fisheries, and environmental monitoring. However, the high dimensionality, large data volumes, and diverse spatial resolutions across multiple channels present significant challenges for data compression and analysis. In this paper, we introduce ImpliSat, a unified framework specifically designed to address these challenges through efficient compression and reconstruction of multispectral satellite data. ImpliSat employs Implicit Neural Representations (INR) to model satellite images as continuous functions over coordinate space, capturing fine spatial details across varying spatial resolutions. Additionally, we propose a Fourier modulation algorithm that dynamically adjusts to the spectral and spatial characteristics of each channel, ensuring optimal compression while preserving critical image details.

Comparison of ground truth, shift modulation, scale modulation, and our proposed method with PSNR across multibands (B1, B5, B2) of the Cairo image. Each row represents a different GSD: 60m, 20m, and 10m, respectively.

Comparison of ground truth, shift modulation, scale modulation, and our proposed method with PSNR across multibands (B9, B8A, B4) of the Merapi image. Each row represents a different GSD: 60m, 20m, and 10m, respectively.

Comparison of ground truth, shift modulation, scale modulation, and our proposed method with PSNR across multibands (B9, B6, B3) of the London image. Each row represents a different GSD: 60m, 20m, and 10m, respectively.

Comparison of ground truth, shift modulation, scale modulation, and our proposed method with PSNR across multibands (B1, B7, B8) of the Hawaii image. Each row represents a different GSD: 60m, 20m, and 10m, respectively.

Comparison of ground truth, shift modulation, scale modulation, and our proposed method with PSNR across multibands (B10, B11, B8) of the Seoul image. Each row represents a different GSD: 60m, 20m, and 10m, respectively.

BibTeX

BibTex Code Here