Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HySpecNet-11k: A Large-Scale Hyperspectral Dataset for Benchmarking Learning-Based Hyperspectral Image Compression Methods

Martin Hermann Paul Fuchs, Begüm Demir

TL;DR

The paper tackles the data bottleneck in learning-based hyperspectral image compression by introducing HySpecNet-11k, a large-scale, unlabeled benchmark dataset of $11{,}483$ patches, each $128\times128$ with $202$ spectral bands, derived from EnMAP imagery. It benchmarks state-of-the-art convolutional autoencoders across spectral- and spatial-domain compression, revealing that spectral-domain compression via a 1D-CAE delivers the strongest rate-distortion performance (e.g., $PSNR$ ≈ $54.85$ dB at $CR$ ≈ $8.08$), while methods employing spatial downsampling lag in reconstruction quality. The dataset provides reproducible splits (patchwise easy and tilewise hard) and accompanying code, enabling robust evaluation of unsupervised learning on hyperspectral data. HySpecNet-11k is poised to advance research in hyperspectral learning by alleviating data scarcity and supporting generalization across diverse geographic regions, with plans to extend the dataset through additional EnMAP tiles.

Abstract

The development of learning-based hyperspectral image compression methods has recently attracted great attention in remote sensing. Such methods require a high number of hyperspectral images to be used during training to optimize all parameters and reach a high compression performance. However, existing hyperspectral datasets are not sufficient to train and evaluate learning-based compression methods, which hinders the research in this field. To address this problem, in this paper we present HySpecNet-11k that is a large-scale hyperspectral benchmark dataset made up of 11,483 nonoverlapping image patches. Each patch is a portion of 128 $\times$ 128 pixels with 224 spectral bands and a ground sample distance of 30 m. We exploit HySpecNet-11k to benchmark the current state of the art in learning-based hyperspectral image compression by focussing our attention on various 1D, 2D and 3D convolutional autoencoder architectures. Nevertheless, HySpecNet-11k can be used for any unsupervised learning task in the framework of hyperspectral image analysis. The dataset, our code and the pre-trained weights are publicly available at https://hyspecnet.rsim.berlin

HySpecNet-11k: A Large-Scale Hyperspectral Dataset for Benchmarking Learning-Based Hyperspectral Image Compression Methods

TL;DR

The paper tackles the data bottleneck in learning-based hyperspectral image compression by introducing HySpecNet-11k, a large-scale, unlabeled benchmark dataset of patches, each with spectral bands, derived from EnMAP imagery. It benchmarks state-of-the-art convolutional autoencoders across spectral- and spatial-domain compression, revealing that spectral-domain compression via a 1D-CAE delivers the strongest rate-distortion performance (e.g., ≈ dB at ≈ ), while methods employing spatial downsampling lag in reconstruction quality. The dataset provides reproducible splits (patchwise easy and tilewise hard) and accompanying code, enabling robust evaluation of unsupervised learning on hyperspectral data. HySpecNet-11k is poised to advance research in hyperspectral learning by alleviating data scarcity and supporting generalization across diverse geographic regions, with plans to extend the dataset through additional EnMAP tiles.

Abstract

The development of learning-based hyperspectral image compression methods has recently attracted great attention in remote sensing. Such methods require a high number of hyperspectral images to be used during training to optimize all parameters and reach a high compression performance. However, existing hyperspectral datasets are not sufficient to train and evaluate learning-based compression methods, which hinders the research in this field. To address this problem, in this paper we present HySpecNet-11k that is a large-scale hyperspectral benchmark dataset made up of 11,483 nonoverlapping image patches. Each patch is a portion of 128 128 pixels with 224 spectral bands and a ground sample distance of 30 m. We exploit HySpecNet-11k to benchmark the current state of the art in learning-based hyperspectral image compression by focussing our attention on various 1D, 2D and 3D convolutional autoencoder architectures. Nevertheless, HySpecNet-11k can be used for any unsupervised learning task in the framework of hyperspectral image analysis. The dataset, our code and the pre-trained weights are publicly available at https://hyspecnet.rsim.berlin
Paper Structure (6 sections, 2 figures, 1 table)

This paper contains 6 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. The HySpecNet-11k Dataset
  3. Learning-based Hyperspectral Image Compression
  4. Experimental Results
  5. Conclusion
  6. Acknowledgements

Figures (2)

  • Figure 1: True color representations of example images from our proposed HySpecNet-11k dataset. Red, green and blue channels are extracted from enmap bands 43.0, 28.0 and 10.0 at wavelengths 634.919nm, 550.525nm and 463.584nm, respectively.
  • Figure 2: Rate-distortion performance of learning-based hsi compression methods on the test set of our proposed HySpecNet-11k dataset (easy split). Rate is visualized in bpppc and distortion is given as psnr in decibel.