· Paraphernalia

PPubMedUndated

Authors not listed by the archive

Abstract

Journal Information: ID: PLoS One, plos Title: PLOS One ISSN: 1932-6203 Publisher: Public Library of Science Location: San Francisco, CA USA Title: An end-to-end deep learning image compression method for satellite images based on entropy model Deep Learning Image Compression Metho Authors: - Yanlong Gao - Haiming Xu - Wei Huang - Hao Bai - Boer Peng - Heyang Xu - Hashmat Fida With the extensive applications of satellite image data in environmental monitoring and geographic surveying and mapping, the amount of data has increased rapidly, which brings great challenges for transmitting and s

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

An end‑to‑end deep‑learning framework compresses VIIRS satellite images about 30:1 while preserving high visual quality.

What’s new

i.Adapted a CDF‑based Gaussian‑mixture likelihood and checkerboard context for VIIRS multispectral blocks
ii.Designed an end‑to‑end compression pipeline with residual bottleneck blocks, simplified attention, hyperprior and context modules
iii.Proposed a block‑wise processing scheme that splits large images into 256×256 sub‑blocks for memory‑efficient compression
iv.Conducted systematic experiments on NASA VIIRS data showing ~30:1 compression ratio and state‑of‑the‑art rate‑distortion performance

How they did it

The authors trained a convolutional encoder‑decoder network with residual bottleneck and attention modules, a quantizer, hyperprior and context models, and an entropy coder using a Gaussian‑mixture likelihood. Training used 300 epochs on 256×256 VIIRS patches with five λ settings, Adam optimizer, and cosine‑annealing learning rate on an RTX 4090 GPU. Experiments evaluated compression ratio, bitrate (bpp), MSE, PSNR, SSIM, and runtime against JPEG, JPEG2000 and several learned baselines.

What they found

Average compression ratio ≈30:1 (data volume reduced to 1.5%‑4% of original)
Bitrate ≈0.51 bpp, MSE 1.45e‑4, PSNR 38.39 dB, SSIM 0.973
60.2% lower bpp than JPEG2000 with 3.24 dB higher PSNR and 0.032 higher SSIM
Outperformed learned baselines (Ballé, Minnen, Cheng, ELIC, Transformer‑CNN) in PSNR/SSIM while being faster than JPEG2000 and most learned methods

Caveats

✕Requires splitting large images into 256×256 tiles, which adds overhead and is inefficient for gigabyte‑scale data
✕Hybrid quantization may cause mismatch between training and inference, slightly degrading reconstruction quality
✕Runtime is slower than simple JPEG compression
✕Block‑wise approach may introduce stitching artifacts at patch boundaries

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

Similar papers in all four archives

The same idea searched in arXiv, PubMed, bioRxiv and chemRxiv at once, so a method can lead you to where it is being applied.

AarXiv4 found

PPubMed4 found

BbioRxiv4 found

CchemRxiv3 found

Searching the other archives…