Scholarlink

<p>The hyperspectral sounders aboard Low Earth Orbit (LEO) satellites produce 515 Gbits of brightness temperature data in each orbital pass, which is many times more than a typical S-band/X-band downlink window can accept. Current compression pipelines, ranging from CCSDS 123.0-B to retrieval-agnostic deep codecs, reduce data volume without preserving the spectral fingerprints required for climate gas retrieval; channel-equal PSNR/SAM loss functions implicitly treat every spectral channel as equally valuable, so high-sensitivity humidity-sounding channels lose the bit budget needed to maintain physical retrieval fidelity, introducing temperature-profile errors of 1.737&#xa0;K at compression ratios above 10:1. This paper proposes a physics-aware end-to-end edge AI pipeline, validated on three Chinese satellite collections: FengYun-3 MWHS-2 (15-channel THz sounder, 118–183&#xa0;GHz), FengYun-3E HIRAS-II (2,275-channel hyperspectral infrared sounder), and GaoFen-5 AHSI (330-band VNIR/SWIR imager). The pipeline comprises three components. First, a Radiative-Transfer-Loss (RTL) Codec—an INT8-quantised 1D + 2D convolutional neural network trained with a Jacobian-weighted RTTOV retrieval loss—achieves a compression ratio of 7.5:1 and limits temperature retrieval error to 0.90&#xa0;K (MWHS-2 synthetic, PSNR = 50.49&#xa0;dB) and 0.93/0.94&#xa0;K on FY-3 real data, satisfying the sub-1&#xa0;K NWP assimilation requirement. Second, a Cloud Filtering and Band Selection (CFBS) module reduces data volume by 38% while adding only + 0.10&#xa0;K retrieval error. Third, a seven-stage data engineering pipeline enables TensorRT INT8 execution on Xilinx Versal AI Core and NVIDIA Orin-Space hardware, with a total power of 5.0 W and a frame latency of 20&#xa0;ms. The combined system reduces the time required to deliver climate anomaly data from 75&#xa0;min, the latency of traditional store-and-forward downlink, to under 12&#xa0;min, a more than 6× improvement that provides a deployment blueprint for next-generation THz CubeSat missions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AI-Powered Edge Image Processing for Compression and Transmission of Hyperspectral Terahertz Climate Data in Low Earth Orbit Satellites

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Zhimin Gu,
  • Hongxin Zhang,
  • Bin Hu,
  • Bo Wang

摘要

The hyperspectral sounders aboard Low Earth Orbit (LEO) satellites produce 515 Gbits of brightness temperature data in each orbital pass, which is many times more than a typical S-band/X-band downlink window can accept. Current compression pipelines, ranging from CCSDS 123.0-B to retrieval-agnostic deep codecs, reduce data volume without preserving the spectral fingerprints required for climate gas retrieval; channel-equal PSNR/SAM loss functions implicitly treat every spectral channel as equally valuable, so high-sensitivity humidity-sounding channels lose the bit budget needed to maintain physical retrieval fidelity, introducing temperature-profile errors of 1.737 K at compression ratios above 10:1. This paper proposes a physics-aware end-to-end edge AI pipeline, validated on three Chinese satellite collections: FengYun-3 MWHS-2 (15-channel THz sounder, 118–183 GHz), FengYun-3E HIRAS-II (2,275-channel hyperspectral infrared sounder), and GaoFen-5 AHSI (330-band VNIR/SWIR imager). The pipeline comprises three components. First, a Radiative-Transfer-Loss (RTL) Codec—an INT8-quantised 1D + 2D convolutional neural network trained with a Jacobian-weighted RTTOV retrieval loss—achieves a compression ratio of 7.5:1 and limits temperature retrieval error to 0.90 K (MWHS-2 synthetic, PSNR = 50.49 dB) and 0.93/0.94 K on FY-3 real data, satisfying the sub-1 K NWP assimilation requirement. Second, a Cloud Filtering and Band Selection (CFBS) module reduces data volume by 38% while adding only + 0.10 K retrieval error. Third, a seven-stage data engineering pipeline enables TensorRT INT8 execution on Xilinx Versal AI Core and NVIDIA Orin-Space hardware, with a total power of 5.0 W and a frame latency of 20 ms. The combined system reduces the time required to deliver climate anomaly data from 75 min, the latency of traditional store-and-forward downlink, to under 12 min, a more than 6× improvement that provides a deployment blueprint for next-generation THz CubeSat missions.