Scholarlink

<p>In this paper, MTAP-Net (Multi-modal Terahertz Atmospheric Profiling Network), a new deep learning model to combine hyperspectral infrared (IR), terahertz (THz), and visible/near-infrared (VIS/NIR) images to obtain high-resolution atmospheric profiles of temperature, humidity, and trace gases, is proposed. The model employs a multi-branch encoder-decoder architecture with cross-modal attention to weight sensor inputs, making it sensitive to noise, content, and scene conditions. Physics-based loss functions enforce vertical smoothness, non-negativity, and consistency with radiative transfer. MTAP-Net is evaluated on synthetic and real IR + THz and has a temperature RMSE of 0.92&#xa0;K (clear-sky) and 1.58&#xa0;K (cloudy), improving on IR-only optimal estimation baselines by 35% and 45%, respectively. Condition RMSE reductions in humidity retrievals are 43–44%. These findings indicate that AI-based multi-modal fusion has great potential to enhance the atmospheric remote sensing operation of weather prediction, climate change modelling, and air quality monitoring systems.</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-Enhanced Multi-modal Image Fusion for Atmospheric Profiling Using Hyperspectral Terahertz and Infrared Satellite Data

  • 期刊论文

  • 发表日期：

    2026年4月25日

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

摘要

In this paper, MTAP-Net (Multi-modal Terahertz Atmospheric Profiling Network), a new deep learning model to combine hyperspectral infrared (IR), terahertz (THz), and visible/near-infrared (VIS/NIR) images to obtain high-resolution atmospheric profiles of temperature, humidity, and trace gases, is proposed. The model employs a multi-branch encoder-decoder architecture with cross-modal attention to weight sensor inputs, making it sensitive to noise, content, and scene conditions. Physics-based loss functions enforce vertical smoothness, non-negativity, and consistency with radiative transfer. MTAP-Net is evaluated on synthetic and real IR + THz and has a temperature RMSE of 0.92 K (clear-sky) and 1.58 K (cloudy), improving on IR-only optimal estimation baselines by 35% and 45%, respectively. Condition RMSE reductions in humidity retrievals are 43–44%. These findings indicate that AI-based multi-modal fusion has great potential to enhance the atmospheric remote sensing operation of weather prediction, climate change modelling, and air quality monitoring systems.